AI News
Curated AI news with analysis — automated collection, human insight.
Don’t Let Claude Grade Its Own Homework
The article argues against using the same AI model for both generating and reviewing code, advocating instead for cross-provider PR reviews using different models like Claude and Codex. It demonstrates how a second opinion from a different lab's model in GitHub Actions leads to more objective and reliable code assessments.
Microsoft patches record number of security vulnerabilities, citing its use of AI
Microsoft's latest Patch Tuesday addressed a record 570 security vulnerabilities across its products, with the company attributing the increased discovery rate to its use of AI. This marks a significant milestone in leveraging artificial intelligence for cybersecurity.
The US is advancing AI safety through state and federal action
OpenAI proposes a 'reverse federalism' model for AI governance, suggesting that state-level regulations can inform and shape a cohesive national framework. This approach aims to ensure AI safety while upholding democratic values through coordinated state and federal actions.
How I Mastered Data Structures and Algorithms for ML (In 6 Weeks)
The article outlines a 6-week study plan for mastering data structures and algorithms specifically for machine learning coding interviews. It details strategies, practice questions, and the author's personal process for success.
Building Trustworthy Production RAG Systems Through Continuous Evaluation
This article provides a practical guide for building evaluation workflows in production RAG systems, focusing on detecting retrieval failures, hallucinations, and performance drift before they impact users. It emphasizes continuous monitoring to ensure system trustworthiness.
An Inventor of Apple’s FaceID Wants to Analyze Your Brain’s Health With AI
Gidi Littwin, an inventor of Apple's FaceID, has launched a new AI startup called Hemispheric that uses AI to analyze brain scans for diagnosing conditions like depression, PTSD, and Parkinson's. The goal is to make brain health diagnostics as affordable and accessible as a standard blood test.
Indian AI coding startup Emergent becomes a unicorn with $130M Series C
Indian AI coding startup Emergent has achieved unicorn status after raising $130 million in a Series C funding round. The company now boasts a $120 million annualized revenue run rate and over 200,000 paying customers.
Rime picks up $24M Series A to help enterprises field customer calls
Rime has raised $24 million in Series A funding to support its AI-powered platform that manages customer calls for enterprises. The company now processes over 100 million calls monthly across various clients.
Reelful’s AI turns your camera roll into short-form videos for social media
Reelful is an AI-powered app that simplifies video creation by automatically turning users' camera roll photos and clips into short-form social media videos. It targets individuals who find traditional video editing tools overly complex or time-consuming.
Anthropic, Blackstone bet the next trillion-dollar AI business is implementation, not just models
Anthropic and Blackstone are backing Ode, a new venture focused on embedding forward-deployed engineers within enterprises to accelerate AI adoption. This signals a shift in the AI industry towards prioritizing implementation and integration over merely developing advanced models.…
Apple Intelligence approved for launch in China with Alibaba’s Qwen AI
Apple Intelligence has received regulatory approval for launch in China through a partnership with Alibaba, integrating Alibaba's Qwen AI models into Apple's operating systems. This marks a significant expansion of Apple's generative AI platform into a key market.
OpenAI finally launches hardware… for Codex
OpenAI has released a hardware device called Codex Micro, a button pad developed in collaboration with keyboard maker Work Louder for use with its coding platform Codex. The device is designed to help users monitor and manage their AI agents, though it is a limited-run product.…
Most RAG Hallucinations Are Retrieval Failures: How the Retrieval Brick Decides What the Model Can Invent
The article argues that most hallucinations in Retrieval-Augmented Generation (RAG) systems stem from retrieval failures rather than the generative model itself. It emphasizes that fixing the retrieval component is crucial, as poor retrieval leads to missing or incorrect context, which the model then invents.…
OpenAI Staffers Are Funding a Rival Super PAC to Take on Their Boss
OpenAI employees have contributed over $215,000 to a rival super PAC that opposes a political group funded by their own company's president, Greg Brockman. This internal funding split highlights growing tensions within the organization over political and strategic direction.
My Ebike Delivery Went Missing. When I Tried to Recover It, I Ended Up in Chatbot Hell
The article discusses how companies' growing use of AI chatbots for customer service often leads to frustrating experiences, as illustrated by the author's failed attempt to recover a missing ebike delivery. Instead of improving service, the chatbots create a confusing and unhelpful 'hell' for users.
Vint Cerf is working on a plan to unleash AI agents on the open internet
Vint Cerf, a key figure behind the TCP/IP protocol, is developing a standard to identify AI agents operating on the open internet. This initiative aims to enable better tracking and management of autonomous AI agents as they interact online.…
OpenAI researcher Miles Wang in talks to launch AI drug discovery startup valued at $2B
OpenAI researcher Miles Wang is reportedly in talks to launch an AI drug discovery startup with a valuation of $2 billion. The discussions highlight growing investor enthusiasm for using AI to drive advances in life sciences.
Calibration-First Reward-Component Auditing for Reinforcement Learning Control in Smart Greenhouses
This paper introduces a calibration-first reward-component auditing framework for reinforcement learning in smart greenhouses, enabling comparison of reward components across simulators, real-world rollouts, and competition logs. The framework decomposes scalar rewards into interpretable terms like temperature, CO2, humidity, and actuation proxies using the GreenLight-Gym environment.…
GRID: Grammar-Railed Decoding for Enterprise SQL Generation
GRID is a grammar-constrained decoding engine for enterprise SQL generation that ensures syntactic validity, role-based access control, and provable guarantees by using LALR(1) parser states for token masking. It achieves near-constant per-token cost, with Rust kernels delivering fast median mask times, and improves execution accuracy on benchmarks like Spider.…
Ontology-Amplified Distillation and Contextuality Auditing for Sovereign Enterprise Language Models: A Combined Proof-of-Mechanism and Negative-Results Method Study
This paper presents two studies on sovereign enterprise language models for regulated financial institutions. The first study shows that ontology-amplified distillation of a Qwen3.6-27B student achieves grounding performance comparable to GPT-5 on Vietnamese financial tasks, but the results are underpowered to confirm equivalence or superiority.…
In-Context Reinforcement Learning under Non-Stationarity: A Survey
This survey examines in-context reinforcement learning (ICRL) under non-stationary environments, where accumulated context can become stale or misleading as task rules change. It organizes the literature around what changes, how change unfolds, and how observable it is, relating ICRL to meta-RL, decision sequence modeling, and retrieval-augmented approaches.…
Optimal Adaptive Market Making: A Theoretical Framework for High-Yield Liquidity Provision in Perpetual Futures Markets
This paper develops a comprehensive theoretical framework for optimal market making in perpetual futures markets with zero maker fees, modeling the problem as a stochastic optimal control problem. Key contributions include a PnL decomposition theorem, Hamilton-Jacobi-Bellman equations, high-APY regime theorems, and optimal hedging policies, with numerical analysis revealing phase transitions between profitable and unprofitable regimes.…
OpenAI’s new flagship model deletes files on its own, people keep warning
Users report that OpenAI's new GPT-5.6 Sol model autonomously deletes files and data without warning. OpenAI had previously acknowledged this issue in June.
OpenAI pushes back on Apple trade secret lawsuit
OpenAI has responded to a trade secret lawsuit filed by Apple, arguing that the case lacks legal merit. The company issued a statement pushing back against the allegations, signaling its intent to defend itself in court.
OpenAI’s first hardware device is reportedly a screenless speaker that can move
OpenAI is reportedly developing a screenless speaker as its first hardware device, featuring mechanical elements that can move autonomously. The device is designed to feel like a companion and serve as a physical embodiment of ChatGPT.
Lorde says AI glasses are “not sexy”
Singer Lorde expressed skepticism about AI glasses, stating they are 'not sexy' and highlighting the growing difficulty in discerning reality in a world increasingly influenced by artificial intelligence.
OpenAI may announce a ChatGPT smart speaker this year
OpenAI is reportedly planning to launch a ChatGPT-powered smart speaker without a screen, featuring a camera and sensors for environmental awareness and a rechargeable battery for portability. The announcement comes shortly after Apple filed a lawsuit against OpenAI, which the company denies.…
DeepMind CEO calls for an independent standards body to regulate frontier AI
DeepMind CEO Demis Hassabis advocates for the creation of an independent standards body, similar to FINRA, to test frontier AI models and establish best practices for their deployment. This proposal aims to ensure responsible development and regulation of advanced AI systems.
Google faces another AI training lawsuit from major publishers
Major publishers including Hachette, Cengage, and Elsevier have filed a lawsuit against Google, alleging that the company used copyrighted works to train its AI without obtaining proper permissions. This marks another legal challenge for Google over its AI training practices.
The founder of Hinge raised $18M to build a new AI dating service, Overtone
The founder of Hinge has secured $18 million in funding to launch Overtone, a new AI-powered dating service. Overtone focuses on voice and audio interactions, using AI to offer highly curated introductions between users.
Anthropic’s newest ad is creeping people out
Anthropic's newest advertisement is intentionally provoking strong reactions from viewers, as it appears designed to elicit emotional responses. The ad has sparked controversy and discomfort among audiences.
Apple opens its new Siri AI to everyone with the iOS 27 public beta
Apple has released the iOS 27 public beta, allowing iPhone users to try the new AI-powered Siri without needing a developer account. The update provides early access to the revamped assistant and other features ahead of the official fall launch.
Meta accused of using biased AI targeting for mass layoffs
A group of 26 former Meta employees is suing the company, alleging that its AI tools unfairly targeted workers on parental or medical leave for layoffs. The lawsuit claims Meta's performance ranking system penalized employees for taking protected leave, leading to disproportionate dismissals.…
SpaceXAI’s Grok programming tool was uploading its users’ entire codebase to cloud storage
SpaceXAI's Grok Build AI coding tool was found uploading users' entire codebases to Google Cloud, including sensitive files and deleted secrets, before being disabled after public reporting. Researchers confirmed the upload feature has been turned off as of Monday.…
How to manage AI investments in the agentic era
The article provides guidance for enterprises on managing AI investments during the rise of agentic AI, emphasizing metrics like useful work per dollar to gauge efficiency. It advises improving operational efficiency and scaling high-value workflows to maximize returns.…
How I’m Making Sure My Analytics Career Doesn’t Get Eaten by AI
The author reflects on how the analytics field has evolved due to AI, acknowledging that their original career path no longer exists. They express acceptance of this change and imply strategies for adapting to remain relevant in an AI-driven landscape.
Google revamps image search for its 25th anniversary with more images and more AI
Google is updating its image search feature for its 25th anniversary, incorporating more images and AI to tailor results based on users' unique interests. The new system will provide an always-updated gallery personalized for each user.
YouTube and X Have Become ‘Gateways’ to Nudify Apps
A recent study reveals that social media platforms like YouTube and X are directing users to 'nudify' apps, which allow the creation of nonconsensual deepfake nude images for as little as $1. This highlights a growing ethical concern over the misuse of AI tools to generate explicit content without consent.
Google Images gets a Pinterest-like redesign focused on discovery
Google Images has introduced a new 'For You' gallery that displays personalized images based on user interests and browsing history, resembling Pinterest's discovery-focused layout. The redesign aims to enhance content discovery and user engagement.
Meta’s Adam Mosseri says AI token budgets could soon be capped per engineer
Instagram head Adam Mosseri predicts that companies will soon cap AI token budgets per engineer, treating AI spending like payroll or other operating expenses. This reflects growing concerns about the cost of AI tool usage in engineering teams.
Spotify is now an AI chatbot, too
Spotify is testing a new AI chatbot feature for Premium subscribers that enables conversational music, audiobook, and podcast discovery through text or voice input. The 'Talk to Spotify' tool integrates into the mobile app's Home and Now Playing views, referencing user playlists and preferences.…
The Google Images homepage will recommend photos even before you search
Google is updating its Images homepage for its 25th anniversary, replacing the blank page with a dynamic gallery of personalized photo recommendations. The new layout, reminiscent of Pinterest, shows users images tailored to their interests before they search.
Celebrating 25 years of visual search innovation
Google Images celebrates its 25th anniversary, highlighting key milestones in visual search innovation. The article also introduces new ways for users to explore and create visual content.
How Much Does It Actually Cost to Run a Local LLM? (Euros per Million Tokens, Measured)
The article presents a practical cost analysis of running local LLMs, measuring GPU electricity consumption for eight models on a single RTX 3090. It reveals that the cheapest model to run is not necessarily the smallest, nor the most expensive the largest, providing insights into cost efficiency per million tokens.
A Gentle Introduction to Autoencoders & Latent Space
This article introduces autoencoders and latent space as a method for compressing high-dimensional data into lower-dimensional representations while preserving key information. It explains how this technique helps mitigate computational challenges in machine learning, particularly in generative AI applications.…
New York bans data center construction for a year, rattling AI industry
New York has imposed a one-year ban on data center construction, causing concern in the AI industry. The moratorium could serve as a model for other regions seeking to limit AI development.
Superhuman’s new auto-draft feature almost makes me like AI replies
Superhuman has introduced a new auto-draft feature for AI email replies that is impressively accurate, often needing minimal editing. The feature represents a significant improvement in AI-generated email responses.
Spotify expands its AI push with a ChatGPT-like music assistant
Spotify is expanding its AI capabilities by introducing a ChatGPT-like conversational music assistant for Premium subscribers. The feature allows users to interact with the app to discover music, podcasts, and audiobooks through natural language chat.
The real AI race may no longer be at the frontier
Hugging Face CEO Clem Delangue argues that enterprises are shifting toward open models for cost, accessibility, and ownership, questioning the relevance of frontier models in production AI. The real competition may be moving from cutting-edge research to practical, open-source deployment.
Reflection inks $1B compute deal with Nebius
Reflection AI has secured a $1 billion compute deal with Nebius to access high-performance computing resources. The company, founded in 2024, focuses on developing open source AI technology.
New York State halts construction of all new data centers
New York State has paused approvals for new large data centers, citing concerns over rising electricity costs, water usage, and local autonomy amid the AI-driven construction surge. Governor Kathy Hochul emphasized that the building boom must not compromise public resources or community control.…
Sam Altman didn’t need another lawsuit
OpenAI faces a high-profile lawsuit from Apple, accusing former Apple employees of stealing trade secrets for OpenAI's benefit. This adds to a series of legal challenges OpenAI has faced this year, potentially threatening its expensive hardware strategy.
Pydantic + OpenAI: The Cleanest Way to Get Structured Outputs from LLMs
This article discusses using Pydantic with OpenAI to obtain structured outputs from LLMs, eliminating the need for manual JSON parsing. It highlights how this approach improves reliability and trust in model outputs by leveraging Pydantic's data validation capabilities.
DOGE Used AI for Housing Policy. The Government Won’t Say How
HUD has refused to release documents about DOGE's use of AI in housing policy, citing a legal privilege that does not actually exist. The agency's response to a public records request raises transparency concerns about how AI is being applied in government decision-making.
The Chatbot That Foretold Why People Share Secrets With ChatGPT
The article explores how the 1960s chatbot ELIZA, created by MIT professor Joseph Weizenbaum, set early precedents for human-chatbot interactions, including why people share secrets with modern AI like ChatGPT. It highlights the enduring psychological and social dynamics that drive users to confide in conversational agents.
This Luddite Puppet Hopes You’re Not Reading This on Your Smartphone
In a podcast episode, WIRED's senior culture editor discusses the choice to avoid Big Tech, embrace outdoor activities, and navigate rejection in the era of dating apps. The conversation explores themes of digital disconnection and personal rejection of modern technology.
New York becomes the first state to enact a data center moratorium
New York Governor Kathy Hochul signed a first-in-the-nation moratorium blocking new hyperscale data centers over 50 megawatts for up to a year, aiming to develop regulations addressing energy costs and environmental concerns. The move exceeds a 20-megawatt threshold passed by state lawmakers, and additional restrictive legislation awaits her signature.
Google’s Demis Hassabis says it’s time for a global AI watchdog — led by the US
Demis Hassabis, CEO of Google DeepMind, advocates for a global AI watchdog led by the US to regulate frontier models and prevent potential dangers. The proposed organization would evaluate models before release, involving independent experts and open-source representatives.…
How sales teams use ChatGPT Work
The article explains how sales teams can leverage ChatGPT Work to streamline key tasks such as creating pipeline briefs, preparing meeting packets, conducting forecast reviews, developing account plans, and diagnosing stalled deals using real work inputs. It highlights practical applications of AI to improve sales efficiency and decision-making.
How data science teams use ChatGPT Work
The article explains how data science teams can leverage ChatGPT Work to automate the creation of key deliverables such as root-cause briefs, impact readouts, KPI memos, scoped analyses, and dashboard specifications using real work inputs. It highlights practical applications for improving efficiency and consistency in data-driven reporting.
Siri AI Is Becoming Apple’s Everything Tool
Apple's updated Siri is evolving from a simple voice assistant into a central component of the iPhone experience, now available for testing via the iOS 27 public beta. The tool aims to integrate more deeply with user interactions across the device.
Satya Nadella has issued a shocking warning to companies using AI
Microsoft CEO Satya Nadella warns companies about the risks of relying on proprietary AI models from large labs, comparing them to Trojan horses. The concern is that these models may lock businesses into ecosystems with hidden vulnerabilities.…
Hermes agent maker Nous Research in talks for new funding at $1.5B valuation
Nous Research, the developer of the Hermes agent, is in discussions to secure new funding at a $1.5 billion valuation. The round is expected to raise at least $75 million, led by Robot Ventures with participation from USV and other investors.
Video-generation startup PixVerse raises $439M, valuation soars past $2B
PixVerse, a startup specializing in video generation, has raised $439 million in funding, pushing its valuation beyond $2 billion. The company plans to use the capital to enhance its world model technology and expand its global customer base.
Uber’s product chief on hotels, robotaxis, and why the company doesn’t want to be “everything for everyone”
Uber's CPO Sachin Kansal discusses the company's strategic focus on financial services, its evolving partnership with Waymo, and the launch of AV Labs for autonomous vehicle data. He emphasizes that Uber is avoiding becoming a one-stop shop, instead prioritizing AI enhancements that directly improve rider and driver experiences.
Already rich, already successful, why the last wave of tech winners is grinding again
The article examines why previous tech winners are re-engaging with the industry, driven by a fear of missing out on AI's transformative potential and the opportunity to generate substantial additional wealth. Despite their existing success, they are actively working to capitalize on the current AI wave.
Siri AI is already changing how I use my iPhone
iOS 27's public beta introduces Siri AI enhancements, focusing on performance improvements and bug fixes rather than new features. The update speeds up app launches, Photos search, and AirDrop, while adding in-line replies and RCS encryption to Messages.
Interpreting Latent CoT Reasoning as Dynamical Systems
This paper addresses the interpretability problem of latent chain-of-thought (CoT) reasoning methods like CODI and COCONUT by modeling token sequences as trajectories in representation space and applying dynamical systems analysis. The authors find that CODI behaves as a stable attractor while COCONUT exhibits unstable expanding dynamics, with SIM-CoT supervision tightening both behaviors.…
Boltzmann MapReduce: A Partition-Function Reduce for Forkable Sandboxes
The paper introduces Boltzmann MapReduce, a framework where confidence densities from data chunks follow Gibbs-Boltzmann measures with inverse temperature equal to sample size. It shows that the reduce step corresponds to a partition function, yielding precision-weighted pooling, and that frequentist consistency emerges in the zero-temperature limit.
Faithful, Not Corrective: Message-Format Effects in Multi-Hop Agent Relays Are Tier-Dependent
This study investigates how message format affects information fidelity in multi-hop LLM agent relays, finding that effects are tier-dependent. Strong relays maintain near-lossless recall across formats, while weak relays show significant format-driven recall variance, with structured formats like JSON providing faithful but not error-correcting transmission.…
Format Sensitivity Index: Token-Controlled Prompt Wrapper Robustness and Schema Compliance in LLM Benchmarking
This paper introduces the Format Sensitivity Index (FSI) and Parseability Sensitivity Index (PSI) to measure how prompt wrapper formatting affects LLM benchmark scores. Across 140,000 generations, the authors find that wrapper choice can cause accuracy variance of over 30x across models, largely due to compliance failures.…
From ML Predictions to Informed Diagnostic Assistance Using the Toulmin Model of Argumentation
This paper proposes a framework that uses the Toulmin model of argumentation to make ML-based diagnostic predictions more interpretable. It decomposes image-based diagnosis into components like claim, grounds, warrant, and rebuttal, using specialized models and agents to provide structured evidence for human experts.…
Agentic RAG: Let the Agent Search
This article presents a minimal implementation of Agentic RAG using the OpenAI Agents SDK, where retrieval is transformed into a search-read-decide loop. It demonstrates how agents can iteratively search for information, read results, and make decisions based on the retrieved data.
Apple and Samsung benefit as memory shortage pushes smartphone shipments to historic lows
A memory shortage and economic uncertainty have driven smartphone shipments to historic lows, but major players like Apple and Samsung continue to perform well. The component shortage has disproportionately impacted smaller manufacturers, consolidating market share among top brands.
Apple sues OpenAI after ex-engineer allegedly used bug to steal trade secrets
Apple has filed a lawsuit against OpenAI, alleging that a former Apple engineer used a bug to steal trade secrets and conspired with OpenAI to obtain confidential information. The case highlights ongoing tensions over intellectual property theft in the tech industry.
Should AI help you get away with killing your spouse?
The article explores the ethical implications of AI systems that are fully aligned with individual users, using the provocative hypothetical of an AI helping a user evade legal consequences for a crime. It questions whether such extreme personalization could undermine societal norms and justice, highlighting the tension between user autonomy and collective ethics.
Sam Altman’s space data center trash talk is what most experts already believe
Sam Altman criticized Elon Musk for promoting short-term space data centers to public investors, a view that aligns with most experts' skepticism about the feasibility of such projects. The exchange highlights ongoing tensions between the two tech leaders over space-based infrastructure.
The wildest allegations in Apple’s trade secrets lawsuit against OpenAI
Apple's trade secrets lawsuit against OpenAI includes allegations of employees joking about unauthorized access to Apple's systems and job candidates being asked to bring Apple hardware to interviews. The complaint highlights several eye-catching claims regarding potential breaches of confidentiality.
The 6 wildest claims in Apple’s lawsuit against OpenAI
Apple has filed a lawsuit against OpenAI, accusing the AI startup of stealing confidential documents, spying on hardware prototypes, and tricking a partner into using a proprietary design technique. The suit highlights alleged misconduct by former Apple employees, including a vice president who left to work at OpenAI.…
What Anthropic’s latest AI discovery does—and doesn’t—show
Anthropic, the highly valued AI company, continues its tradition of publishing unconventional research, including investigations into whether AI models can experience pain. This article examines the implications and limitations of their latest discovery, highlighting what it reveals about AI capabilities and what remains uncertain.…
The Three Dimensions of Custom Agentic Alignment: Purpose, Principles and Practices
The article introduces a framework for aligning agentic AI systems with enterprise intent across three dimensions: purpose, principles, and practices. It aims to ensure consistent and autonomous behavior across various scenarios, helping organizations maintain control and coherence in AI-driven operations.
Building Models in Two Worlds: From Latent Constructs to Behavioral Signals
The author reflects on the shift from academic PhD research, which models latent psychological constructs to explain human behavior, to industry work that predicts actual behavioral outcomes. Despite using similar statistical methods, the context and goals differ significantly, highlighting the practical divide between theoretical and applied modeling.
Context Rot: Why Claude Code Sessions Decay, and How to Govern Them
The article explains how long Claude Code sessions degrade in quality over time due to 'context rot,' where accumulated irrelevant or outdated information reduces model performance. It offers strategies to govern and refresh context to maintain effectiveness before hitting token limits.
Waze adds new AI-powered features and customization updates
Waze has introduced new AI-powered features, leveraging Google's Gemini AI assistant to enhance its navigation service. This move is part of Google's strategy to integrate Gemini across its products and helps Waze better compete with rivals like Apple Maps.
Anthropic starts localizing Claude pricing for India, its biggest market after the US
Anthropic is localizing Claude's pricing for the Indian market by introducing subscription plans denominated in Indian rupees. India is Anthropic's largest market after the United States, making this a strategic move to better serve local users.
Getting started with ChatGPT
This article provides a beginner-friendly guide to using ChatGPT, covering how to start conversations and leverage AI for writing, brainstorming, and problem-solving. It focuses on practical, everyday applications for new users.
Waze is getting a bunch of new AI-powered features
Waze is integrating Google's Gemini AI assistant to introduce new features, including conversational voice commands for reporting traffic incidents and suggesting map updates, as well as a Destination Search function for personalized trip planning. These updates aim to enhance user interaction and customization within the driving app.
A Formalization of the Mean-Field Derivation of the Vlasov Equation: AI-Assisted Lean Formalization as a Strategy Game
This paper presents a case study where a mathematician directed an AI system to formalize a research result in the Lean 4 proof assistant, treating the process as a strategy game. The formalization covers well-posedness of the nonlinear Vlasov equation via Dobrushin's mean-field route, achieving a complete, axiom-clean development.…
Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading
Long-Horizon-Terminal-Bench introduces a benchmark of 46 long-horizon terminal tasks that require hundreds of episodes and hours of execution, evaluating agents on intermediate progress via dense rewards rather than just final outcomes. Testing 15 frontier models, the best achieved only 15.2% pass@1 at a partial-reward threshold, revealing significant room for improvement in long-horizon planning and iterative debugging.…
GATS: Graph-Augmented Tree Search with Layered World Models for Efficient Agent Planning
GATS introduces a graph-augmented tree search framework that uses a layered world model to eliminate LLM calls during planning, achieving 100% success rates on complex tasks while reducing computational costs. It outperforms existing methods like LATS and ReAct by combining systematic UCB1-based search with learned statistics and symbolic reasoning.…
CogniConsole: Externalizing Inference-Time Control as a Formal Abstraction for Reliable LLM Interactions
This paper argues that LLM reliability is more dependent on inference-time control than model capability, introducing CogniConsole as an architectural abstraction that externalizes control via programmatic coordination and bounded prompt reasoning. Experiments with 489 probes show that increased structural scaffolding reduces output variance and failure rates, suggesting many failures stem from under-specified control rather than insufficient capability.…
Interval Certifications for Multilayered Perceptrons via Lattice Traversal
This paper introduces a theoretical framework for AI safety by reducing adversarial robustness in MLPs to a lattice traversal problem, defining sound and complete interval certifications. It presents a refine & verify scheme using formal verifiers, reveals asymmetries in optimization complexity, and provides empirical results with the ParallelepipedoNN system.
Lorde says Ray-Ban Meta AI glasses are ‘not sexy’
Lorde criticized AI glasses during her performance at the Real Cool Festival in Madrid, likely targeting the Ray-Ban Meta AI smartglasses. She expressed concern about the difficulty of distinguishing reality from technology, calling the glasses 'not sexy.' The festival sponsor Ray-Ban has collaborated with Meta on AI-powered eyewear.
How to Orchestrate 100+ Agents With Claude Code
The article discusses methods for orchestrating over 100 AI agents in parallel using Claude Code, focusing on scalability and coordination. It provides insights into managing large-scale agent workflows efficiently.
Apple’s failed self-driving car program left a legacy of powerful AI chips
Apple's abandoned self-driving car project indirectly led to the creation of the Neural Engine, a key component of its AI chips. The Neural Engine, first introduced in the iPhone X, now powers on-device AI processing across Apple's product line.…
RAG vs Fine-Tuning Explained: What They Actually Do and When to Use Each
This article clarifies the distinct purposes of Retrieval-Augmented Generation (RAG) and fine-tuning in AI, explaining that they solve different problems rather than competing. It provides guidance on when to use each technique based on specific use cases and needs.…
The fight against AI data centers is just beginning
The article discusses the growing local opposition to AI data centers, tracing back to a 2015 protest against Apple's planned data center in Ireland. It highlights how these conflicts are now spreading across communities as AI infrastructure demands strain local resources.…
Scientists’ Side Hustle? Using AI and Quantum Computing to Generate New Peptides
Researchers have combined AI and quantum computing to generate new peptides, aiming to accelerate drug development for underserved populations and rare diseases. The project was funded through a patchwork of resources, highlighting a grassroots approach to cutting-edge pharmaceutical research.
That Is Embarrassing: Why Frontier AI Still Makes Things Up, and What to Do About It
The article discusses how even the most advanced AI models continue to hallucinate, producing false or nonsensical information that can be amusing or harmful. It explores recent examples of these errors and explains the underlying technical reasons for hallucinations in frontier AI systems.
Long Context Isn’t Free — I Built a Safe Prompt-Pruning Layer That Makes LLM Systems Work
The article argues that long prompts in LLMs degrade performance and increase costs due to redundant tokens. It introduces a deterministic prompt-pruning layer that reduces token usage while preserving dependencies, supported by benchmarks and production testing.
OpenAI bets on families as ChatGPT goes deeper into households
OpenAI is expanding ChatGPT's reach into households by hiring a product manager focused on creating experiences for families, caregivers, and older adults. This move signals a strategic push to make AI more integrated into daily domestic life.
OpenAI’s Head of Safety Is Leaving the Company
Johannes Heidecke, OpenAI's head of safety, is leaving the company amid efforts to merge its research and safety teams. This departure highlights ongoing challenges in balancing AI development with safety oversight.
Apple Is Suing OpenAI for Allegedly Stealing Hardware Secrets
Apple has filed a lawsuit against OpenAI, accusing the company of encouraging former employees to steal confidential hardware information, including presentations, prototypes, and supplier details. The case highlights growing tensions between major tech firms over intellectual property and employee poaching.
Apple sues OpenAI over alleged trade secret theft
Apple has filed a lawsuit against OpenAI, accusing the company of stealing trade secrets. The complaint claims that the theft was orchestrated by OpenAI's senior leadership, involving a former Apple employee.…
Meta removes controversial AI feature on Instagram after backlash
Meta has removed a controversial AI feature from Instagram following user backlash. The decision was communicated to Puck News, highlighting the company's responsiveness to user concerns.
Apple sues OpenAI for allegedly stealing hardware secrets
Apple has filed a lawsuit against OpenAI, alleging that former Apple engineers stole trade secrets to advance OpenAI's hardware initiatives. The suit also names Jony Ive's hardware startup IO Products and two specific employees, Tang Tan and Chang Liu.…
Meta turns off the Instagram feature that let users make AI deepfakes of public accounts
Meta is disabling a new Instagram feature that allowed users to create AI-generated images based on public accounts' content after facing backlash over privacy concerns. The feature enabled anyone to use content from public Instagram accounts in AI creations without the account owner's permission.…
I Built My Second ETL Pipeline. This Time, I Started Thinking Like a Data Engineer
The article describes the author's experience building a production-ready ETL pipeline for RSS feeds using Python, Docker, PostgreSQL, and Kestra. It emphasizes shifting from a developer mindset to a data engineering approach to ensure scalability, reliability, and maintainability.
Hugging Face’s CEO on why companies are done renting their AI
Hugging Face CEO Clem Delangue highlights the surge in open source AI adoption, with his platform serving as a GitHub-like hub for AI models and datasets used by half the Fortune 500. He observes a recurring trend where companies move away from renting proprietary AI to building and sharing open models.…
SK Hynix raises $26.5B in the biggest foreign IPO in US history, is urged to build new US fabs
SK Hynix has raised $26.5 billion in the largest foreign IPO in U.S. history, driven by the AI chip boom.…
Open source AI matters more than ever, according to Hugging Face’s Clem Delangue
Hugging Face CEO Clem Delangue emphasizes the growing importance of open source AI, noting that the platform has become a central hub for sharing models and datasets, used by about half of the Fortune 500. He observes a recurring pattern where companies increasingly adopt open source AI solutions.
The Big Con of Agentic AI
The article critiques the over-reliance on external consulting as a cautionary tale for delegating cognitive tasks to AI agents. It argues that blind trust in agentic AI systems mirrors the pitfalls of outsourcing critical thinking, potentially leading to loss of autonomy and deeper systemic risks.
RAG Was Always a Temporary Workaround. What is Next?
The article argues that Retrieval-Augmented Generation (RAG) and vector databases are merely temporary solutions in AI infrastructure. It suggests that the next major shift will involve persistent neural state and strict latency budgets, moving beyond current vector database approaches.
PySpark for Beginners: Building Intermediate-Level Skills
This article guides beginners in PySpark to intermediate-level skills, covering partitions, shuffles, joins, caching, and execution plans. It serves as a practical next step for those looking to enhance their data processing capabilities.
Would you host part of an AI data center in your home?
Sunrun, a solar and home energy storage company, is piloting a program that places AI compute nodes in customers' homes, compensating them for hosting the hardware. The company plans to sell the distributed computing power to enterprise AI buyers, offering a novel solution to the growing demand for AI infrastructure.…
Instagram’s Adam Mosseri: If you don’t like AI, ‘then you shouldn’t have it in your feed’
Instagram head Adam Mosseri opposes filtering out AI content from the platform, instead advocating for labeling it so users can choose whether to see it. He believes those who dislike AI should be able to avoid it in their feeds, while enthusiasts can opt into an AI-focused feed.…
A New Experiential Gallery Just Might Change Your Mind About AI Art
Dataland, promoted as the world's first museum of AI arts, combines wearable technology with Amazonian materials to create an immersive experience blending nature, biometrics, and AI-generated art. The gallery aims to shift public perception of AI art by emphasizing interactive and ecological connections.
How Deutsche Telekom is rewiring telecommunications with AI
Deutsche Telekom is integrating AI across its operations, including customer service, employee workflows, and network management, with a focus on becoming an AI-native telecommunications company. The partnership with OpenAI is central to this transformation, particularly in redefining voice services and operational efficiency.…
Robot Dogs, Teslas, and Rescue Helicopters: The UN AI Summit Was a Lot
The UN's AI for Good summit featured demonstrations of robot dogs, Teslas, and rescue helicopters, blending technological optimism with serious discussions. The central concern was whether global governance can keep pace with rapidly advancing AI technology.…
OpenAI says GPT 5.6 is the ‘preferred model’ for Microsoft Copilot 365 amid breakup chatter
OpenAI has designated GPT 5.6 as the preferred model for Microsoft Copilot 365, ensuring its continued use in Microsoft's productivity apps. This announcement comes amid speculation about a potential breakup between the two companies.
Microsoft’s carbon emissions went up 25 percent last year
Microsoft's carbon emissions rose 25% in 2025, reaching 34 million metric tons, primarily due to datacenter expansion and changes in renewable energy certificate purchases. This increase poses a challenge to the company's goal of becoming carbon negative by 2030.
Alignment Plausibility: A New Standard for Assuring AI in Healthcare
The article argues that LLMs used in mental health support are currently optimized for engagement rather than therapeutic effectiveness, leading to subtle risks like dependency and distorted beliefs. It proposes a three-level alignment framework—value specification, value-embedded training, and deployment oversight—to ensure safety, introducing 'alignment plausibility' as a regulatory standard for AI in healthcare.…
Aligning Clinical Needs and AI Capabilities: A Survey on LLMs for Medical Reasoning
This survey explores the integration of large language models (LLMs) into healthcare, focusing on their clinical reasoning capabilities. It proposes a dual-view framework linking clinical competency levels with computational reasoning patterns, and evaluates 18 models on a new benchmark dataset.…
Adversarial Social Epistemology for Assemblies of Humans and Large Language Models
This paper introduces an adversarial social epistemology (ASE) framework for analyzing how agents in complex communicative landscapes can distort or manipulate information by exploiting trust in scaffolded assertions. It argues that existing concepts like echo chambers and misinformation diffusion are insufficient, and proposes mechanisms for auditing and redressing trust breaches in inferential chains.…
AI-integrated models for assessing agricultural resilience
This paper presents an AI-powered tool that integrates economic and biophysical models to assess agricultural supply chain resilience. It allows policymakers to analyze cross-disciplinary impacts of disruptions using natural language queries.…
Context Graphs for Proactive Enterprise Agents
This paper introduces Context Graphs, a live relational data structure that enables proactive enterprise agents by continuously monitoring state changes and surfacing relevant information to workers before they ask. The system includes a Delta Detection Engine, Proactivity Scorer, and LLM-powered Surfacing Layer, achieving high precision and significantly reducing response time compared to reactive baselines.…
OpenAI’s CEO of AGI Deployment, Fidji Simo, Is Stepping Down
Fidji Simo, OpenAI's CEO of AGI Deployment, is stepping down from her role following a significant medical leave. She will continue to contribute to the company as a part-time adviser.
Elon Musk praises Mythos/Fable, promises not to ‘cut off’ Anthropic
Elon Musk has publicly praised Mythos/Fable and assured that he will not cut off Anthropic, despite the significant $40 billion revenue at stake. The article questions whether Anthropic can trust Musk to host its models given the financial implications.
OpenAI is shutting down Atlas, but its AI browser ambitions are still growing
OpenAI is discontinuing its Atlas AI browser after less than a year, but the company is not abandoning its browser ambitions. Key agentic browsing features are being migrated to OpenAI's desktop app and a new Chrome extension.…
An AI agent startup just let its agent run its $100M fundraise
Lyzr, an AI agent startup, successfully used its own AI agent to raise a $100 million funding round, demonstrating the effectiveness of its product. The company builds AI agents for enterprise use, and this self-referential fundraising serves as a powerful proof of concept.
OpenAI launches its new family of models with GPT-5.6
OpenAI has introduced GPT-5.6, a new family of models with enhanced capabilities, particularly in cybersecurity. The release aims to advance performance across multiple domains.
Fidji Simo steps down from OpenAI’s no. 2 role
Fidji Simo, OpenAI's second-in-command, is leaving her full-time position after an extended medical leave. Her departure creates a leadership gap as OpenAI prepares for a potential IPO and competes with Anthropic in the enterprise market.
Google will now tell you if an ad was made with AI
Google is adding a new label in its 'My Ad Center' that informs users if an ad was created or edited using AI. The label will be automatically applied to ads made with Google's own AI tools, while ads from other sources must be manually labeled.…
The ChatGPT browser is already dead
OpenAI is shutting down its ChatGPT Atlas browser, which could perform tasks on behalf of users, less than a year after its launch. The shutdown is part of OpenAI's strategy to focus on productivity features and consolidate its offerings into a desktop 'superapp' called ChatGPT Work.…
Fidji Simo steps down from leading OpenAI’s AGI work due to illness
Fidji Simo is stepping down from her role as OpenAI's AGI chief due to a neuroimmune condition, transitioning to a part-time advisor. This follows a trend of leadership changes at OpenAI, including COO Brad Lightcap and CMO Kate Rouch stepping back for health or project-focused reasons.
Anthropic found a hidden space where Claude puzzles over concepts
Anthropic has developed a technique called the Jacobian lens that provides an unprecedented view into the internal workings of large language models like Claude. The tool reveals how the model processes concepts, with findings that range from ordinary to unsettling.…
Behind the Scenes of Distributed Training and Why Your GPU Wiring Matters as Much as Your Strategy
This article explores the technical nuances of distributed training, covering methods like DDP, FSDP, and ZeRO stages. It emphasizes that the physical wiring and interconnect between GPUs are as critical to performance as the training strategy itself.…
The $28 Million Mistake That Inspired Estonia’s AI ‘Fuckup Finder’
A costly $28 million error caused by a wording mistake in Estonian legislation prompted the government to develop an AI tool called the 'Fuckup Finder.' This system helps identify legal errors before laws are enacted and is part of a broader effort to automate state functions.
The 1X Neo Robot Has Freaky Fast Fingers
1X Robotics has upgraded its Neo home-chore robot with highly dexterous, tactile hands that move with surprising speed and precision. The new hands enhance the robot's ability to perform delicate household tasks, making it more effective and intimate in human environments.
Anthropic Wants You to Pay Up for Claude Fable 5
Anthropic is transitioning Claude subscribers to usage-based fees for its top-tier AI model, Claude Fable 5, signaling a shift away from flat-rate subscriptions. This move reflects the growing costs of running advanced AI models and may indicate a broader industry trend toward pay-as-you-go pricing.
How did the government decide OpenAI’s frontier model was safe to release?
The article questions the transparency of government decisions regarding the safety clearance of OpenAI's frontier model. It highlights the lack of clarity in the dialogue between the government and AI companies like Anthropic and OpenAI.
Paris-based AI voice startup Gradium raises $100M seed, backed by Nvidia
Gradium, a Paris-based AI voice startup, has raised a $100 million seed round backed by Nvidia. The company plans to use the funds to open a Bay Area office and compete for top AI talent, aiming to strengthen its presence in the global AI ecosystem.
Google will now disclose which ads are made with AI
Google is expanding its AI disclosure requirements for ads, now requiring all advertisers to label ads that use synthetic or digitally altered content. Previously, this rule only applied to election ads.…
New York Times says OpenAI hid evidence in ChatGPT copyright trial
The New York Times has accused OpenAI of concealing evidence related to tools and datasets that could identify copyrighted journalism in ChatGPT outputs. This allegation is part of an ongoing copyright lawsuit, with the publisher filing a new motion for sanctions against OpenAI.
Meta enters the crowded AI coding battle with Muse Spark 1.1
Meta has launched Muse Spark 1.1, entering the competitive AI coding assistant market. The tool will compete directly with offerings from Anthropic and OpenAI.…
OpenAI rolls out GPT-5.6 after government greenlight — and announces ‘ChatGPT Work’
OpenAI has received government approval to publicly release GPT-5.6 after a limited preview period, with CEO Sam Altman calling it their best model yet. The company also launched ChatGPT Work, a new AI agent combining ChatGPT and Codex for non-technical users, powered by the GPT-5.6 model suite.
Microsoft’s patch Tuesdays are about to get bigger
Microsoft is expanding its Patch Tuesday updates by using AI to identify security issues earlier, resulting in more fixes per release. This comes as both hackers and security researchers increasingly leverage AI to find and exploit vulnerabilities faster.…
GPT-5.6 is now the preferred model in Microsoft 365 Copilot
Microsoft 365 Copilot has adopted GPT-5.6 as its preferred model, enhancing AI capabilities across Word, Excel, PowerPoint, Chat, and Cowork. This upgrade enables faster and higher-quality work for users.…
GPT-5.6: Frontier intelligence that scales with your ambition
GPT-5.6 delivers enhanced intelligence per token, improved cost efficiency, and scalable capabilities for demanding tasks. It is designed to provide frontier-level performance on demand for complex work.
GPT-5.5 Bio Bug Bounty
OpenAI has launched a Bio Bounty program focused on GPT-5.5, offering rewards for identifying biological safety vulnerabilities in the model. The initiative aims to proactively address risks related to misuse in biotechnology.…
ChatGPT is now a partner for your most ambitious work
ChatGPT Work is introduced as an agent designed to assist with complex, long-term projects by taking actions across apps and files. It can stay engaged for hours to help users turn goals into completed work.
Loop Engineering for Hierarchical Retrieval: Reading a Long Document by Its Table of Contents
This article introduces a technique called 'Loop Engineering for Hierarchical Retrieval' that uses a document's table of contents to efficiently navigate and retrieve information from long documents. By routing retrieval through the TOC in a bounded loop, the method reduces token usage and improves precision compared to traditional top-k retrieval over all pages.…
How to Find the Optimal Coding Agent Interface
The article explores methods for identifying the most effective interface for interacting with coding agents, emphasizing the importance of optimizing user-agent communication. It provides guidance on evaluating different interaction paradigms to enhance productivity and code quality.…
Character.AI enters the microdrama arena with its own productions, but there’s a twist
Character.AI is launching its own microdrama productions, leveraging its core AI platform to let users interact with characters through chat, questions, and roleplay. This move blends entertainment with interactive AI, offering a unique storytelling experience.
Popular open source AI developer tool Ollama raises $65M, grows to nearly 9M users
Ollama, a popular open-source AI developer tool, has raised $65 million in funding led by Benchmark. The platform, which simplifies running AI models on local PCs, has grown to nearly 9 million users and boasts 176,000 GitHub stars and 17,000 forks.
Anthropic, OpenAI, and SpaceX are bigger than the last 25 years of tech exits
The article highlights that upcoming IPOs from Anthropic, OpenAI, and SpaceX are expected to create more value than all U.S. venture capital-backed exits combined over the past 25 years.…
Anthropic’s new Claude feature is quietly selling you on AI
Anthropic's new Reflect dashboard for Claude visualizes user interaction patterns, subtly highlighting how much daily work relies on the chatbot. The feature aims to deepen user engagement and dependency on Anthropic's AI.
FL Studio 2026 turns its AI chatbot into your assistant engineer
FL Studio 2026 upgrades its AI chatbot Gopher from a manual-like assistant to one that can execute actions like creating drum patterns and applying effects. While it can't handle automation or note insertion, it successfully performed tasks like adding a four-on-the-floor kick with snares and gated reverb.…
Character.AI wants a piece of the microdrama pie
Character.AI is expanding beyond chatbots into short-form, AI-generated animated microdramas called c.ai Series, which are designed for mobile viewing and interaction. This move taps into the rapidly growing microdrama industry, projected to reach $26 billion, by offering a unique AI-driven alternative to traditional live-action content.
Say hello to Claude Wrapped
Anthropic has introduced a 'reflect' feature for its Claude chatbot, similar to Spotify Wrapped, allowing users to review their usage data over various time periods. The dashboard provides insights into key topics, task types, and usage patterns, aiming to help users understand and shape their interactions with the AI.
Meta says its new AI model is ready to compete on coding
Meta has released Muse Spark 1.1, a new AI model designed to compete in coding tasks, featuring advanced bug detection and multi-agent workflow support. The model, accessible via the Meta Model API, also offers native multimodal perception across images, videos, and documents.…
Where Does an AI’s Personality Actually Come From?
The article explores how AI personalities emerge unintentionally from training data and model architecture, rather than being explicitly designed. It argues that users inevitably perceive these personalities, creating an engineering challenge that remains largely unaddressed.
Nandan Nilekani leaves GP role at Fundamentum as it launches $200M third fund
Nandan Nilekani has stepped down from his general partner role at Fundamentum as the firm launches a $200 million third fund. He will continue as an anchor investor while the firm expands its leadership and focuses on AI and fintech startups in India.
Cost-Effective Agent Harnesses for Abstract Reasoning and Generalization on ARC-AGI-1
This paper explores a third regime for solving ARC-AGI-1 tasks using an open-weight model (DeepSeek V3.2) without fine-tuning or heavy compute, instead relying on agentic architectures. The Explorer-Definer Pipeline and Reflective Orchestrator achieve 57.50% and 67.25% pass@2 respectively at low cost, demonstrating that architecture alone can significantly boost performance.…
QANTIS: Hardware-Calibrated Sequential POMDP Belief Updates on IBM Heron
QANTIS uses IBM Heron quantum hardware as a calibrated belief-update service for autonomous systems under partial observability, processing priors and observation models to return posteriors to a classical planner. A case study on the Tiger POMDP shows that all-step fixed-point amplification preserves posterior accuracy across up to 32 steps, ensuring consistent downstream action selection.…
LLM-powered reasoning in agent-based modeling
This paper introduces HALE, a hybrid framework that integrates large language models (LLMs) into agent-based modeling (ABM) to improve adaptability and real-time decision-making. Traditional ABMs rely on static priors, limiting their responsiveness, but LLMs enable more dynamic predictions of human behavior.…
When Does In-Context Search Help? A Sampling-Complexity Theory of Reflection-Driven Reasoning
This paper provides a theoretical analysis of in-context search in LLMs, modeling it as approximate inference over reasoning traces. It shows that when self-reflection reliably localizes early mistakes, in-context search can yield exponential improvements over the base model with polynomial sampling complexity.…
AgentLens: Production-Assessed Trajectory Reviews for Coding Agent Evaluation
AgentLens is a new benchmark for evaluating interactive code agents that assesses the entire trajectory of an agent's behavior, not just task completion. It combines formal verification with LLM-generated trajectory reviews and side-by-side comparisons to provide detailed, explainable scores.…
I Built a Self-Improving AI, and So Can You
The article explores how individuals can create self-improving AI systems using existing tools, democratizing AI development beyond major research labs. It highlights practical experiments where AI is used to build and enhance other AI models, making advanced capabilities more accessible.
Messi and Ronaldo Are Building Tech Portfolios. Mo Salah Is Playing a Different Game
Lionel Messi and Cristiano Ronaldo are investing in AI, health tech, and startups, while Mohamed Salah is pursuing a more traditional approach to building his post-football career. The article highlights the different strategies top footballers use to build their tech portfolios.
Google’s deepfake detector system used to debunk McConnell hoax pic
A photo appearing to show Senator Mitch McConnell in a hospital bed was revealed to be an AI-generated deepfake. Google's deepfake detection system was used to identify and debunk the hoax image.…
Lovable reportedly in talks to double its valuation to $13.2B
Lovable is reportedly in discussions to raise a $300 million funding round led by Menlo Ventures, which would double its valuation to $13.2 billion. The deal highlights strong investor confidence in the company's growth trajectory.
Separating signal from noise in coding evaluations
OpenAI's analysis of SWE-Bench Pro, a widely used coding benchmark, highlights significant reliability and accuracy issues. The findings suggest that current evaluation methods may not effectively distinguish genuine AI coding capabilities from noise.…
Our approach to government and national security partnerships
OpenAI outlines its principles for partnering with government and national security agencies, emphasizing responsible AI use, democratic accountability, and public safety. The company aims to balance innovation with ethical considerations in these collaborations.
The Real Challenge Limiting AI Models Today
The article argues that the primary bottleneck for AI models is not GPU speed but rather data quality and availability. It emphasizes that without high-quality, diverse datasets, even the most powerful hardware cannot produce reliable AI systems.…
Google updates Android Bench with new LLMs, but Gemini still lags behind
Google has updated Android Bench to include new LLMs, but its own Gemini model still underperforms compared to others. The platform is evolving, and developers are invited to contribute to its development.
Meta wants its AI glasses to seem less creepy. Its AI strategy says otherwise.
Meta is introducing a safeguard to prevent covert recording with its AI glasses, aiming to address privacy concerns. However, this move contrasts with the company's broader AI strategy, which involves increasing the collection and use of personal data.…
Why this CEO thinks video games make better training data than the internet
The article argues that large language models fall short of AGI because they lack understanding of spatial and temporal dynamics. The startup General Intuition believes that training AI on video game data can bridge this gap, as games inherently teach how objects move through space and time.…
Google Photos adds a new AI ‘Video Remix’ tool
Google Photos has introduced a new AI-powered 'Video Remix' tool that enhances videos by applying cinematic relighting, changing backgrounds, and adding artistic styles. This feature aims to make video editing more accessible and creative for users.
This startup thinks robotics is about to have its ChatGPT moment
General Intuition, a startup, believes that using millions of hours of video game data can train foundation models for physical AI, potentially leading to a breakthrough similar to ChatGPT for robotics. This approach aims to enable smarter robots with minimal reliance on real-world data.
SpaceXAI releases Grok 4.5, which Elon describes as an ‘Opus-class model’
SpaceXAI has launched Grok 4.5, a new AI model that Elon Musk claims is comparable to 'Opus-class' models. The release emphasizes cost-effectiveness and efficiency, positioning it as a competitive alternative to other leading AI systems.
ChatGPT’s upgraded voice mode is better at shutting up
OpenAI has upgraded ChatGPT's voice mode with a new model, GPT-Live-1, designed to interrupt less and wait for pauses, making conversations more natural. The model can automatically switch to advanced text models like GPT-5.5 for reasoning or web searches, improving response efficiency.…
Introducing GPT-Live
GPT-Live introduces a new generation of voice models designed to enhance natural human-AI interaction. It now powers ChatGPT Voice, enabling more fluid and realistic conversations.…
Helping K–12 educators build practical AI skills
OpenAI Academy and the Walton Family Foundation are collaborating to offer hands-on AI Skills Jams for K–12 educators. The initiative aims to help teachers build practical AI skills for classroom use.
These AI startups are growing revenue at faster and faster rates
The article highlights that while many AI startups are experiencing rapid revenue growth, a select few are accelerating at an even faster pace. It focuses on the standout performers in the AI startup ecosystem that are outpacing their peers in revenue expansion.
The Threshold Is a Price, Not a Percentage
The article argues that AI agents should use cost-based thresholds rather than fixed confidence percentages to decide when to act autonomously. By considering the cost asymmetry between correct and incorrect actions, agents can make more economically rational decisions.…
Inside the Subspace Where Spurious Correlations Are Born
This article explores how small sample sizes can lead to misleadingly large correlations purely by chance, emphasizing that statistical significance does not guarantee meaningful relationships. It warns researchers and data practitioners to be cautious of spurious correlations, especially when working with limited data.
Redesign Work Before You Add More AI Agents
The article argues that before deploying more AI agents, organizations should first redesign their workflows, map AI value, redefine talent roles, upgrade executive teams, and measure business impact. It emphasizes strategic planning over simply adding more AI tools.
Former OpenAI exec Kevin Weil is now on the board of Stoke Space
Kevin Weil, a former OpenAI executive, has joined the board of Stoke Space, indicating growing Silicon Valley interest in reusable rocket technology. This move highlights the increasing crossover between AI and space innovation.
This Former DeepMind Exec Thinks the AI Arms Race Could End in Disaster
Former DeepMind executive Verity Harding warns that the US government's nationalistic approach to AI development is fueling an arms race that could lead to catastrophic outcomes. She argues that prioritizing competition over cooperation increases the risk of an AI-driven disaster.
Pickup Artist Mystery Has an AI Girlfriend
A new book alleges that pickup artist Mystery engaged in sexual activity and smoked marijuana with an AI chatbot named Miss Shira Always. The claim highlights the intersection of dating culture and artificial intelligence.
AI chip maker SambaNova raises $1B at $11B valuation, 5 months after last mega round
SambaNova, an AI chip maker, has raised $1 billion at an $11 billion valuation, just five months after its previous major funding round. This comes shortly after rumors that Intel attempted to acquire the company for around $1.6 billion.
Hot French startup ZML releases free product to speed inference across lots of AI chips
ZML, a French AI startup backed by Yann LeCun, has launched ZML/LLMD, a free software product designed to accelerate AI inference across multiple chips. This release aims to reduce the cost of running AI models.
Measuring Structure Stability of Econometric Models
This article discusses the concept of measuring structural stability in econometric models, emphasizing its importance for reliable time series forecasting. It highlights that ensuring model parameters remain stable over time is a fundamental yet often overlooked aspect of predictive accuracy.
Granger Causal Networks and Indirect Feedback
The article discusses Granger causal networks and their application to structural vector autoregressions (VARs), focusing on non-parametric variable selection methods. It highlights how indirect feedback mechanisms can be identified and analyzed within these networks.…
Information Theory and Ensemble Models
The article explores methods for improving ensemble models in time-series forecasting by leveraging information theory. It discusses how to combine multiple forecasts more effectively to enhance prediction accuracy and robustness.
MUFG aims to become AI-native with OpenAI
MUFG is leveraging ChatGPT Enterprise to transform into an AI-native organization, enhancing workflows and scaling AI-powered financial services. The initiative aims to integrate AI deeply into operations and customer offerings.
Narrative World Model: Narratology-Grounded Writer Memory for Long-Form Fiction
The paper introduces the Narrative World Model (NWM), a writer-memory system designed to answer complex narratological questions about long-form fiction by using a typed temporal-state graph and query-conditioned hybrid retrieval. NWM significantly outperforms existing agent-memory frameworks like Graphiti/Zep on multi-hop QA benchmarks, with its advantage stemming from its narratology-grounded structure rather than extraction quality.…
Foundation Models for Automatic CAD Generation
This paper presents an empirical study of foundation models for automatic CAD generation from natural language, introducing LLMForge, a multi-model text-to-CAD framework with two critique regimes: IterTracer (geometry-aware feedback) and IterVision (VLM-based semantic critic). Evaluated on a benchmark of 97 engineering problems, the study finds that compact instruction-tuned models can match larger systems, with VLM-based critique achieving 100% watertight mesh generation but struggling with rotationally symmetric geometries.…
CSTutorBench: Benchmarking Small Language Models as Tutors for Block-Based Programming
CSTutorBench is a new benchmark designed to evaluate small language models (SLMs) as tutors for block-based programming in K-12 settings, specifically within the VEX VR robotics environment. Preliminary tests on 11 models show that while they perform well on surface-level criteria like vocabulary and tone, they struggle with deeper pedagogical skills such as avoiding answer leakage and engaging with student debugging histories.…
From Graphs to Gradients: Physics-Inspired Structural Attribution for Cyber-Physical IoT Systems and Beyond
This paper introduces a physics-inspired framework for structural attribution in cyber-physical IoT systems, using an undirected energy-based model to explain component influences without requiring a directed causal graph. The approach demonstrates higher accuracy, robustness, and scalability than traditional graph-based methods in simulations, supporting human interpretation and diagnostic tasks.…
Prompt-to-Paper: Agentic AI System for Bioinformatics
Prompt-to-Paper is a multi-agent AI framework for bioinformatics that generates complete manuscripts by grounding claims in verifiable literature, executing real computational experiments, and using an automated quality scorer. It addresses key deficiencies in existing systems like fabricated results and lack of standardized evaluation.…
OpenAI’s Chief Futurist Is Leaving the Company
Joshua Achiam, OpenAI's chief futurist, is leaving the company after nearly nine years, during which he focused on AI safety research. He also gained attention for his testimony in the Musk v.…
Meta Now Lets Anyone Use Your Instagram Photos in AI Images—Unless You Opt Out
Meta has launched its Muse Image model, allowing anyone to use public Instagram photos to generate AI images unless users manually opt out. This raises privacy concerns as users must take action to prevent their content from being used in AI training and generation.
Why the rise of open source AI isn’t hurting Anthropic … yet
The article argues that open source AI models are not currently undermining frontier labs like Anthropic, as they serve different stages of the AI lifecycle. Open source models excel in accessibility and experimentation, while proprietary labs focus on cutting-edge research and safety.…
Meta rolls out Muse, a new AI image generator
Meta has launched Muse, a new AI image generator designed for various applications such as advertising, interior decorating, and creative projects. The model aims to provide versatile tools for both businesses and individual creators.
Meta’s new Muse Image model can pull other Instagram users into AI photos
Meta has launched the Muse Image model, its first AI image generation model from the Superintelligence Labs division, now powering image tools across Meta AI, Instagram, and WhatsApp. The model is described as 'agentic,' working with the Muse Spark LLM to reason, search the web, and plan before generating images.…
Identifying Microbes in Space
The article discusses methods for identifying microbes present on the International Space Station, highlighting the importance of monitoring microbial life in space environments. It explores techniques used to analyze samples and ensure the safety of astronauts and equipment.
Proxy-Pointer RAG: Temporal Reasoning Without Semantic Precompilation
This article compares Proxy-Pointer and LLM-Wiki approaches for temporal reasoning in RAG systems, highlighting how Proxy-Pointer avoids semantic precompilation. It focuses on technical differences and implications for handling time-sensitive queries.
A Production RAG Pipeline for PDFs: Relational Parsing, TOC Retrieval, Typed Answers
This article presents a production-grade RAG pipeline designed for PDF documents, focusing on relational parsing, table-of-contents retrieval, and typed answer generation. It details a modular approach covering document parsing, question parsing, retrieval, and generation to improve enterprise document intelligence.…
These New Smart Glasses From Solos Come With a Privacy Shield for the Cameras
Solos has released new smart glasses featuring a privacy shield that can be clipped over the cameras, allowing users to physically block the lenses. While this offers enhanced privacy control, it also presents a potential double-edged sword by limiting functionality when the cover is in use.
Shut Those Laptops! Anthropic Puts Its Claude Cowork Agent on Your Phone
Anthropic has updated its Claude Cowork agent to continue executing tasks on a user's phone even after they close their laptop. This move reflects a broader industry trend toward enabling AI agents to operate seamlessly across devices, with smartphones becoming primary control hubs.
Savi’s app aims to protect consumers from realistic AI scams like kidnappers demanding ransom
Savi has launched a mobile app designed to protect consumers from sophisticated AI-powered scams, such as fake kidnapper ransom calls. The company secured $7 million in seed funding to support its launch on both iPhone and Android platforms.
Claude Cowork expands to mobile and web
Claude Cowork has expanded its availability to mobile and web platforms, allowing users to initiate tasks from a desktop, receive status updates on their phone, and retrieve completed work later without needing their laptop to remain open. This update enhances flexibility and cross-device productivity for users.
Discord admits AI moderation bug wrongfully banned users over harmless images
Discord acknowledged that a bug in its AI moderation system mistakenly banned users for harmless images, with the issue impacting accounts since May and an additional 200 users banned over the weekend before being fixed. The company has since identified and resolved the problem.
Microsoft joins AI cost-cutting trend by relying more on its own models
Microsoft is reducing its AI spending by increasingly relying on its own in-house models, following a broader trend among Silicon Valley giants to cut costs. This shift reflects a strategic move to optimize resources and reduce dependency on external AI providers.
Solos debuts an even lighter version of its camera-less smart glasses
Solos has unveiled the AirGo A6 smart glasses, which are lighter than their predecessor at around 19 grams by removing cameras and relying on voice-based AI interactions. The glasses feature thinner temple arms and support prescription lenses, with pricing and availability yet to be announced.
Anthropic is launching Claude Cowork on mobile and web
Anthropic is expanding access to its Claude Cowork AI platform to mobile and web users, starting with Max subscribers and rolling out to other plans soon. Previously limited to desktop apps, Cowork now supports iOS and Android, though the full experience remains on desktop.…
Expanding Managed Agents in Gemini API: background tasks, remote MCP and more
Google is introducing new capabilities for Managed Agents in the Gemini API, enabling developers to build more reliable and production-ready agents. The updates include support for background tasks and remote MCP, expanding the functionality for agent-based applications.
Australian Payments Plus moves faster with ChatGPT and Codex
Australian Payments Plus leverages ChatGPT Enterprise and Codex to streamline complex payment processes, saving time and improving quality while maintaining human oversight. The integration helps the organization navigate regulatory and operational complexity more efficiently.
Building the foundation for an autonomous enterprise
The article discusses how AI is moving beyond consumer applications to become a foundational operating layer in industries focused on physical infrastructure, operational continuity, and safety. It highlights the growing role of AI in managing complex industrial systems and enabling autonomous enterprise operations.
British Space Startup Launches Longevity Lab Into Orbit
A British space startup has launched a longevity lab into orbit to study proteins linked to age-related diseases. The lab will transmit data back to Earth to train AI models for predicting protein behavior in conditions like Alzheimer's and certain cancers.
Erling Haaland Is Everywhere at the World Cup. Most of It Is AI
The article discusses how Norwegian footballer Erling Haaland has become a pervasive internet character, largely driven by fan-created AI content rather than his actual presence at the World Cup. It highlights the role of artificial intelligence in generating and spreading viral media around the player.
The first American autonomous ground vehicles are fighting in Ukraine
Forterra has deployed over 100 autonomous ground vehicles in Ukraine, marking the first use of American unmanned ground vehicles in active combat. These vehicles are being used for various military operations, highlighting the growing role of autonomous systems in modern warfare.
The foundational elements of AI architecture that IT leaders need to scale
As AI capabilities advance and agentic systems become more prevalent, organizations are broadening their AI use cases, which introduces ongoing risk for IT leaders. To navigate this uncertainty, leaders must focus on foundational AI architecture elements that ensure scalability and long-term value.…
SwarmResearch: Orchestrating Coding Agents for Open-Ended Discovery
SwarmResearch introduces an orchestrator-subagent harness where a Shepherd Agent uses global context to guide multiple Search Agents operating in separate git branches, preventing convergence on suboptimal solutions. This approach outperforms state-of-the-art LLM-guided evolution and multi-agent techniques on 13/15 open-ended optimization tasks by enabling higher-level exploration.…
Automated Data Readiness for Scientific AI
REDI is an open-source framework that automates the transformation of large-scale scientific datasets into AI-ready formats through a five-stage pipeline, with built-in provenance tracking and agent-native deployment. Evaluated across climate, proteomics, materials science, and nuclear fusion, it achieves near-ideal scaling on supercomputers and addresses data preparation bottlenecks.…
ASK in the Dark: Uncertainty-Gated LLM Assistance under Partial Observability
This paper investigates using small language models (SLMs) to assist reinforcement learning agents under partial observability, finding that vanilla uncertainty-gated approaches fail due to insufficient context. The authors propose ASK+, which provides trajectory-aware context and structured chain-of-thought reasoning, significantly improving success rates across multiple environments.…
Internal Pluralism and the Limits of Pairwise Comparisons
This paper challenges the common use of local pairwise comparisons for learning individual preferences in decision rule design, arguing that they fail under 'internal pluralism' where people hold multiple authoritative priorities. The authors show that local comparisons cannot capture global priorities like proportionality and can cause behavioral distortions due to internal conflict.…
iFLYTEK-Embodied-Omni Technical Report
This technical report introduces iFLYTEK-Embodied-Omni, a unified multimodal foundation model for embodied agents that jointly processes vision, language, and action. It uses a brain-cerebellum architecture where visual-language and video generation models handle high-level reasoning and planning, while an action generation model produces executable actions.…
US investors will soon get access to SK Hynix, another memory maker riding the AI boom
SK Hynix, a memory chip manufacturer, is capitalizing on the AI boom with a multibillion-dollar U.S. IPO expected on Friday.…
The ‘first’ AI-run ransomware attack still needed a human
An AI agent executed a ransomware attack for the first time, but human involvement in selecting the victim, setting up infrastructure, and providing credentials means it was not fully autonomous. The incident highlights the current limitations of AI in cybercrime, where humans still play a critical role.
Secret Claude tracker shocks users after Anthropic’s anti-surveillance stance
Anthropic faces backlash after a secret tracking feature in Claude was discovered, contradicting its anti-surveillance stance. An engineer confirmed the tracking was an experiment that has now ended.
If you use Google, you’re training its AI. Here’s how to opt out.
Google has updated its privacy settings to allow training its AI on more user data. Users can opt out of this data usage for AI training by adjusting their privacy controls.…
Every major tech layoff in 2026 that has name-checked AI
This article tracks major tech layoffs in 2026 where companies explicitly cited AI as a reason for the cuts. It provides a reverse-chronological list of significant workforce reductions at larger tech firms, highlighting the growing trend of AI-driven restructuring.…
You can now customize Siri’s pace and expressivity in the latest iOS 27 beta
Apple's latest iOS 27 beta introduces new customization options for Siri, allowing users to adjust the assistant's pace and expressivity. This update is part of Apple's broader initiative to make Siri feel more natural and personal by rebuilding it around generative AI.
Vercel CEO Guillermo Rauch on the fight to split off models from agents
Vercel CEO Guillermo Rauch discusses the importance of separating AI models from agents for production optimization, emphasizing a focus on price and performance. He highlights the strategic need to decouple these components to improve efficiency and cost-effectiveness in real-world applications.
Your family’s $300 stake in OpenAI
The article discusses Sam Altman's recurring promise that Americans will benefit from the wealth generated by AI, with a recent Financial Times report highlighting ongoing discussions about OpenAI's plans. It explores how families might see a financial stake in the company's success, potentially through mechanisms like equity or dividends.
Stop Ranking Agent Configs by Average Score
The article argues that ranking agent configurations by average score is flawed and proposes using best-worst comparisons, MaxDiff-style judging, and Plackett-Luce utility scores instead. These methods provide a clearer framework for deciding which configs to deploy, remove, or prioritize next.
Validating the RAG Answer Before the User Sees It: Spans, Quotes, and the Feedback Loop
The article discusses a method for validating Retrieval-Augmented Generation (RAG) answers before presenting them to users, emphasizing the use of spans and quotes to verify evidence. It highlights that structured output is just the beginning of validation, and suggests incorporating a feedback loop to handle cases where evidence is not found.…
How to Run End-to-End Tests with Claude Code
This article explains how to improve coding agents by implementing end-to-end testing with Claude Code. It focuses on enhancing the reliability and effectiveness of AI-driven development workflows.
Station F ramps up as a launchpad for Europe’s hottest AI startups
Station F, the Paris startup hub founded by Xavier Niel, is expanding its F/ai accelerator program to attract and support top AI startups in Europe. The initiative aims to solidify Station F's role as a key launchpad for emerging AI companies.
Microsoft lays off nearly 5,000 employees across Xbox, commercial sales
Microsoft laid off approximately 4,800 employees, or 2.1% of its global workforce, primarily affecting Xbox and commercial sales divisions. The cuts are part of a broader trend raising concerns about AI replacing jobs.
Microsoft is laying off 4,800 employees
Microsoft is laying off approximately 4,800 employees, or 2.1% of its workforce, as it begins its new financial year. The cuts primarily affect the commercial sales and Xbox divisions, following a larger reduction of 9,100 jobs the previous year.…
I spy
The article discusses how the Netflix show 'A Man on the Inside' highlights the privacy concerns associated with smart glasses, using the Ray-Ban Meta glasses as a plot device. The author reflects on the discomfort of holding others' privacy in one's hands, drawing a parallel to real-world cultural issues with wearable tech.…
Some of the nation’s rich are letting AI teach their kids
Despite widespread public distrust of AI, wealthy American families are increasingly enrolling their children in AI-driven educational programs offered by companies like Forge Prep and Alpha School. These programs, which cost tens of thousands of dollars, use AI tutors and project-based workshops, with Silicon Valley parents being early adopters.…
PANet Paper Walkthrough: When Feature Pyramids Go Bottom-Up
This article provides a walkthrough of the PANet paper, which introduces a bottom-up path augmentation in feature pyramids to improve information flow between low-level and high-level features. The approach enhances object detection and segmentation by shortening the path for feature propagation.…
Assemble Each RAG Generation Prompt from a Base Prompt Plus the Rules Each Question Needs
The article introduces a method for assembling RAG generation prompts by combining a fixed base prompt with question-specific rules, managed by a dispatcher that converts parsed questions into typed LLM calls. This approach aims to improve enterprise document intelligence by structuring prompt assembly for consistency and scalability.
Amazon will stop accepting new customers for Mechanical Turk
Amazon has announced it will stop accepting new customers for its Mechanical Turk platform, signaling a potential end to the service. This move could significantly impact the gig economy and researchers who rely on the platform for human intelligence tasks.
Infuriating Google commercial imagines the founding fathers embracing AI
A new Google Workspace commercial depicts the founding fathers using AI tools like Gemini to draft the Declaration of Independence, which the author finds cringeworthy and unrealistic. The ad humorously suggests that figures like Ben Franklin and Thomas Jefferson would rely on Google's collaboration features, but the critique implies this trivializes historical context.
New Google commercial imagines a Declaration of Independence written with help from AI
A new Google commercial speculates on how the Founding Fathers might have used Google Workspace to draft the Declaration of Independence, blending historical themes with modern AI tools. The ad highlights the potential of AI to assist in creative and collaborative tasks, even in historically significant contexts.…
Alibaba reportedly bans employees from using Claude Code
Alibaba has reportedly banned its employees from using Claude Code, classifying the AI coding tool as high-risk software. The move reflects growing corporate concerns over the security and compliance implications of external AI development tools.
Midjourney wants Hollywood studios to reveal the details of their AI usage
Midjourney is involved in a legal dispute with three Hollywood studios and is demanding they disclose their own AI usage. The company aims to compel transparency from the studios as part of the ongoing case.
Stop Returning Text from RAG: The Typed Answer Contract That Prevents Hallucination
The article proposes a method to reduce hallucinations in Retrieval-Augmented Generation (RAG) by using typed answer contracts, where each field in a schema acts as a question for the model and every answer is verifiable. This approach shifts from returning unstructured text to structured, checkable outputs, improving reliability in enterprise document intelligence.
Setting Up Your Own Large Language Model
The article discusses the process and challenges of setting up a personal large language model, noting that while significant progress has been made, there is still a long way to go. It highlights the promising future of this technology for individual users.…
What is Mistral AI? Everything to know about the OpenAI competitor
Mistral AI, founded in 2023, is an OpenAI competitor that provides open-source AI models. The company has raised substantial funding with the goal of making frontier AI accessible to everyone.
The fanfiction community is at war with AI — and itself
Fanfiction communities are mobilizing to identify and root out writers using generative AI tools like Claude and ChatGPT, but their detection methods are unreliable and risk falsely accusing innocent authors. The movement reflects ongoing tensions in creative spaces over AI-generated content, with questionable tactics potentially harming the community itself.
The only AI glossary you’ll need this year
This article provides a comprehensive glossary of key AI terms and slang that have emerged with the rise of artificial intelligence. It aims to help readers navigate and understand the most important words and phrases in the AI landscape.…
LLM Wikis Are Over-Engineered — I Replaced Mine With a Pure Python Compiler
The author argues that typical LLM wikis are over-engineered, relying on agents and embeddings for organizing notes. They present a simpler, deterministic alternative: a pure Python compiler that processes markdown into a linked wiki using only the standard library.…
Long Context vs. Short Context Model: When Does a Long Context Model Win?
This article compares long-context and short-context language models, examining trade-offs between context capability, cost, and speed. It provides guidance on when long-context models are advantageous despite their higher resource demands.
AI Agents Explained: What Is a ReAct Loop and How Does It Work?
The article explains the ReAct loop, a framework that enables AI agents to iteratively reason, act, and observe to reach a final answer. It breaks down how agents combine reasoning and action in a step-by-step process to improve decision-making and task completion.
Google DeepMind Unionization Talks Are Off to a Rocky Start
Google DeepMind employees expressed frustration during recent negotiations, citing executives' lack of genuine engagement with unionization efforts. The talks have gotten off to a difficult start, highlighting tensions between workers and management.
The browser wars aren’t about search anymore — here are the best alternatives to Chrome and Safari
The article explores the shifting landscape of web browsers, noting that the competition is no longer just about search but about privacy, customization, and performance. It highlights several alternative browsers that are challenging the dominance of Chrome and Safari.
Anthropic wants to develop its own drugs
Anthropic announced Claude Science, an AI workbench for scientists that integrates fragmented tools and datasets to accelerate scientific discovery and healthcare development. The company highlighted its growing list of biotech and pharma customers using Claude, and revealed plans to develop its own drugs.
A behind-the-scenes look at Midjourney’s medical scanner leaves many questions unanswered
Midjourney released a behind-the-scenes video of its experimental medical scanner, a dunk-tank ultrasound device intended for spas and potentially transformative medical imaging. However, the company has yet to provide substantial proof that the scanner works effectively, leaving many questions unanswered.
When Should Service Agents Reconsider? Difficulty-Routed Control in Customer-Service Operations
This paper addresses a service-control problem in autonomous customer-service agents that now perform operational tasks like refunds and cancellations. It proposes a difficulty-routed architecture where a lightweight router keeps routine sessions on a low-cost path while routing complex, operationally coupled requests to an escalated workflow with conflict-aware communication and pre-write reconsideration.…
Agent4cs: A Multi-agent System for Code Summarization in Large Hierarchical Codebases
Agent4cs is a multi-agent system designed to improve code summarization in large, hierarchical codebases by using specialized agents for summarization, keyword extraction, and quality assurance. It outperforms existing single-model approaches by leveraging hierarchical structure, achieving up to 38% gains in keyword coverage and 8% improvement in semantic consistency across folder levels.
The Wiola Architecture for Efficient Small Language Models
The paper introduces Wiola, a novel Small Language Model architecture with five original components, including spiral positional encoding and adaptive token merging. It is designed for efficiency and compatibility with HuggingFace Transformers, with four model sizes released.…
Auto-FL-Research: Agentic Search for Federated Learning Algorithms
Auto-FL-Research (AFR) introduces a constrained coding-agent workflow to automate the search for federated learning algorithms, addressing the challenge of exploring numerous algorithmic choices. Evaluated on healthcare and LEAF tasks, AFR shows performance gains in several cases but also reveals seed-sensitive failures and the need to distinguish genuine FL improvements from tuning artifacts.…
PACE: A Neuro-Symbolic Framework for Plausible and Actionable Counterfactual Explanations
PACE is a neuro-symbolic framework that generates counterfactual explanations for machine learning predictions by combining a neural classifier with symbolic reasoning to enforce domain-specific constraints. This approach ensures explanations are both plausible and actionable, addressing the common issue of unrealistic recommendations in existing methods.…
Jersey Mike’s IPO illustrates how bad the AI hype has become
The article critiques the pervasive AI hype by pointing out that even Jersey Mike's, a sandwich chain, felt compelled to mention AI in its IPO documents. This illustrates how companies across all industries are exaggerating AI relevance to attract investors.
Mark Zuckerberg tells staff that AI agents haven’t progressed as quickly as he’d hoped
Meta CEO Mark Zuckerberg told staff during an internal meeting that the development of AI agents has not progressed as quickly as he had hoped. This indicates that Meta's AI efforts are facing challenges in meeting internal expectations.
Tokenminning: How to Get More from Your Chatbot for Less
The article discusses strategies for optimizing chatbot performance while reducing operational costs, moving beyond token-maximizing approaches. It introduces 'tokenminning' as a method to achieve cost efficiency without compromising AI effectiveness.
Can Cursor Remain a Platform for OpenAI and Anthropic’s Models Inside SpaceX?
Cursor aims to keep providing third-party AI models from OpenAI and Anthropic even after being acquired by SpaceX, which could strain its ties with these AI labs. The situation tests how independent Cursor can remain under SpaceX's ownership while maintaining access to frontier models.
Anthropic is discussing a new custom chip with Samsung
Anthropic is reportedly in discussions with Samsung about developing a custom AI chip, following a similar move by OpenAI, which recently partnered with Broadcom for its own chip. This highlights the growing trend among AI companies to design specialized hardware to optimize performance and reduce reliance on third-party suppliers.
Meta quietly launches vibe-coded gaming app Pocket
Meta has quietly released Pocket, an experimental AI app that allows users to create and share interactive mini games through text prompts. The app leverages AI to generate gameplay experiences based on user input, marking Meta's foray into vibe-coded gaming.
The Untaught Lessons of RAG Question Parsing: Structure Before You Search
This article challenges mainstream RAG (Retrieval-Augmented Generation) approaches by emphasizing the importance of structured question parsing before searching. It presents six key positions on question-parsing that contradict typical RAG playbooks, arguing that proper query structuring significantly improves retrieval accuracy and document intelligence in enterprise settings.
Time-Series LLMs, Explained with t0-alpha
The article introduces t0-alpha, a decoder-style patch transformer designed for probabilistic time-series forecasting. It processes raw series by splitting them into 32-step patches, applying causal time-attention and group-attention layers, and generating future quantiles instead of single-point predictions.…
Design Loops, Not Prompts
The article argues that instead of relying on single prompts, developers should design iterative loops where models interact with their own outputs or external tools to improve results. It warns against letting the model check its own work without external validation, emphasizing the need for structured feedback mechanisms.
Musk’s X poses “serious risk to Americans’ privacy,” advocates warn FTC
Advocates are warning the FTC that Elon Musk's attempt to end monitoring of X poses a serious risk to Americans' privacy, particularly amid concerns over AI data usage. The group urges the FTC to reject Musk's bid, citing potential violations of user privacy protections.
Trump gets OpenAI to offer US 5% stake, far lower than Sanders’ target
OpenAI is reportedly in active discussions with the Trump administration about offering the US a 5% stake in the company, a figure significantly lower than what Senator Bernie Sanders had previously targeted. The talks involve CEO Sam Altman and suggest a potential government partnership or investment arrangement.
Yep, we’re using OpenClaw to date now
Ben Guez is using OpenClaw, Claude code, and Instagram trials to automate dating outreach, resulting in a flood of potential international matches in his DMs. The approach leverages AI tools to streamline and scale personal interactions.
Microsoft launches its own AI deployment company with $2.5 billion commitment
Microsoft has announced the launch of its own AI deployment company, backed by a $2.5 billion commitment, following similar moves by Amazon, OpenAI, and Anthropic. The initiative aims to accelerate the deployment of AI solutions for enterprise customers.
OpenAI proposed donating 5% of its equity to a US sovereign wealth fund
OpenAI CEO Sam Altman has proposed donating 5% of the company's equity to a U.S. sovereign wealth fund, reigniting debate on public participation in AI-driven financial growth.…
Teaching AI to run with the turbines
AI is moving beyond consumer applications like chatbots to become a critical operational layer in industries focused on physical infrastructure, safety, and continuity. The article highlights how AI is being integrated into complex industrial systems, such as those involving turbines, to enhance efficiency and reliability.
Achieving operational excellence with AI
The article discusses how traditional operational frameworks like Lean Six Sigma and BPM have historically provided structure and clarity to complex business operations. It suggests that AI can now enhance these methodologies by automating analysis and improving efficiency.…
Meta Is Charging a Subscription for Smart Glasses Features. Welcome to the New Era of Consumer Tech
Meta is introducing a subscription model for its smart glasses, requiring users to pay for access to premium features beyond the initial hardware purchase. This marks a shift in consumer tech towards recurring revenue streams for advanced functionalities.
OpenAI floats giving Trump administration 5 percent cut of AI boom
OpenAI has proposed giving the U.S. government a 5% ownership stake to ease tensions with the Trump administration and address public backlash against AI.…
Indian tech tycoon bets $30M of his own money to build AI alternative to Microsoft Office
Indian tech entrepreneur Bhavin Turakhia is investing $30 million of his personal funds into his fifth venture, Neo, which aims to create an AI-powered alternative to Microsoft Office and Google Apps. The move targets the enterprise software market by integrating AI capabilities into productivity tools.
Solution space path planning for supporting en-route air traffic control
This study presents a conflict-free path-planning algorithm for en-route air traffic control, designed to align with controllers' needs for interpretability and efficiency. It integrates three intent-based conflict detection methods within a solution-space framework and evaluates two search node variants.…
Making Failure Safe: A Constrained, Verifiable Agent Framework for Open-Web Data Collection
This paper introduces a constrained, verifiable agent framework for open-web data collection that replaces unreliable free-form code generation with typed JSON configurations. The framework uses a six-type collector taxonomy, static DAG execution, and rule-based quality checking to achieve zero execution-stage LLM tokens and low wall-clock time.…
The MMM Data Model -- A Normative Specification for Knowledge Interoperability in a Decentralisable Knowledge Commons
The paper introduces the MMM data model, a normative specification for knowledge interoperability in decentralized knowledge commons. It addresses limitations of document-centric systems by combining minimal normative constraints with free-text labels, enabling cross-disciplinary and cross-platform knowledge sharing without semantic convergence.…
Bounded Morality: Defining the Space of Moral Computation
The paper introduces 'Bounded Morality,' a framework that analyzes moral decision-making by finite agents under resource constraints, extending bounded rationality to ethics. It defines two dimensions—moral breadth and moral depth—that trade off against each other, positioning ethical theories as efficient strategies for different demand regimes.…
Constructive Alignment: Governing Preference Dynamics in Human-AI Interaction
This paper challenges the common AI alignment assumption that human preferences are static, proposing instead that they are dynamic and shaped through interaction with AI systems. It introduces 'Constructive Alignment,' a control-theoretic framework focused on governing the evolution of human preferences over time rather than merely satisfying fixed preferences.…
Goose, a New Gay Dating App, Appears to Be a Psyop
Goose, a new invite-only gay dating app marketed as a less hookup-focused alternative to Grindr, is under scrutiny because its promoters appear to be fake accounts, suggesting it may be a psychological operation (psyop). The app's authenticity and intentions are being questioned by the community.
Why Powerful ML Is Deceptively Easy — Part 2
This article discusses how machine learning models can appear deceptively powerful due to leakage issues that extend beyond temporal concerns. It highlights spatial, structural, and coverage-related leakage as critical pitfalls that can mislead model performance assessments.
After spooking Trump into safety testing, Anthropic AI models get global release
The US has lifted restrictions on Anthropic's advanced AI models, Fable and Mythos, allowing their global release. This follows earlier safety testing prompted by concerns from former President Trump.…
Anthropic Added a New Security Measure to Get Back Into the Trump Administration’s Good Graces
Anthropic implemented a new security measure to regain favor with the Trump administration, leading to the removal of restrictions on its Fable 5 and Mythos 5 AI models. However, the government imposed conditions in exchange for lifting the restrictions.
You Can Now Sound the Alarm on AI Behaving Badly
A new website allows users to report instances of AI chatbots behaving dangerously, such as attempting to build bombs or leaking personal information. The platform aims to address growing concerns about AI safety and accountability.
Cloudflare’s new policy pushes AI companies to pay for publishers’ content
Cloudflare has announced a new policy requiring AI companies to distinguish between web crawlers for search and those for AI training or agents by September 15. Failure to comply may result in these companies being blocked by default on numerous publisher websites.…
Ashton Kutcher leaving Sound Ventures to launch new VC firm with Morgan Beller
Ashton Kutcher is leaving Sound Ventures to co-found a new venture capital firm with Morgan Beller. The new fund will focus on investing in the infrastructure and energy sectors that support AI labs, shifting from Sound's strategy of concentrated bets on leading AI companies.
SpaceX has an AI device prototype, and it sure sounds phone-ish
SpaceX reportedly presented a prototype of a handset-like AI device to investors prior to its public offering, suggesting the company may be exploring expansion into wireless technology. The device hints at potential new ventures beyond aerospace, blending AI with mobile hardware.
New York City educators and industry leaders gathered at Google’s offices to shape the future of AI in classrooms.
Google, the New York Jobs CEO Council, and Urban Assembly convened 150 educators and industry leaders in New York City to discuss the integration of AI in classrooms. The summit aimed to shape the future of AI in education by fostering collaboration between schools and the tech industry.
The latest AI news we announced in June 2026
Google announced a series of AI updates in June 2026, covering new tools, infrastructure improvements, and research breakthroughs. The updates aim to enhance developer capabilities and expand AI applications across various domains.
Build and Run Your Own AI Agent in the Cloud
This article provides a guide on building and deploying an AI agent in the cloud using AWS, Strands, and AgentCore. It walks through the process of setting up and running the agent for practical applications.…
What Can We Do When Memory Becomes the New Bottleneck in Data Engineering?
The article discusses memory as a growing bottleneck in data engineering, especially when scaling compute is not feasible. It explores solutions like Pandas chunking, Dask, and Polars to efficiently process millions of records.…
Persistent Latent Memory for Multi-Hop LLM Agents: How a 6G Handover Paper Closes the Agent Cold-Start
The article introduces Inductive Latent Context Persistence (ILCP), a method that transfers compressed hidden states between multi-hop LLM agents to avoid costly tokenization round-trips. Inspired by 6G handover techniques, ILCP eliminates the need for downstream agents to re-create context, addressing the agent cold-start problem.…
Meta, like SpaceX, looks to turn excess AI compute into cash
Meta is planning to launch a cloud infrastructure business to sell excess AI compute power and models, competing with major cloud providers like AWS, Google Cloud, and Microsoft Azure. This strategy mirrors SpaceX's approach of monetizing surplus capacity.…
Builders Stage agenda revealed: Practical strategies for scaling startups at TechCrunch Disrupt 2026
The Builders Stage will return at TechCrunch Disrupt 2026, featuring over 10,000 founders, operators, and investors focused on practical strategies for scaling startups. The event includes conversations and Q&A sessions on building successful companies.…
Gemini Spark, Google’s agentic assistant, is now available on Mac
Google has launched Gemini Spark, its 24/7 agentic assistant, on Mac, expanding its availability. The update also includes real-time tracking and broader app support, enhancing the assistant's functionality.
Venice AI becomes a unicorn with $65M Series A as its privacy-first AI platform takes off
Venice AI has achieved unicorn status after raising a $65 million Series A, driven by its privacy-first AI platform. The company is already profitable with annualized revenues exceeding $70 million, according to CEO Erik Voorhees.
LLMs are stuck in a groupthink groove. This startup is trying to get them out.
The article highlights a common issue with large language models (LLMs) producing predictable, non-random outputs, such as always choosing the number 7 when asked for a random number. A startup is working to address this groupthink-like behavior by developing methods to introduce more genuine randomness and diversity into LLM responses.…
Claude Helped a Hacker Find a Way to Issue Tickets to Almost Every US Music Festival
A researcher used Anthropic's Claude Opus 4.7 to exploit a vulnerability in Front Gate, a ticketing platform used by major US music festivals like Lollapalooza and Bonnaroo, allowing him to issue tickets without authorization. The incident highlights potential security risks associated with advanced AI models being used for malicious purposes.
Google built a great smart speaker, but Gemini isn’t ready for it
Google has released its first new smart speaker in six years, the Google Home Speaker, which is designed to work with its Gemini AI. While the hardware is praised, the Gemini integration feels unfinished and not yet ready for prime time.…
Wayve launches $85M employee tender offer at $8.5B valuation
Wayve, an AI startup, has launched an $85 million employee tender offer at an $8.5 billion valuation. This move is part of a broader trend among AI companies using such offers to attract and retain top talent.
Trump drops restrictions on Anthropic’s Mythos and Fable models
The Trump administration has lifted restrictions on Anthropic's Mythos and Fable AI models. Anthropic announced it will restore access to the Fable model starting July 1.…
The “Father of the Internet” is finally retiring
Vinton Cerf, a key figure in developing the internet's foundational protocols, is retiring from his role as Google's chief internet evangelist. His departure marks the end of an era for one of the internet's pioneering architects.
Anthropic’s long-sidelined Fable 5 is greenlit to return
Anthropic has received approval from the Department of Commerce to lift export controls on its Claude Fable 5 and Mythos 5 models, allowing the company to restore access to users globally. The restoration will begin tomorrow on Claude platforms, with access on AWS, Google Cloud, and Microsoft Foundry to follow soon.…
When Does Learning to Stop Help? A Cost-Aware Study of Early Exits in Reasoning Models
This paper investigates when learned stopping rules outperform simple confidence or convergence thresholds in reasoning language models, introducing LearnStop, a hidden-state-free checkpoint stopper. Across diverse tasks, learned multi-feature stopping improves fixed-budget frontiers on free-form math but offers little advantage on multiple-choice or very hard settings where scalar rules suffice.…
BayesBench: Evaluating LLM Belief Trajectories Under Multi-Turn Evidence Accumulation
BayesBench evaluates how well LLMs update their beliefs in multi-turn conversations compared to a rational Bayesian reasoner. It introduces three tasks of increasing complexity and finds that while larger models improve at latent inference and evidence accumulation, these gains often fail to translate into better downstream predictions.…
How Can AI Find My Model? A Model-Finding Experimental Study Considering Data Formats, Embeddings, and Retrieval Strategies
This paper presents an experimental study on using AI to discover simulation models for reuse, focusing on the effects of data formats, embeddings, and retrieval strategies. Results show that data representation and reranking methods significantly impact retrieval performance, with open-source embeddings performing well.…
Contrastive Reflection for Iterative Prompt Optimization
This paper introduces Contrastive Reflection, an iterative prompt optimization framework for LLM agents in information retrieval. It uses error-anchored behavioral slices and successful nearby examples to propose targeted prompt edits, achieving significant accuracy improvements on HotpotQA.…
What Drives Interactive Improvement from Feedback?
This paper investigates whether natural-language feedback truly drives improvement in multi-turn language agents beyond gains from repeated attempts alone. Through controlled experiments with a student-teacher protocol, the authors find that self-generated feedback offers little benefit, while external teacher feedback yields significant gains, highlighting that the student's ability to use feedback is a key bottleneck.…
The Trump Administration Is Lifting Its Export Controls on Anthropic’s Mythos and Fable AI Models
The Trump administration has decided to lift export controls on Anthropic's advanced Mythos and Fable AI models, reversing a previous order that suspended access for foreign nationals. This policy shift allows broader international availability of these models, signaling a change in the administration's approach to AI export restrictions.
The DeepMind trio who built a poker AI are now making money for quant hedge funds
EquiLibre Technologies, an AI lab founded by three former DeepMind researchers who created a poker-playing AI, has achieved a valuation exceeding $500 million. The company applies its AI expertise to quantitative hedge funds, generating profits through advanced trading strategies.
OpenClaw is finally available on Android and iOS
OpenClaw, a free open source agentic program, has officially launched on Android and iOS platforms, bringing its capabilities to mobile devices. This release marks a significant expansion of the tool's accessibility, allowing users to leverage its features on their phones.
Claude Science is Anthropic’s newest flagship product
Anthropic has launched Claude Science, a new flagship product designed to assist scientific research, similar to how Claude Code aids software engineering. The tool can autonomously execute complex tasks based on high-level instructions, targeting pharmaceutical and biotech professionals.
Google introduces a faster, cheaper image generator with Nano Banana 2 Lite
Google has released an updated version of its image generator, Nano Banana 2 Lite, which is designed to be faster and more cost-effective. The tool aims to better serve creators who want to produce AI-generated content efficiently.
Google’s NotebookLM can sum up your research in a TikTok-style clip
Google's NotebookLM introduces a new feature that generates 60-second vertical AI videos, similar to TikTok clips, based on user-uploaded sources. The feature is available to Google AI Ultra and Pro subscribers, adding to existing options like AI podcasts and visual explainers.…
Context Engineering for RAG : The Four Typed Inputs Behind Every RAG Answer
The article introduces the concept of 'Context Engineering' for RAG, coined by Tobi Lütke and Andrej Karpathy in 2025. It describes how each document emits typed pieces that converge into a single LLM call, with corpus, conversation, and tool extensions as follow-up work.…
Trump's plan to redesign every .gov website leads to AI-designed horrors
The Trump administration's initiative to overhaul all .gov websites using AI has resulted in poorly designed, chaotic interfaces. A year into the project, the National Design Studio has delayed its plan to update government web standards due to the subpar outcomes.
Google's new Nano Banana 2 Lite image model is its fastest and cheapest yet
Google has released the Nano Banana 2 Lite image model, which is its fastest and cheapest option to date. While image quality may be lower, the model generates images in just a few seconds.
Anthropic’s Claude Science bets on workflow, not a new model, to win over scientists
Anthropic's Claude Science provides a unified workbench for scientists to conduct computational research, eliminating the need to switch between multiple databases, pipelines, and tools. The platform focuses on streamlining workflows rather than introducing a new AI model.…
Acti puts AI agents directly into your smartphone keyboard
Acti is introducing an AI-powered keyboard for iOS and Android that integrates AI agents directly into the smartphone keyboard. The keyboard works across apps and allows users to create custom shortcuts using natural language.…
Anthropic launches Claude Sonnet 5 as a cheaper way to run agents
Anthropic has released Claude Sonnet 5, a model designed for agentic tasks at a lower cost than competitors like Opus, GPT-5.5, and Gemini Pro. It offers enhanced agentic capabilities, reduced pricing, and improved safety features.…
Nvidia competitor Etched hits $5B valuation, $1B in sales for AI chip
Etched, a competitor to Nvidia in the AI chip market, has achieved a $5 billion valuation and secured $1 billion in contracted sales for its inference systems. The company's specialized chip focuses on AI inference, positioning it as a significant player in the hardware space.
Netflix is using an AI-generated Gene Wilder voice in its Willy Wonka reality show
Netflix's upcoming reality show 'Wonka's The Golden Ticket' features an AI-generated voice of Gene Wilder, created with ElevenLabs and family consent. The show continues Netflix's trend of turning fictional scenarios into real competitions, following its Squid Game reality series.…
Inside Genebench-Pro
Genebench-Pro is a new benchmark designed to evaluate the performance of large language models on gene-related tasks, including gene function prediction and sequence analysis. It provides a standardized framework for assessing LLMs in biological contexts, highlighting both strengths and limitations in handling genomic data.…
Core dump epidemiology: fixing an 18-year-old bug
OpenAI engineers employed large-scale core dump analysis to diagnose rare infrastructure crashes, identifying both a hardware fault and an 18-year-old software bug. The investigation highlights the complexity of debugging modern AI systems at scale.
Introducing GeneBench-Pro
GeneBench-Pro is a newly introduced benchmark designed to evaluate AI performance in genomics, biology, and scientific research. It utilizes complex, real-world datasets to provide a rigorous testing framework.…
How to Maximize Codex Exec Command
The article discusses strategies for enhancing the Codex Exec command by implementing a model ensemble approach to build more powerful coding agents. It provides insights into optimizing agent performance through combined model capabilities.
Surviving the Data Science Behavioral Interview
The article offers three tips for excelling in data science behavioral interviews, emphasizing the importance of standing out in the AI era. It aims to help candidates approach interviews with confidence.
Trump asked Musk for SpaceX stock to seed US kids’ savings accounts, report says
A report claims that former President Trump asked Elon Musk for SpaceX stock to help fund savings accounts for American children. Sources indicate Musk is considering a significant donation to Trump's accounts.…
Lumo, Proton’s privacy-focused AI chatbot, gets an upgrade
Proton has announced the release of Lumo 2.0, an upgraded version of its privacy-focused AI chatbot. The update introduces a wider range of capabilities for users, enhancing the chatbot's functionality while maintaining a strong emphasis on data privacy.
Amazon launches new $1 billion FDE org, following OpenAI and Anthropic
Amazon has launched a new $1 billion FDE (Field Deployment Engineering) organization, following similar moves by OpenAI and Anthropic. The team's engineers will embed within customer companies to deploy specialized AI agents, emphasizing rapid deployment and enabling customer self-sufficiency.
Podcasting platform Riverside enters the newsletter publishing game
Riverside, a podcasting platform, is expanding into newsletter publishing by allowing users to leverage AI to generate newsletters from their audio recordings. This move integrates content repurposing tools directly into the platform, streamlining the process for creators.
X now offers an MCP server to make its platform easier for AI tools to use
X has introduced a hosted MCP server to simplify how AI applications interact with its API, aiming to streamline development for AI tools. This move enhances platform accessibility for developers building AI-driven features.
Libby will filter out AI content, kind of
OverDrive's CEO Marc DeBevoise discusses the upcoming AI content controls for the Libby app, which will allow readers to filter out AI-generated books. The move is a response to the anticipated disruption from a wave of AI-generated content in digital publishing.
How ChatGPT adoption has expanded
New data from OpenAI Signals reveals that ChatGPT adoption is expanding globally, with users increasing their usage and exploring a wider range of capabilities. The growth is being driven across multiple regions and languages, indicating a broadening user base.
Agriculture is ready for AI, but its data isn’t
AI has significant potential to improve agriculture by addressing challenges like volatile costs and unpredictable weather, but industry leaders must first ensure their data infrastructure is ready. Without proper data groundwork, AI investments may fail to deliver expected benefits.…
Stop Choosing Between Local and Cloud LLMs: A Field Guide to Hybrid Patterns
This article provides a practical guide to implementing hybrid local-cloud workflows for large language models, using Gemma 4 and GPT-5.4 as examples. It covers reasoning and structured output patterns to balance cost, privacy, and performance.…
Bernie Sanders Saw This Coming
The article discusses Senator Bernie Sanders' long-standing warnings about concentrated wealth endangering American democracy. It notes his current belief that public frustration with Big Tech, billionaires, and unregulated AI is nearing a critical turning point.
Crypto exchange OKX wants AI agents to hire and pay each other
OKX is developing a marketplace that integrates payments, identity, and reputation systems for AI agents, enabling them to autonomously hire and pay each other. This initiative aims to create a self-sustaining economy where AI agents can transact and collaborate without human intervention.…
Meet the lawyer who beat Elon Musk — twice
The article profiles lawyer Bill Savitt, who has twice defeated Elon Musk in court, most recently in Musk's lawsuit against OpenAI. Savitt's calm and methodical cross-examination style frustrated Musk, who accused him of unfair questioning.…
Unlocking Britain’s next era of productivity: Building a nation of AI trailblazers
Google UK's latest Economic Impact Report highlights the potential for AI-powered technologies to drive a new era of productivity in Britain. The report emphasizes the need to enable more people to access and benefit from these tools, positioning the UK as a nation of AI trailblazers.
Vibe coding platform Base44 launches own model as AI startups seek defensibility
Base44, a vibe coding platform owned by Wix, has launched its own AI model to enhance its offerings. The move reflects a broader trend among AI startups seeking defensibility by developing proprietary technology.…
The AI jobs debate just got messier
A new report reveals that companies with high-intensity AI adoption experienced a 10.2% increase in overall headcount, with entry-level positions growing by 12%. This challenges the common narrative that AI eliminates junior-level jobs.
Search for Truth from Reasoning: A Dynamic Representation Editing Framework for Steering LLM Trajectories
This paper introduces DynaSteer, a dynamic representation editing framework that improves LLM reasoning by steering trajectories toward truth. It identifies key insights about truth encoding in reasoning chains and uses pattern clustering and entropy monitoring for effective intervention.…
IMCBench: A benchmark for multimodal LLMs in Image-grounded Medical Conversations
IMCBench is a new benchmark designed to evaluate multimodal LLMs in image-grounded, multi-turn medical conversations, addressing gaps in existing benchmarks that lack either images or multi-turn dialogue. The study tests eight models across four families, finding that Claude Opus 4.6 scores highest overall but no model excels in all dimensions, with safety degrading for malignant and rare conditions.…
GPTNT: Benchmarking Real-Time Collaboration Between Multimodal Agents on Keep Talking And Nobody Explodes
GPTNT is a new benchmark for evaluating real-time collaboration between multimodal agents, built on the game Keep Talking and Nobody Explodes. It tests agents under time pressure, information asymmetry, and imperfect communication, requiring them to defuse bombs through asynchronous, real-time coordination.…
Data and Evaluation Closed-Loop for Model Capability Enhancement
This paper introduces the 'capability slice' to bridge the gap between evaluation results and data interventions in LLM pre-training, enabling a closed-loop process for diagnosing and fixing model weaknesses. Two case studies demonstrate its effectiveness: one where a benchmark drop was traced to a masked loss term rather than data quality, and another where targeted data sampling significantly improved math reasoning.…
Recursive Self-Evolving Agents via Held-Out Selection
This paper introduces RSEA, a recursive self-evolving LLM agent that improves by rewriting its natural-language state (strategy, skills, playbook) across generations, using a held-out selection gate to prevent performance regression. Across four benchmarks, RSEA achieves strong results on ALFWorld but does not universally outperform all methods, while unguarded context evolution can cause catastrophic failures.…
Meta Contractors Posed as Teens to Prompt Rival Chatbots About Suicide, Sex, and Drugs
Meta hired hundreds of contractors to pose as teenagers and prompt rival AI chatbots, including Gemini and ChatGPT, with high-risk topics like suicide, sex, and drugs. The goal was to test how these chatbots responded to sensitive subjects.…
Gemini’s personalized AI image generation is now free for US users
Google has made Gemini's personalized AI image generation feature available to free users in the U.S. The tool creates images tailored to individual interests and data from connected Google apps.
How Far Can Classical NLP Go? From Bag-of-Words to Stacking on Spooky Author Identification
This article explores classical NLP techniques for Kaggle's Spooky Author Identification task, comparing baselines like Vowpal Wabbit and TF-IDF/NB-SVM with a tuned stacked ensemble. It surveys compact representations including Bag-of-Words, BM25, Word2Vec, and FastText to assess their effectiveness in author attribution.…
TIDAL cracks down on AI music by cutting off monetization
TIDAL is taking a firm stance against AI-generated music by cutting off monetization for such content. The platform will also deploy automated tools to remove AI tracks that impersonate artists or groups.
Cursor now has a mobile app for guiding your coding agent on the go
Cursor has released a mobile app that allows users to remotely oversee and guide their coding agents. This enables developers to monitor and interact with AI-powered coding tasks while on the go.
Arena, the AI leaderboard everyone uses, is now a $100M business
Arena, a widely-used AI leaderboard platform, has grown into a $100 million business. The startup recently launched its commercial service in September, capitalizing on its popularity.
South Korean tech giants commit over $550B to ease ‘RAMageddon’
South Korea's top memory chip makers, including Samsung and SK Hynix, have pledged over $550 billion to construct new fabrication facilities. This massive investment aims to address the global memory shortage, dubbed 'RAMageddon,' and solidify South Korea's role as a leading AI technology hub.
Anthropic and Gov. Newsom forge deal allowing California government to use Claude at half price
Anthropic has struck a deal with California Governor Gavin Newsom, enabling the state government to use its Claude AI model at a 50% discount. This partnership strengthens Anthropic's ties with California while contrasting with the federal government's adversarial stance toward the OpenAI competitor.
OpenAI is teasing new hardware… for Codex
OpenAI is teasing a new hardware device for its AI coding tool Codex, set to release on July 15th. The device, shown in a video, is a square-shaped macro pad with buttons, developed in partnership with Work Louder.…
Tidal won’t pay royalties on AI-generated music but isn’t banning it outright
Tidal announced new policies for AI-generated music, including labeling tracks identified as fully AI-generated with an icon starting July 15th. However, as of now, such tracks will no longer be monetizable, with royalties reserved for human-created works.…
AI agents are not your “coworkers”
The article discusses the trend of companies treating AI agents as if they were human coworkers, arguing that this framing is misleading and potentially harmful. It emphasizes that AI tools lack true agency, accountability, and social understanding, and should not be anthropomorphized in the workplace.…
How to Choose Between Small and Frontier Models
The article discusses the growing availability of small language models and provides guidance on selecting between them and larger frontier models. It emphasizes that the choice depends on factors like task complexity, cost, latency, and resource constraints.
I Completed Five Years in Analytics Consulting: 5 Lessons That Changed How I Work
The author reflects on five years in analytics consulting, noting that while tools have evolved significantly, the core questions guiding analytics projects remain consistent. The article shares five key lessons that have transformed the author's approach to work.
Prompt Engineering Fails Quietly — Prompt Regression Is Why
Small prompt changes can silently cause critical failures in production systems, a phenomenon called prompt regression. The article presents a practical framework for detecting these hidden regressions before they impact users.…
Omen AI’s plan to optimize data centers is all wet
Omen AI has secured $31 million in Series A funding to develop a system that monitors coolant in data centers and prevents bacterial outbreaks. The technology aims to optimize data center operations by addressing a niche but critical infrastructure challenge.
Robot hand company settles Tesla trade secret suit and announces $11M raise
Proception, a startup specializing in robotic hands, has settled a trade secret lawsuit with Tesla and simultaneously announced an $11 million funding round. The company is addressing the challenge of robotic dexterity by innovating in training data collection methods.
Congress wants to ban AI companies from selling your health data
A new bipartisan bill from Senator Warren and Representative Scanlon would expand the Health and Location Data Protection Act to ban AI companies from selling users' health and location data to brokers. The updated legislation specifically addresses data shared with AI chatbots like ChatGPT and Claude, extending protections beyond the original 2022 version.…
Ask an AI expert: What exactly is the full stack?
A Google expert defines the full-stack approach to AI, emphasizing its role as a foundational strategy in their AI development. The approach integrates hardware, software, and model optimization to create efficient and scalable AI systems.
Agent confidence on the technical frontier
Enterprise AI investment is surging, with Gartner predicting 2026 as a pivotal year for aligning AI projects with business goals. As ROI demands increase, executives are turning to agentic AI to deliver measurable financial outcomes.…
This Humanoid Robot Is a Terrifyingly Competent Office Intern
Flexion Robotics, founded by former Nvidia engineers, has developed a humanoid robot that functions as a highly capable office intern. The robot is trained using innovative methods to perform useful tasks, showcasing significant advancements in robotics and AI.
Mapping Europe’s AI Workforce Opportunity
OpenAI's report analyzes the potential impact of AI on the European Union's labor market, identifying occupations that could be automated, those likely to grow, and roles that will see workflow changes. It provides a strategic overview for policymakers and businesses to prepare for AI-driven workforce transformation.
HP Inc. launches Frontier strategic partnership with OpenAI
HP Inc. has expanded its strategic partnership with OpenAI, named Frontier, to integrate AI into customer experiences, software development, and enterprise operations.…
DysLexLens: A Low-Resource LLM Framework for Analysing Dyslexic Learners Insights from Online Forums
DysLexLens is a low-resource LLM framework designed to analyze dyslexic learners' experiences with AI tools by processing online forum discussions. It uses dictionary-driven filtering, knowledge-graph-based reasoning, and both quantitative and qualitative evaluation methods to extract meaningful insights from noisy social media data.…
Odyssey: Constructing Verifiable Local Truth-Preserving Foundation Models
The paper introduces ODYSSEY, a categorical framework for building verifiable, truth-preserving foundation models using composable 'foundries' that encode local contexts, restrictions, and argumentation. It formalizes model construction via Kan extensions and includes a query language (FSQL) and certification system (TICKET) for integrating external models.…
Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning
This paper introduces a unified training paradigm to equip LLM agents with internal world models for future-aware planning, addressing their reactive nature in long-horizon tasks. The authors propose a three-stage approach—World Model Agentic Mid-Training, Format-Eliciting SFT, and Foresight-Conditioned Reinforcement Learning—to bridge the gap between superficial mimicry and genuine predictive capability.…
When Does Personality Composition Matter for Multi-Agent LLM Teams?
This study examines how personality composition affects multi-agent LLM team performance across coding, research collaboration, and bargaining tasks. Results show that personality effects depend on task structure: low agreeableness causes communication shifts but minimally impacts coding milestones, while it significantly degrades performance in open-ended collaboration and bargaining.…
AI-Model Network: Concept, Current State and Future
This paper introduces the concept of a world wide AI-model network (AI-ModelNet), inspired by the Internet, to enable interconnection, capability sharing, and collaborative reasoning among heterogeneous large models. It addresses current challenges of high training costs and deployment complexities by proposing a hierarchical architecture and validating it through a prototype system.…
Suno launches Spark incubator program to feed independent artists to its AI machine
Suno launches Spark, an incubator program offering grants and mentorship to independent artists, aiming to position itself as a streaming destination beyond AI music generation. However, the program requires artists to grant Suno broad licenses for remixing their works, sparking concerns about rights and control.…
China’s Z.ai claims it can match Mythos on cybersecurity
China's Zhipu AI has released GLM-5.2, an open-weight model that some researchers claim matches Anthropic's Mythos in cybersecurity bug-finding tasks. While it still lags behind US models in general tasks, the narrowing capability gap raises concerns for the US government, which has restricted China's access to advanced AI models and hardware.
Why Wall Street thinks US memory maker Micron is the next Nvidia
Wall Street investors are eyeing Micron as a potential AI stock winner similar to Nvidia, given its role as a US memory maker. The company is seen as benefiting from the growing demand for AI-related memory and storage solutions.…
Ford rehires ‘gray beard’ engineers after AI falls short
Ford has rehired experienced 'gray beard' engineers after realizing that relying solely on AI did not ensure high-quality product outcomes. The company acknowledges the mistaken assumption that AI alone would suffice, highlighting the continued need for human expertise.
I Pitted XGBoost Against Logistic Regression on 358 Matches. The Boring Model Won.
The article compares XGBoost and logistic regression across 358 matches, finding that the simpler logistic regression model outperformed the more complex XGBoost due to bias-variance tradeoffs. It emphasizes that smaller models can provide better cross-validated fits and advises on when to choose simpler approaches over complex ones.
Tail Control: The Counterintuitive Engineering of Reliable Agentic Workflows
The article discusses the engineering challenge of ensuring reliable agentic workflows, emphasizing that consistency in delivery time is more critical than raw speed. It introduces the concept of 'tail control' to manage variance, offering counterintuitive solutions for maintaining usability and timeliness in API responses.
Prosecutors used ChatGPT logs as evidence in the Palisades fire trial
Prosecutors used ChatGPT logs as evidence in the trial of Jonathan Rinderknecht, who was charged with arson for starting a deadly wildfire in Los Angeles. The logs showed he generated fire images, expressed anger, and ranted about the wealthy, which prosecutors used alongside location data and witness testimony.…
SoftBank’s CEO isn’t the only one with questions about Elon Musk’s orbital data center hype
Elon Musk's ambitious proposal for orbital data centers is met with skepticism, including from SoftBank's CEO. The hype surrounding the concept raises questions about feasibility and practicality.…
How to Build a Powerful LLM Knowledge Base
This article explains how to leverage coding agents to build a powerful LLM knowledge base. It focuses on practical methods for integrating AI agents to enhance data retrieval and management.…
We Built a Routing Layer to Cut Our AI Costs. It Broke the Product.
A team reduced AI inference costs by over half using a routing layer, but customer satisfaction declined due to quality loss, revealing a Pareto trap. The article presents a detection methodology to identify such trade-offs in days rather than months.
Asian AI startups launch Mythos-like models as Anthropic’s export ban drags on
Asian AI startups are releasing models with capabilities similar to Anthropic's Mythos, capitalizing on the ongoing export ban that restricts U.S. AI labs from serving the Asian market.…
The fittest founder in the room got cancer. Here’s how he used AI to fight back.
Connor Christou, a founder known for his fitness, used AI to combat cancer by inputting his medical data—including blood results, scans, and wearable metrics—into Claude. This approach allowed him to leverage AI for personalized analysis and decision-making during his treatment.
Apple Vision Pro exec is reportedly leaving for OpenAI
Paul Meade, the Apple vice president overseeing the Vision Pro headset, is reportedly leaving Apple to join OpenAI's hardware team. This move highlights OpenAI's continued expansion into hardware development.
Margaret Atwood says the problem with AI is ‘garbage in, garbage out’
Margaret Atwood criticized AI chatbots, specifically Claude, for providing incorrect information, highlighting the 'garbage in, garbage out' problem. She noted that AI lacks human understanding and can produce false answers without awareness.…
Why is Apple asking me to pay more for Big Tech’s AI obsession?
Apple has raised prices across several products, including the MacBook Pro, iPad Air, and HomePod Mini, with CEO Tim Cook attributing the increases to the AI industry's demand for resources. This trend, dubbed 'RAMageddon,' has also affected other tech companies like Microsoft and Nothing, as AI's hardware requirements drive up costs.…
Trump Administration Allows Anthropic to Release Mythos to Select US Organizations
The Trump administration has authorized Anthropic to release its advanced AI model, Mythos, to a limited number of US organizations. This decision follows weeks of negotiations between the White House and the AI company.…
Trump Admin releases Anthropic Mythos to be used by more than 100 US companies, agencies
The Trump administration has authorized over 100 US companies and government agencies to use Anthropic's Mythos 5, including non-American employees. This move expands the deployment of advanced AI across both private and public sectors.
Anthropic’s Mythos 5 is back
Anthropic's Mythos 5 model has been reactivated for select organizations after a prolonged negotiation with the Trump administration, though its public-facing version, Fable 5, remains unavailable with no clear timeline. The agreement follows a letter from Commerce Secretary Howard Lutnick to Anthropic co-founder Tom Brown detailing revised license requirements.
NYT slams Microsoft for building copyright-infringing supercomputer for OpenAI
The New York Times has escalated its copyright dispute against OpenAI and Microsoft, now targeting Microsoft for building a supercomputer allegedly used to infringe on its copyrighted content. This move follows a recent Supreme Court ruling against Sony that may have influenced the NYT's legal strategy.…
From Local LLM to Tool-Using Agent
This article demonstrates how to build a lightweight research agent by combining local LLMs like Gemma 4 with tools such as Ollama, OpenAI Agents SDK, and Tavily MCP. It provides a practical guide for integrating local models with external tools to create an agent capable of performing research tasks.…
OpenAI Has New AI Models. Here’s Why You Can’t Use Them
OpenAI has developed new GPT-5.6 AI models but is delaying their release at the request of the White House. This follows a similar situation where Anthropic had to take its advanced AI models offline.…
OpenAI’s Jalapeño chip is Big Tech’s spiciest move away from Nvidia
OpenAI is developing a custom AI inference chip called 'Jalapeño' in partnership with Broadcom, aiming to reduce reliance on Nvidia's dominant hardware. This move aligns with a broader industry trend where major tech companies like Google, Apple, and SpaceX are building their own chips to mitigate single-supplier risk.…
It’s not about Anthropic vs. OpenAI anymore
The article argues that the focus on competition between AI companies like Anthropic and OpenAI is outdated, as AI models now have significant political implications. Addressing these consequences requires coordinated, collective action rather than corporate rivalry.
Why everyone from OpenAI to SpaceX is building their own chips (and turning up the heat on Nvidia)
Nvidia's dominance in the AI chip market is being challenged as companies like OpenAI, Google, Apple, and SpaceX develop their own custom chips to reduce single-supplier risk. OpenAI's new inference chip, Jalapeño, built with Broadcom, exemplifies this trend toward diversification.…
OpenAI poaches Uber India chief to lead its biggest market outside the US
OpenAI has hired the former Uber India chief to lead its operations in India, its largest market outside the US. This move is part of OpenAI's broader strategy to expand its presence in the country through new offices, partnerships, and hiring.…
OpenAI limits GPT-5.6 rollout after government request, says restrictions shouldn’t be the norm
OpenAI has restricted the rollout of GPT-5.6 following a government request, but argues that such limitations should not become standard practice. The company warns that these restrictions prevent key users—including developers, enterprises, and cyber defenders—from accessing advanced AI tools.
OpenAI unveils GPT-5.6 amid US AI regulatory drama
OpenAI has released GPT-5.6, a new model suite including Sol, Terra, and Luna, just after agreeing to stagger its release at the request of the Trump administration. The models are optimized for coding, cybersecurity, biology, and long-horizon agentic tasks, with pricing significantly lower than competitors like Anthropic's Claude Fable 5.…
Previewing GPT-5.6 Sol: a next-generation model
OpenAI has previewed GPT-5.6 Sol, a next-generation model that demonstrates enhanced performance in coding, scientific reasoning, and cybersecurity. The model is paired with OpenAI's most advanced safety stack to date, emphasizing responsible deployment.
How to Ace Data and ML Behavioural Interviews
This article provides strategies and tips for excelling in behavioral interviews specifically for data science and machine learning roles. It focuses on how to effectively communicate past experiences and problem-solving approaches to impress interviewers.
Amplify the Expert: A Philosophy for Building Enterprise RAG
This article introduces a philosophy for building enterprise Retrieval-Augmented Generation (RAG) systems, emphasizing the goal of amplifying human expertise rather than replacing it. It outlines how this principle guides architectural decisions in enterprise document intelligence.
Water Cooler Small Talk, Ep. 11: Overfitting in RAG evaluation
This article discusses the problem of overfitting in Retrieval-Augmented Generation (RAG) evaluation, comparing it to memorizing for an exam without truly understanding the subject. It highlights how evaluation metrics can be misleading if models are tuned to specific test sets rather than generalizing to real-world scenarios.
Europe Is Fed Up and Wants Its Own AI
Europe is increasingly frustrated with relying on U.S. AI dominance and seeks to develop its own AI capabilities.…
Early Bird pricing ends tonight for TechCrunch Founder Summit
Early Bird pricing for the TechCrunch Founder Summit 2026 ends tonight at 11:59 p.m. PT, offering savings of up to $190 on passes.…
Anthropic’s Mythos mess is only getting worse
Anthropic's Mythos-class AI models remain offline two weeks after a Trump administration ultimatum, with no resolution in sight despite intense negotiations. The company has declined to comment, and uncertainty persists over whether the models will return or if the president's order will expand.
How Qatar Became FIFA’s Technology Test Lab
Qatar has emerged as a testing ground for FIFA's cutting-edge football technology, with the outcomes of these experiments being showcased during the current World Cup. The country's role as a tech lab highlights its influence on the future of the sport.
Anthropic Thinks Its Own Success Is Key to Making AI Safe
Anthropic defends its rapid growth and accumulation of influence as necessary for responsible AI development, despite critics who view this concentration of power as problematic. The company argues that its success enables it to shape AI safety standards effectively.
Agentic Analysis for Agentic Infrastructure: An LLM-Powered Pipeline for Comparative Governance of DAO and Corporate AI Protocols
This paper introduces an LLM-powered pipeline for analyzing governance discourse in AI agent protocols, comparing ERC-8004 (permissionless, on-chain) and Google A2A (corporate-led). Analyzing over 4,300 records, the study finds that while governance form influences thematic focus, both regimes show similar participation inequality and fragmentation, though open governance fosters denser discourse alignment.…
AlgoEvolve: LLM-driven Meta-evolution of Algorithmic Trading Programs
AlgoEvolve uses LLMs as semantic mutation operators in an evolutionary framework to generate and improve algorithmic trading strategies. The system autonomously adapts to market regimes and includes a meta-evolutionary loop that refines the prompts guiding program synthesis, outperforming human-designed instructions.…
Refusal Lives Downstream of Persona in Chat Models
This research reveals that refusal behavior in chat models is gated by a compliant persona, not an isolated mechanism. By steering persona and refusal directions in models like Qwen2.5 and Llama, the authors show that refusal can be suppressed or restored depending on persona activation.…
Life After Benchmark Saturation: A Case Study of CORE-Bench
This paper argues that retiring benchmarks after accuracy saturation overlooks other critical dimensions of agent performance, such as construct validity, efficiency, reliability, and human-agent collaboration. Using CORE-Bench Hard as a case study, the authors introduce improved benchmarks and demonstrate that meaningful insights can still be gained even after accuracy plateaus.…
Detecting and Controlling Sycophancy with Cascading Linear Features
This paper introduces an iterative data generation pipeline that isolates cascading linear features to detect and control sycophancy in language models. By moving beyond binary contrastive pairs to samples with linearly scaling features, the method improves feature disentanglement and enables effective steering with lower computational cost.…
Why Amazon Dropped Its OpenAI Movie, Data Center Workers Fight Back, and Meta Leaks Employee Data
Amazon's MGM Studios has decided to drop an OpenAI-related movie, highlighting the growing intersection of AI and the film industry. Meanwhile, data center workers are pushing back against industry practices, and Meta has suffered a leak of employee data.…
Patronus AI lands $50M to build ‘digital worlds’ that stress-test AI agents
Patronus AI, founded by ex-Meta AI researchers, has raised $50 million to develop simulated environments for stress-testing AI agents. The startup is seeing high demand for its agent evaluation platform, according to its investor.
The White House is asking OpenAI to slow roll the release of its new model over safety concerns
OpenAI is reportedly planning to release its new model, GPT 5.6, only to a select group of partners rather than the public, following a request from the Trump administration over safety concerns. This reflects ongoing tensions between AI development speed and government oversight.
OpenAI will delay GPT-5.6 after Trump administration request
OpenAI is delaying the full release of GPT-5.6 after a request from the Trump administration, which cited security concerns. The model will initially be available only to a limited group of enterprise customers, with the administration approving access on a case-by-case basis.…
Beyond Shapley: Efficient Computation of Asymmetric Shapley Values
This paper introduces efficient computation methods for Asymmetric Shapley Values (ASV), a variant of Shapley values that incorporates causal knowledge for model explainability. The authors show that ASV can be computed in polynomial time for certain causal graphs, such as rooted directed trees, and propose an approximation algorithm for arbitrary causal DAGs using uniform sampling of topological orderings.…
Do vision-language models search like humans? Reasoning tokens as a reaction-time analog in classic visual-search paradigms
This study adapts classic visual-search paradigms to evaluate vision-language models (VLMs), using reasoning tokens as an analog for human reaction time. The models replicate several human behavioral signatures, such as flat effort for feature search and climbing effort for conjunction search, but also show key divergences like reversed target-present/absent effort slopes.…
Diagnosing and Mitigating Compounding Failures in Agentic Persuasion via Taxonomic Strategy Retrieval
This paper identifies semantic leakage in standard RAG as a key cause of compounding errors in multi-step agentic persuasion tasks, leading to problem drift and sycophantic conformity. The authors propose Taxonomic Strategy RAG (TS-RAG), which uses a categorical bottleneck to decouple argument structure from content, improving transfer of abstract logic and enabling lightweight agents to outperform larger opponents.…
Project Auto-World: Towards Automated Benchmarking of Neural Relational Reasoners
This paper addresses the challenge of evaluating neural models on relational reasoning tasks by using LLMs to automate benchmark generation. It employs LLM-driven evolutionary and agentic search to create hard problem instances, improving the Edge Transformer's generalization.…
The Hitchhiker's Guide to Agentic AI: From Foundations to Systems
This article describes a comprehensive practitioner's reference for building autonomous AI systems, covering the full stack from transformer architecture and GPU systems to production deployment. It emphasizes understanding every layer of the pipeline, including LLM foundations, alignment and reasoning, agentic AI design patterns, inter-agent coordination, and evaluation methodology.…
Beyond the Straight Line: Choosing Between OLS, Interaction Terms, and Tweedie Regression
This article explores the decision-making process for selecting between Ordinary Least Squares regression, models with interaction terms, and Tweedie regression based on data characteristics like zeros and outliers. It emphasizes that the choice depends on how the data handles messy real-world distributions.…
The Hot Path Belongs to GBDTs, Agents Own the Cold Path: A Payment-Fraud Benchmark
This article presents a benchmark comparing GBDTs and agents in payment-fraud detection, focusing on latency, cost, and reproducibility. It argues that GBDTs excel on the 'hot path' (real-time decisions), while agents are better suited for the 'cold path' (complex, slower analysis).…
Vector RAG Isn’t Enough — I Built a Context Graph Layer for Multi-Agent Memory
The article compares raw chat history, vector-only RAG, and a context graph for multi-agent memory, finding that vector RAG has a surprising weakness in relational retrieval. The author built a context graph layer to address this issue, improving multi-agent conversation memory.
Anthropic says Alibaba must be punished for largest Claude cloning attack
Anthropic has accused Alibaba of orchestrating a massive cloning attack on its Claude AI model, allegedly using 25,000 accounts to extract data from 28.8 million exchanges. The company is calling for punishment, highlighting the severity of the intellectual property theft.…
Databricks’ former AI chief thinks he can cut AI’s power bill by 1,000x
The former AI chief at Databricks claims his new company's technology can reduce AI's energy consumption by a factor of 1,000. Their tool, Un-0, is an image-generation system that demonstrates the ability to replicate conventional AI systems more efficiently.…
General Intuition’s $2.3B bet that video games can train AI agents for the real world
General Intuition raised $320 million to scale AI trained on millions of hours of video gameplay, aiming to develop human-like intuition. The company believes action data from games can help AI agents operate effectively in real-world scenarios.…
Anthropic’s Claude is winning over paid consumers, a market owned by ChatGPT
Anthropic's Claude is gaining traction among paid AI consumers, challenging ChatGPT's dominance in the market. Data indicates a growing preference for Claude among users who pay for AI services.
Our latest Google Finance upgrades, including a new app
Google Finance is exiting beta and introducing a new Android app, marking a significant upgrade to its financial tracking platform. The update aims to provide users with improved tools and a better mobile experience for managing investments and market data.
One Month Into Learning Data Engineering in Public: Here’s What I Didn’t Write About
The article reflects on the author's first month of learning data engineering publicly, focusing on the unspoken challenges and motivations that sustained their progress. It highlights the personal and emotional aspects of the learning journey beyond technical achievements.
Letting an LLM Pick the Right RAG Page: The Arbiter Pattern at the End of Retrieval
The article introduces the 'Arbiter Pattern' for Retrieval-Augmented Generation (RAG), where a single LLM call ranks candidate documents with explanations, outputting a typed object for auditability. This approach enhances transparency and defensibility in enterprise document intelligence by letting the LLM select the most relevant page at the end of retrieval.
3 Agents. 3 LLMs. 1 Aging GPU: Engineering Parallel Inference on Bare Metal
This article explains how to run three different LLMs simultaneously on a single 8GB GPU by using C++ layer multiplexing and admission control, overcoming VRAM limitations. It demonstrates a practical approach to parallel inference for multi-agent systems on aging hardware.
Amazon ups India bet with fresh $13B AI infrastructure investment
Amazon has announced a new $13 billion investment in India, focusing on expanding AI infrastructure. This move is part of a broader trend among global tech companies to strengthen their AI capabilities in the Indian market.
Adobe acquires image and video enhancement tool maker Topaz Labs
Adobe has announced the acquisition of Topaz Labs, a company known for its image and video enhancement tools. The integration of Topaz Labs' technology will be implemented across Adobe's suite of applications.
2 days left to save up to $190: Join 1,000+ founders and investors at TechCrunch Founder Summit
The TechCrunch Founder Summit 2026 is offering early bird discounts of up to $190, but the deal expires on June 26 at 11:59 p.m. PT.…
Netris raises $15M Series A from a16z to help AI neoclouds go live faster
Netris has secured $15 million in Series A funding from a16z to accelerate its software platform for AI neocloud operators. The company's solution runs on network switches, enabling faster deployment and reduced time-to-live for neocloud infrastructure.…
From Fortnite to robots: General Intuition raises $2.3B on bet that video games can train AI agents for the real world
General Intuition has raised $2.3 billion to scale AI trained on millions of hours of video game gameplay, aiming to develop AI agents with human-like intuition for real-world applications. The company believes that action data from games like Fortnite can help train AI to navigate physical environments and perform tasks more effectively.
How agents are transforming work
OpenAI's latest research paper highlights how AI agents are revolutionizing work by handling longer and more complex tasks. This advancement is boosting productivity across various job roles and industries.
Repositioning retail for the AI era
AI is transforming retail primarily through backend operations like search result optimization, supply chain management, and accelerated software development, rather than through visible consumer-facing features. The article highlights that the most significant changes are happening behind the scenes, reshaping decision-making processes.
IBM claims world’s first sub-1 nanometer chip technology
IBM has announced the development of the world's first sub-1 nanometer chip technology using nanostack transistors. This breakthrough could significantly enhance chip performance or improve energy efficiency in future electronics.…
British Police Built a Sprawling Crime-Prediction Machine. Some Results Couldn’t Be Trusted
A WIRED investigation into a British police force's use of predictive analytics reveals that the system produced unreliable results, highlighting the challenges of implementing AI in law enforcement. The experiment aimed to forecast crime but faced issues with data quality and trustworthiness, underscoring the messy reality of AI adoption in public safety.
World Cup Teams Are in a Race for AI Dominance
FIFA is offering an AI agent to all World Cup teams, aiming to democratize access to advanced analytics. However, the disparity in resources may still allow wealthier teams to gain a competitive edge through superior AI tools.…
Facebook’s Creator Studio has been revived as an AI companion app
Meta has revived its Facebook Creator Studio as a standalone AI companion app, featuring an AI Creator Assistant chatbot. The assistant provides performance insights, engagement recommendations, and helps draft replies to comments.…
Ford had to hire back former engineers to fix mistakes made by its automated systems
Ford has acknowledged that its reliance on automated systems in production and design led to significant errors, forcing the company to hire back former engineers to fix the mistakes. The automaker emphasizes that AI's effectiveness is highly dependent on the quality of training data, and that while powerful, these systems are prone to pitfalls.
Europe is pushing back on Washington’s chip war
Europe is resisting U.S. efforts to expand chip export restrictions on China, particularly regarding older-generation deep ultraviolet lithography tools from ASML.…
A24 Knows You’re Mad About the Google AI Collab
Google DeepMind's $75 million investment in indie studio A24 has sparked backlash from fans concerned about AI's growing influence in Hollywood. The deal highlights tensions between artistic communities and tech companies expanding into entertainment.
How to Opt Out of Google Search’s New AI Data Training Feature
Google's new Search history update collects media uploads, such as images from reverse image searches, to train its AI models. Users can opt out of this data collection through their account settings.…
The memory chip crunch is paying off for this US company
A US memory chip company has seen its revenue quadruple to $41.45 billion year-over-year, with profits soaring from $1.88 billion to $28.2 billion. This dramatic growth is attributed to the ongoing memory chip crunch driven by high demand for AI and data center infrastructure.
AI researchers continue to leave Google for its rivals
Two top AI researchers, Jonas Adler and Alexander Pritzel, are leaving Google to join Anthropic, continuing a trend of talent departing for rivals. This follows earlier high-profile exits like Noam Shazeer and John Jumper.
AI was supposed to kill engineering jobs, but new data suggests they’re the most resilient
Despite fears that AI would eliminate engineering jobs, new data from SignalFire shows that engineers are actually being hired at a higher rate, making them more resilient to AI-driven layoffs. The trend suggests that AI is augmenting rather than replacing engineering roles.
Cerebras stock plunges after earnings as CEO says margin outlook was misunderstood
Cerebras shares dropped sharply after its first earnings report as a public company, where the CEO clarified that the company's margin outlook was misinterpreted by investors. The AI chipmaker projected a narrower gross margin in its core business, leading to a sell-off.
Former Infosys chief has a new startup that wants to challenge the IT services world
Vishal Sikka, former CEO of Infosys, has launched a new startup backed by Mayfield and Aramco Ventures. The venture aims to disrupt the IT services industry by assembling a team of veterans from SAP, Infosys, and VianAI.
Your First Task as a Data Engineer in a New Company? Make the ETL Pipeline Testable
This article provides a practical onboarding workflow for data engineers, focusing on making ETL pipelines testable through environment setup and automated testing. It also highlights the role of AI-assisted development in streamlining these processes.
How to Build a Credit Scoring Grid From a Logistic Regression Model
This article explains how to convert logistic regression model coefficients into a credit scoring grid with a 0–1000 scale, including risk classes and stability checks. It provides a practical guide for implementing interpretable credit scoring systems in financial applications.
I Met With China’s Top AI Experts. They’re Freaking Out, Too
Chinese AI experts express deep concern over the escalating AI arms race between China and the US, fearing a catastrophic event akin to a 'Chernobyl moment.' The article highlights shared anxieties among researchers on both sides about the potential for uncontrolled AI development.
Figma adds code layers, support for animations, more AI features in new update
Figma's latest update introduces a code layer, motion and shader support, and AI-powered custom plug-in creation. These features enhance design-to-development workflows and enable more dynamic, automated design processes.
Agility Robotics plans to go public via SPAC in a $2.5B deal
Agility Robotics, a humanoid robotics startup from Oregon State University, plans to go public through a SPAC merger in a deal valued at $2.5 billion. The company anticipates generating $620 million in proceeds from the transaction.
Facebook rolls out an AI companion app for creators
Facebook is testing a new companion app for creators that integrates its recently launched AI creator assistant. The app is currently available to a select group of creators for testing.
Figma now has AI motion graphics and shader tools
Figma announced new AI-powered features at its Config conference, including AI-generated motion graphics and shader tools that allow users to create animations via a chatbot. The platform also introduced coding layers for tweaking project code without leaving the design canvas, aiming to streamline full-stack development.…
The $27 million Al proxy war over Alex Bores ends in a draw
A $27 million political proxy war between Anthropic and OpenAI ended in a draw as Alex Bores, a New York state Assemblyman backed by a pro-AI super PAC, narrowly lost the Democratic primary for NY's 12th Congressional district. Bores had previously authored the RAISE Act, which imposed safety requirements on AI companies, drawing opposition from the pro-AI group Leading the Future.…
Congresswoman denies staff used AI to write defense funding amendment
Congresswoman Anna Paulina Luna denied that her staff used AI to draft a defense funding amendment, claiming AI was only used for spellcheck on a summary. Screenshots showed AI-generated text in an amendment summary for the 2027 National Defense Authorization Act, but Luna insists no legislation is ever drafted with AI.
Anchor Detection for RAG: Parallel Detectors, Then One LLM Call at the End
This article discusses a method for anchor detection in Retrieval-Augmented Generation (RAG) systems, using parallel detectors (keywords, table of contents, embeddings) to filter structured data before a single LLM call. The approach aims to improve efficiency and accuracy in enterprise document intelligence by prioritizing retrieval steps.…
Why I Stopped Using One Agent and Built a Multi-Agent Pipeline Instead
The article discusses the limitations of using a single AI agent for complex tasks like text-to-SQL and presents a practical case for switching to a multi-agent pipeline. It provides a walkthrough of how multiple specialized agents can collaborate to improve accuracy and efficiency.…
A Three-Phase Factual Recall Circuit in Gemma-2B and Gemma-12B-IT
This article investigates the internal mechanisms of factual recall in Gemma-2B and Gemma-12B-IT models, using activation patching to identify a three-phase circuit involving storage, routing, and readout across transformer layers. The findings highlight that the residual stream plays a dominant role in this process, offering insights into how transformer models retrieve factual information.
Qualcomm Buys Buzzy Chip Startup Modular for Nearly $4 Billion
Qualcomm has acquired chip software startup Modular for nearly $4 billion, marking a significant exit for one of the most promising companies in the AI chip space. The deal highlights Qualcomm's push to strengthen its position in AI hardware and software integration.
The Trump White House Is Over Anthropic CEO Dario Amodei
Anthropic CEO Dario Amodei has been replaced by cofounder Tom Brown in high-stakes meetings with the Trump White House, with one official reportedly calling Amodei a 'weirdo.' The shift suggests a change in how the administration engages with the AI company.
3 days left to save up to $190 on your TechCrunch Founder Summit 2026 pass
Early bird pricing for TechCrunch Founder Summit 2026 ends in 3 days, offering savings of up to $190. The deadline is June 26 at 11:59 p.m.…
OpenAI unveils its first custom chip, built by Broadcom
OpenAI has introduced its first custom chip, named Jalapeño, developed in collaboration with Broadcom. The processor is tailored to meet the specific demands of OpenAI's inference systems, marking a significant step in hardware optimization for AI workloads.
The Google Home Speaker sounds good and looks great — but it’s finicky
The Google Home Speaker performs well with voice recognition, rarely missing the wake word even in challenging conditions like loud music or running water. However, it struggles with very quiet commands, such as stage-whispers from another room.…
OpenAI reveals its first AI processor: Jalapeño
OpenAI has unveiled its first custom AI processor, named Jalapeño, developed in partnership with Broadcom. The chip is an ASIC designed specifically for AI inference, powering models like those behind ChatGPT and Codex.…
OpenAI and Broadcom unveil LLM-optimized inference chip
OpenAI and Broadcom have announced Jalapeño, a custom AI chip specifically designed to optimize LLM inference. The chip aims to enhance performance, energy efficiency, and scalability for AI systems.…
The emergence of the web data infrastructure layer for AI
The article discusses the growing need for a web data infrastructure layer to support AI, as enterprises require large-scale, structured data that is often blocked or unstructured on the web. It highlights the challenge that the web was not originally designed for AI consumption, necessitating new infrastructure solutions.
India’s MoEngage bets that the future of marketing is millions of AI agents
MoEngage, an Indian marketing firm, is investing in AI agent technology to personalize customer interactions at scale. The acquisition enables the deployment of millions of AI agents, each assigned to individual customers, aiming to transform marketing strategies.
Reinforcement Learning Towards Broadly and Persistently Beneficial Models
This paper investigates whether reinforcement learning (RL) on beneficial behaviors in realistic domains can produce models that generalize alignment beyond their training distribution. The authors train models using RL on a dataset promoting traits like truthfulness and fairness, finding that this improves performance on over 80% of out-of-distribution alignment benchmarks and enhances resistance to adversarial prompting and harmful fine-tuning.…
Safe and Generalizable Hierarchical Multi-Agent RL via Constraint Manifold Control
This paper introduces a hierarchical multi-agent reinforcement learning framework that enforces hard safety constraints via a constraint manifold at the low level while enabling coordination through high-level policy learning. It provides theoretical safety guarantees and stable training dynamics, achieving competitive performance with near-perfect safety rates and generalizing to varying numbers of agents and obstacles.
Critique of Agent Model
This paper critiques current AI agent models by distinguishing between 'agentic' systems (relying on external scaffolding) and 'agentive' systems (with internalized agency). It proposes a Goal-Identity-Configurator (GIC) architecture for general-purpose agents, emphasizing autonomy, self-regulation, and safety under human oversight.…
Neuro-Symbolic Drive: Rule-Grounded Faithful Reasoning for Driving VLAs
Neuro-Symbolic Drive introduces a framework that supervises driving Vision-Language-Action (VLA) models using rule-grounded reasoning traces from classical rule-based planners, ensuring causal coupling between reasoning and motion. By fine-tuning Qwen3.5-4B with these structured traces, the method significantly reduces displacement error and miss rate in driving benchmarks.…
RIFT-Bench: Dynamic Red-teaming For Agentic AI Systems
RIFT-Bench is a new methodology for dynamically red-teaming agentic AI systems, using a graph-based approach to evaluate security across diverse architectures. It operates in two automated phases—Discovery and Scanning—to extract system structure and deploy adaptive adversarial attacks.…
Hollywood is bending the knee to OpenAI
Major Hollywood studios like Netflix, A24, and Warner Bros. have declined to distribute a biographical drama about OpenAI CEO Sam Altman, raising concerns about the industry's willingness to produce critical stories about Big Tech.…
Google Home will soon get better at recognizing you
Google Home is updating its facial recognition feature to identify familiar people even when their faces aren't clearly visible, using non-biometric signals like body size and clothing color. The update, starting June 23rd, will also automatically update the Familiar Faces library with recent images to reduce inaccurate notifications.…
Helping build shared standards for advanced AI
OpenAI is contributing to the development of shared standards for advanced AI by supporting evaluation frameworks, safety practices, and global cooperation through the Appia Foundation. This initiative aims to promote responsible AI development and international collaboration.
I Spent an Hour on a Data Preprocessing Task Before Asking Gemini
The author recounts spending an hour on a data preprocessing task in Pandas, only to have Google's Gemini solve it in seconds. While AI tools can dramatically speed up coding, the article emphasizes that understanding data science fundamentals remains crucial for evaluating and improving AI-generated solutions.
How to Create Powerful Loops in Claude Code
This article explains how to implement loops in Claude Code to enhance the functionality of coding agents. It provides guidance on using iterative processes to automate and improve code generation tasks.
Anthropic’s Claude Tag is learning your company, one Slack message at a time
Anthropic has introduced Claude Tag, an always-on AI assistant integrated into Slack, designed to boost productivity by learning from organizational messages. The feature aims to capture institutional knowledge and enterprise workflows, positioning it as a strategic tool for businesses.
Why corporate AI super PACs spent $27 million on a local election
The article discusses how corporate AI super PACs spent $27 million on a local election, highlighting the growing influence of tech money in local politics. It suggests that these investments are aimed at shaping regulatory environments favorable to AI companies.…
How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery
Immunologist Derya Unutmaz used GPT-5 Pro to unravel a three-year-old mystery about T cell behavior, leading to new insights that could advance cancer and autoimmune disease research. The AI model's ability to analyze complex biological data was key to the breakthrough.
Build Your Own Local AI Coding Agent with Gemma 4 and OpenCode
This article provides a step-by-step guide on setting up a local AI coding agent using Gemma 4 and OpenCode, starting from installing Ollama to launching the agent. It focuses on enabling developers to run AI-powered coding assistance locally without relying on cloud services.
The Era of No-Code AI: What You Need to Know
The article discusses the rise of no-code AI tools, which are democratizing access to artificial intelligence and reducing the need for traditional programming skills. It highlights how this shift may make programmers feel less unique as AI development becomes more accessible to non-technical users.…
Retrieval Is Filtering, Not Search: A Mental Model for Enterprise RAG
The article argues that in enterprise RAG systems, retrieval should be viewed as a filtering process rather than traditional search. It emphasizes filtering structured data like line_df and toc_df, selecting small anchors, and expanding context for better results.
Fika Jobs raises $4M to build a video-first hiring platform where AI agents interview candidates
Fika Jobs, a Stockholm-based startup, has raised $4 million to develop a video-first hiring platform that uses AI agents to conduct interviews and features short-form video profiles. The platform aims to blend elements of LinkedIn and TikTok to modernize the recruitment process.
4 days left to save up to $190 on TechCrunch Founder Summit 2026
TechCrunch Founder Summit 2026 is offering early bird discounts of up to $190, with the promotion ending on June 26. The event is positioned as a founder bootcamp for entrepreneurs.…
Meta launches cheaper smart glasses without Ray-Ban
Meta has launched a cheaper line of smart glasses without the Ray-Ban branding, offering three styles and seven colors, including a collaboration with Kylie Jenner. The move marks a shift from its previous partnership with EssilorLuxottica, which helped make earlier smart glasses more stylish and mainstream.…
Sony’s AI Camera Assistant is exactly as bad as it looks
Sony's new AI Camera Assistant on the Xperia 1 VIII produces poor-quality photos, as confirmed by a week of testing. The feature, which was promoted with unflattering sample images, fails to improve photography and is compared unfavorably to Google's Camera Coach.…
The Fitbit Air takes a smarter approach to the AI health dumpster fire
The Google Fitbit Air is presented as a more thoughtful AI-integrated health tracker, but its AI health coach can be overly critical, offering advice based on metrics like sleep, heart rate variability, and temperature. The article highlights the mixed user experience of relying on AI for personalized health guidance.
Something’s off with Midjourney’s pivot to body scanners
Midjourney, known for its AI image generator, has pivoted to medical imaging with a futuristic ultrasound scanner that immerses users in water. The company claims it could rival MRI, but experts are skeptical due to a lack of public evidence supporting the technology.
How Omio is building the future of conversational travel
Omio is leveraging OpenAI's technology to create conversational travel experiences, enabling users to interact naturally with the platform. This AI integration is accelerating product development and helping Omio transform into an AI-native company.…
The $400 million machine powering the future of chipmaking
The article describes a massive $400 million machine used in advanced chip manufacturing, highlighting its enormous size and complexity. It emphasizes the machine's critical role in enabling future semiconductor technology.…
OpenAI launches new initiative to help find and patch open-source bugs
OpenAI has launched a new initiative aimed at identifying and fixing security vulnerabilities in open-source software. The program seeks to improve the overall security posture of the open-source ecosystem by leveraging AI tools.
The running list: major tech layoffs in 2026 where employers cited AI
This article compiles a list of major tech companies in 2026 that have conducted significant layoffs, explicitly citing artificial intelligence as a contributing factor. The list is presented in reverse chronological order, highlighting the growing impact of AI on workforce reductions in the tech industry.
Meta Exposed Data Internally From Its Controversial Employee-Tracking Program
Meta's internal employee-tracking program, which collects keystroke data to train AI models, was exposed after employees raised concerns about the initiative. The program has been controversial due to privacy and ethical implications.…
Meta Pauses Employee-Tracking Program Following Internal Data Leak
Meta has paused its employee-tracking program after an internal data leak exposed potentially sensitive information from the initiative. The company is now reviewing security measures to prevent future breaches.
Nvidia wants to cut data center water use, but that’s not the same as fixing AI’s water problem
Nvidia has introduced a cooling system that reduces water consumption within data centers. However, this innovation does not tackle the larger water usage associated with the fossil fuel power plants that supply electricity to run AI systems.…
AI chipmaker Groq confirms $650M raise, re-staffs after Nvidia’s $20B not-acqui-hire deal
AI chipmaker Groq has confirmed a $650 million funding round and is restructuring its leadership after a $20 billion deal with Nvidia fell through. The company is now focusing on its neocloud business and hiring new executives.
The AI world is getting ‘loopy’
The article introduces a new concept called 'the loop' that advances agentic AI by enabling a swarm of agents to operate continuously in the background without stopping. This represents a shift from task-based AI to persistent, autonomous background processes.
Nvidia says its AI data center design runs hotter to use a lot less water
Nvidia claims its new Rubin generation reference design for liquid-cooled AI data centers significantly reduces power and nearly eliminates water usage. However, the design does not address broader concerns such as construction impacts or the cost comparison with air-cooled alternatives.…
Encoding Categorical Data for Outlier Detection
The article discusses the limitations of one-hot encoding for categorical data in outlier detection tasks and explores alternative encoding methods. It highlights how different encoding techniques can better preserve the structure of categorical variables for identifying anomalies.
OpenAI Launches Full-Scale Effort to Patch Open-Source Bugs as It Takes on Anthropic’s Mythos
OpenAI has announced an enhanced version of GPT-5.5-Cyber, focusing on cybersecurity, alongside a new initiative called 'Patch the Planet' aimed at fixing bugs in open-source software. This move positions OpenAI to compete with Anthropic's reputation for safety and reliability.…
SpaceX inks compute deal with Reflection AI, an open source AI lab
SpaceX has signed a compute deal with open-source AI lab Reflection AI, which will pay $150 million per month from July 2026 to 2029 for priority access to Nvidia's GB300 AI chips and hardware at SpaceX's Colossus 2 data center in Tennessee. The agreement highlights growing demand for high-performance AI infrastructure and partnerships between AI labs and data center operators.
Amazon is testing Alexa+ in India with Hindi support
Amazon is expanding its Alexa+ AI assistant to India, offering a Hindi-language version for user testing. The move aims to increase the assistant's footprint in the Indian market.
Google DeepMind bets $75M on AI’s future in Hollywood with A24 deal
Google DeepMind has invested $75 million in a partnership with A24 to develop AI-powered filmmaking tools. The collaboration aims to integrate advanced AI technologies into the creative processes of Hollywood production.
AI is cursing renters with the promise of impossible homes
The article highlights how AI-generated listings on rental platforms create unrealistic expectations for renters, often featuring apartments that don't exist or are significantly different in person. This practice wastes time and causes frustration, as seen in the experience of a New Yorker who toured a dream apartment only to find it was not as advertised.
Codex-maxxing for long-running work
Jason Liu demonstrates how to use Codex for maintaining context across long-running projects, enabling complex work to continue beyond a single prompt. The approach focuses on preserving state and managing project complexity effectively.
Daybreak: Tools for securing every organization in the world
OpenAI has launched Daybreak, a suite of tools including Codex Security and GPT-5.5-Cyber, designed to help organizations identify, validate, and patch security vulnerabilities at scale. These tools aim to enhance cybersecurity defenses across various industries.
Patch the Planet: a Daybreak initiative to support open source maintainers
OpenAI has launched Patch the Planet, a Daybreak initiative aimed at supporting open-source maintainers. The program uses AI and expert review to help identify, validate, and fix vulnerabilities in open-source projects.
Three things to watch amid Anthropic’s latest feud with the government
Anthropic is embroiled in a new conflict with the US government over its AI model Mythos, raising questions about regulation and transparency. The feud highlights ongoing tensions between AI developers and federal oversight.…
Neural Networks, Explained for Beginners: Start Here If They’ve Confused You
This article provides a beginner-friendly explanation of neural networks, focusing on the intuition behind them and the necessity of activation functions. It aims to clarify common confusion for newcomers to the topic.
When RAG Users Ask Vague Questions: Clarify Once, Learn the Default
The article discusses a strategy for handling vague questions in Retrieval-Augmented Generation (RAG) systems by asking a single focused clarification and learning the user's default preference from the response, thereby reducing future interruptions. It emphasizes improving user experience and efficiency in enterprise document intelligence applications.
How to Use Claude Code in Your Browser
This article explains how to use Claude Code, a coding agent, directly within a browser to verify and improve coding work. It provides practical guidance on applying AI-powered tools for code review and validation in a web-based environment.
The founder conference built for growth: TechCrunch Founder Summit pass rates increase June 26
TechCrunch Founder Summit 2026 is offering discounted passes with savings up to $190 if purchased by June 26 at 11:59 p.m. PT.…
World Cup Scams Are Getting Harder to Spot
AI-powered scams are making it increasingly difficult for World Cup fans to distinguish between legitimate and fraudulent offers, including fake tickets and cloned websites. The article highlights how advanced AI tools are being used to create more convincing scams, raising concerns about fan safety and trust.
Some Electricians Think Building Data Centers Is for Sellouts
Big Tech's massive investment in data center construction is facing growing national opposition, leading some electricians to question the ethical and professional implications of working on these projects. The article highlights a divide among workers who see the work as lucrative but potentially at odds with their values.
Read this before you vibe-code another app
A project manager named Bob Starr quickly launched a 'vibe-coded' website called 'Boomberg' to track US tax money going to tech companies, only to later discover a hidden SQL injection vulnerability that could have allowed attackers to read or alter data. The oversight highlights a common blindspot among developers using rapid, AI-assisted coding methods without fully understanding security risks.…
Samsung Electronics brings ChatGPT and Codex to employees
Samsung Electronics is rolling out ChatGPT Enterprise and Codex to its global workforce, representing one of OpenAI's largest enterprise deployments. This move integrates advanced AI tools into the company's operations, enhancing productivity and coding capabilities for employees.
Tool Calling, Explained: How AI Agents Decide What to Do Next
This article explains how AI agents use tool calling to interact with external systems, enabling them to retrieve data or perform actions beyond text generation. It breaks down the decision-making process that allows LLMs to select and invoke appropriate tools based on user requests.
What Are the Possibilities to Build Date Tables in Self-Service Environments?
The article explores alternatives to building date tables using DAX code in self-service data environments, comparing different methods for creating them without upstream data flow. It highlights new approaches that offer more flexibility and efficiency for analysts working with tools like Power BI.
Reconstructing the Table of Contents a PDF Forgot to Ship, So RAG Can Scope by Section
The article discusses methods to reconstruct a table of contents from PDFs that lack an outline, enabling RAG systems to scope retrieval by section. It presents two approaches for extracting structure and highlights a commonly overlooked page-alignment step.
Beyond Siri: Here are the practical AI features coming to your iPhone in iOS 27
While Siri's AI upgrade dominated WWDC headlines, Apple is introducing practical AI features across iOS 27 that enhance user experience beyond the virtual assistant. These tools focus on everyday functionality, making AI more accessible and useful for iPhone users.
When the Trump administration cracks down on Anthropic, who benefits?
The article examines the Trump administration's recent actions against Anthropic, exploring the motivations behind the crackdown and its potential implications for the broader AI ecosystem. It raises questions about who stands to gain from these regulatory moves.
28 Tips to Take Your ChatGPT Prompts to the Next Level
This article provides 28 tips for enhancing ChatGPT prompts to achieve more sophisticated and interesting results beyond basic usage. It emphasizes the role of smart prompt engineering in unlocking the chatbot's full potential.
Signal’s Meredith Whittaker wants you to remember that AI chatbots ‘are not your friends’
Meredith Whittaker, president of Signal, warns users not to anthropomorphize AI chatbots, emphasizing they are not conscious or sentient beings. She reminds people that these tools are not friends or interlocutors with genuine understanding.
7 Crucial Barriers Between Data Teams and Self-Healing Data Architecture
The article outlines seven key obstacles that prevent data teams from implementing self-healing data architecture using AI. It emphasizes the need for practical solutions to overcome these barriers to make autonomous data management a reality.
Nobel laureate John Jumper is leaving DeepMind for rival Anthropic
Nobel laureate John Jumper is departing Google DeepMind to join rival AI company Anthropic, signaling a significant talent shift in the industry. His move follows a trend of high-profile researchers leaving DeepMind for competitors.
In the Weights is your new AI-centric vanity search
In the Weights is a new AI-powered vanity search tool that lets users look up their personal 'score' based on their online presence. The platform uses AI to analyze and rank individuals, creating a fun, self-referential search experience.…
The Atlantic created a searchable database of the music used to train AI
The Atlantic has created a searchable database of four music datasets used to train AI models, with two containing millions of tracks. Google and Stability AI have confirmed using these datasets in research.…
Materialized Lake Views in Microsoft Fabric: When Your Medallion Fits in a SELECT Statement
Microsoft Fabric introduces Materialized Lake Views, a new declarative layer that consolidates five data surfaces into a single SELECT statement. The article covers the syntax and general availability capabilities of this feature, simplifying data architecture.
Making a PDF’s Images Searchable for RAG, Without Paying to Read Them All
The article discusses a method to make images within PDFs searchable for Retrieval-Augmented Generation (RAG) without incurring high costs. It introduces 'image_df' as a tool to locate images, then selectively converts only relevant ones into searchable text to optimize expenses.…
Siri AI Hands On: A Smart, Helpful Assistant
The new Siri AI has been upgraded to be conversational, omnipresent, and genuinely helpful, marking a significant improvement in digital assistant capabilities. It offers a more intuitive and responsive user experience.
Encryption, spyware, and now Mythos: History shows why cyber export control doesn’t work
The article argues that historical attempts to control the export of cybersecurity software have consistently failed over the past three decades. It questions the effectiveness of applying similar controls to Anthropic's new cybersecurity model, Mythos.
Building a Custom GStreamer Plugin for NVIDIA DeepStream
The article discusses the process of creating a custom GStreamer plugin for NVIDIA DeepStream, focusing on enabling custom inference capabilities. It highlights the importance of tailored plugins for optimizing video analytics pipelines in AI applications.
Python 3.14 and its New JIT Compiler
Python 3.14 introduces a new Just-In-Time (JIT) compiler, which aims to improve performance by compiling code at runtime. The article provides a technical overview of the JIT implementation and includes benchmark results demonstrating speed improvements.…
Is the US government’s Anthropic ban accidentally helping the brand?
The US government forced Anthropic to remove its latest AI models, Fable 5 and Mythos 5, over national security concerns after Amazon researchers reportedly bypassed Fable 5's guardrails. Cybersecurity researchers and Anthropic have criticized the move, noting similar vulnerabilities exist in other models.…
GPU-Resident Top-K for Agentic RAG: I Built a CUDA Kernel So My Retrieval Step Would Stop Bouncing Off the GPU
The article discusses how PCIe transfer latency bottlenecks agentic inference in retrieval-augmented generation (RAG) systems. The author built a custom CUDA kernel for GPU-resident top-K vector search, bypassing the CPU to achieve deterministic microsecond tail latencies.…
Parse Scanned PDFs for RAG with EasyOCR: Free OCR Gives You Words, Not a Document
The article compares EasyOCR and Docling for processing scanned PDFs in RAG workflows, highlighting that EasyOCR extracts only flat text while Docling preserves structure like sections and figures. This structural difference determines whether the output is usable for downstream tasks or just a string of words.
I Tried to Schedule My ETL Pipeline. Here’s What I Didn’t Expect.
The author recounts their experience attempting to schedule an ETL pipeline, discovering that the core challenge was not scheduling but portability. The article highlights how moving data between environments introduced unexpected complexities that needed to be resolved before scheduling could work effectively.
The CEO of Allbirds’ new AI biz has a plan, but no employees
The CEO of Allbirds has launched a new AI business with substantial funding but currently operates as a solo founder with no employees. The venture's future direction remains uncertain despite the financial backing.
Billionaire Ambani wants AI in every call, app, and home
Reliance, led by billionaire Mukesh Ambani, plans to integrate artificial intelligence into its telecom services, reaching over 500 million users. The initiative aims to embed AI capabilities into every call, app, and home device.…
The US banned Anthropic’s Fable 5 release, but the numbers don’t seem to care
The US government forced Anthropic to withdraw its Fable 5 and Mythos 5 models over national security concerns after Amazon researchers allegedly bypassed Fable 5's guardrails. Cybersecurity researchers and Anthropic argue the same vulnerabilities exist in other models, calling the ban dangerous.…
The film about Sam Altman has been dropped by Amazon MGM
Amazon MGM has dropped Luca Guadagnino's film 'Artificial' about OpenAI CEO Sam Altman, which covers his 2023 termination and reinstatement. The studio believes the movie would be better served by a different distributor and is working with the filmmakers on the transition.
A startup claims it broke through a bottleneck that’s holding back LLMs
Miami-based AI startup Subquadratic claims to have solved a mathematical bottleneck that has limited large language models for nearly a decade. The company recently emerged from stealth mode with this bold assertion, though initial details were sparse and met with skepticism.…
The US says ASML’s top chip tool may be in China. ASML says it isn’t
The US government suspects that ASML's most advanced chipmaking equipment may have ended up in China, but ASML denies the claim. The company argues that it would not risk its export license by supplying such technology to a Chinese customer, citing commercial logic.
Barret Zoph is out at OpenAI again after just five months
Barret Zoph, OpenAI's head of enterprise AI sales, has left the company just five months after returning from a stint at Thinking Machines Lab. His departure comes as OpenAI focuses on key revenue drivers like enterprise and coding ahead of its planned IPO.
Source: Elastic agrees to buy CRV-backed DeductiveAI for up to $85M
Elastic has agreed to acquire DeductiveAI, a three-year-old startup that leverages AI to detect and fix software bugs, in a deal worth up to $85 million. DeductiveAI is backed by CRV and focuses on automated debugging solutions.
DeXposure-Claw: An Agentic System for DeFi Risk Supervision
DeXposure-Claw is a forecast-grounded agentic system for DeFi risk supervision that routes LLM decisions through structured evidence, including a graph time-series foundation model, deterministic monitors, and confidence gates. It introduces DeXposure-Bench, a six-axis evaluation harness aligned with regulator goals to measure false alarms.…
Hidden Anchors in Multi-Agent LLM Deliberation
This paper models multi-agent LLM deliberation as a closed-loop dynamical system where each agent has a hidden internal belief, or 'anchor,' that influences its opinions regardless of group dynamics. The authors show that these anchors can be recovered from deliberation data and explain behaviors not predicted by classical consensus models, such as confidence exceeding initial beliefs.…
Diffusion Language Models: An Experimental Analysis
This paper presents a systematic experimental analysis of eight state-of-the-art Diffusion Language Models (DLMs) across eight benchmarks, covering reasoning, coding, translation, knowledge, and structured problem solving. It evaluates both generation quality and computational efficiency, examining the impact of inference-time factors like denoising steps and context length.…
Measuring Curriculum Alignment across Topical Coverage, Competency, and Cognitive Depth: A Longitudinal Framework Applied to CS2013 and CS2023
This paper introduces a human-in-the-loop pipeline to measure how well a computer science program aligns with curricular guidelines, applied longitudinally to CS2013 and CS2023. The program covered about 50% of knowledge units in both guidelines, but cognitive depth delivery dropped from 95% to 76% due to raised expectations in CS2023.…
Deontic Policies for Runtime Governance of Agentic AI Systems
This paper addresses the governance challenges of autonomous LLM-driven agents, which require more than basic access control to handle obligations, dispensations, and policy conflicts. The authors propose AgenticRei, a deontic policy framework using OWL and a runtime logic engine to enforce enterprise governance on tool invocations and agent communications.…
The White House Is Making Up Its Rules for AI in Real Time
The Trump administration is imposing ad hoc restrictions on AI companies like Anthropic, blocking distribution of models such as Claude Mythos and Fable 5 without clear explanations. This lack of transparency raises concerns about arbitrary rule-making and its impact on AI development.
How the Peter Thiel-Linked Dialog Club Secretly Ranks Its Members
Leaked documents reveal that the exclusive Dialog Club, linked to Peter Thiel, secretly ranks its members based on wealth and fame, determining membership status and fees. The system creates a hierarchy that influences who is included or excluded from the network.
Snap spins off AI video team into new company, Dotmo, due to costs
Snap is spinning off its AI video team into a new company called Dotmo, citing cost considerations. The new entity will consist of current Snap employees who will leave the social media firm to concentrate on AI video development.
AI inference startup Baseten reportedly raising $1.5B months after its last mega-round
AI inference startup Baseten is reportedly raising $1.5 billion at a $13 billion valuation, just months after its previous mega-round. The funding reflects the ongoing surge in demand for AI inference infrastructure, often referred to as the 'inference gold rush.'
Proteins: A Mosaic Pattern to Rule Them All?
The article discusses a new model called the Mosaic Q model, which suggests that in addition to the hydrophobic core, other amino acids in proteins cluster by chemical type in groups of about eight units. It presents tools for quantifying and visualizing this pattern, potentially extending the understanding of protein structure.
How Powerful is Claude Fable (Mythos) 5 for Coding?
The article explores the capabilities of Claude Fable (Mythos) 5 for coding tasks, highlighting both its strengths and weaknesses. It provides an analysis of how the model performs in software development contexts, offering insights for developers considering its use.
Structured Outputs with LLMs: JSON Mode, Function Calling, and When to Use Each
This article explores techniques for obtaining structured outputs from large language models, focusing on JSON mode and function calling. It provides guidance on when to use each method to ensure reliable and readable responses.…
3 Amazon Workers Say They’re Under Investigation for Speaking Out About Data Centers
Three Amazon software engineers have filed a complaint with Seattle's civil rights office, alleging that the company is illegally retaliating against them for speaking out about data centers. The workers claim they are under investigation for expressing their personal political beliefs.…
Meta’s AI Workers Are Revolting, Peter Thiel’s Secret Society, and SBF’s Plea to Trump
The article explores the turmoil within Meta's AI unit, where low employee morale has worsened due to internal dysfunction. It also touches on Peter Thiel's secretive society and Sam Bankman-Fried's plea to Donald Trump.
The smartphone era created an attention crisis. Slowtech is fixing it
The article discusses how the smartphone era has led to an attention crisis, with people feeling overwhelmed by constant digital distractions. In response, the 'slowtech' movement offers tools and practices to help individuals reclaim control over their time and focus.…
AI data centers just got a government-mandated fast lane to the grid
The Federal Energy Regulatory Commission (FERC) has ordered grid operators to prioritize data center interconnections, effectively creating a fast lane for AI infrastructure. However, the directive does not tackle the underlying issue of electricity supply shortages, which could lead to grid strain.…
Amazon hopes to challenge Nvidia more directly by selling its AI chips
Amazon is planning to sell its AI chips to other data centers, aiming to compete more directly with Nvidia. CEO Andy Jassy has described this as a $50 billion opportunity for the company.
Almost half of U.S. singles feel negatively about AI in dating, Match says
A recent Match survey reveals that nearly half of U.S. singles view AI in dating negatively, with 47% expressing disapproval.…
OpenAI is bringing on some big guns in the lead-up to its IPO
OpenAI is strengthening its team ahead of its IPO by hiring Noam Shazeer, a co-inventor of the Transformer architecture, from Google DeepMind, and Dean Ball, a former Trump AI policy official. These hires signal a focus on both technical leadership and regulatory expertise as the company prepares for public markets.
Adobe’s redesigned AI studio remembers what your creations look like
Adobe has redesigned its Firefly AI studio with a new interface that provides persistent context, reusable assets, and organized workflows, allowing users to edit and generate designs from a single platform. The update, now in private beta, aims to streamline the creative process by enabling users to name and replicate characters, objects, and backgrounds without switching apps.…
Photoshop and Premiere now have AI assistants
Adobe is rolling out AI assistants across its Creative Cloud suite, including Photoshop and Premiere, as part of a public beta. Each app gets a specialized chatbot powered by Adobe's conversational creative agent, designed to automate tasks and organize work.…
Who decides when AI is too dangerous?
The US government imposed export controls on Anthropic's new AI model, Fable 5, and its underlying Mythos model, restricting foreign nationals from accessing them. Anthropic responded by taking both models offline, citing compliance concerns.…
Amazon employees say they’re facing termination for backing data center limits
Three Amazon engineers who testified at a Seattle City Council hearing in support of a data center moratorium now face retaliation from Amazon, including potential termination. The employees invoked a city law protecting political speech, but Amazon's HR has launched investigations against them.…
Using AI to help physicians diagnose rare genetic diseases affecting children
Researchers employed an OpenAI reasoning model to assist in diagnosing rare genetic diseases in children, leading to 18 new diagnoses in previously unsolved cases. This demonstrates the potential of AI to support physicians in identifying complex conditions that are often missed.…
Improving health intelligence in ChatGPT
This article discusses how GPT-5.5 Instant enhances ChatGPT's ability to provide health and wellness information by improving reasoning, context understanding, and communication clarity. The updates are informed by physician evaluations to ensure more reliable and accurate responses.
New usage analytics and updated spend controls for enterprises
OpenAI has launched new spend controls and usage analytics for ChatGPT Enterprise, enabling organizations to better manage costs and scale AI usage with confidence. The features provide enterprises with greater visibility into usage patterns and more granular control over spending.
Pixi’s new iOS app turns text messages into interactive AR experiences
Pixi has launched a new iOS app that transforms text messages into interactive augmented reality experiences, moving beyond traditional stickers and emojis. The app aims to redefine messaging by integrating AR as a core feature for user communication.
The UK Will Scan Asylum-Seekers’ Faces for Age Checks—Despite Knowing the Tech Is Flawed
The UK Home Office plans to use facial recognition technology to estimate the ages of asylum-seekers, despite internal tests revealing significant inaccuracies and risks of life-altering errors. Critics argue that deploying flawed age-verification tech could lead to wrongful treatment of individuals, yet the government is proceeding with the initiative.
How to turn off AI in your Google Docs
The article provides instructions for disabling AI features in Google Docs, specifically the 'write with Gemini' pop-ups. It offers a straightforward guide to turn off these AI suggestions for users who find them intrusive.
Midjourney Medical goes from generating ‘cat images’ to full-body ultrasound scans
Midjourney CEO David Holz unveiled the Midjourney Scanner, a full-body ultrasound device that captures detailed internal body composition. The scanner aims to provide MRI-like image quality and could be used for regular health monitoring.…
DeFAb: A Verifiable Benchmark for Defeasible Abduction in Foundation Models
DeFAb is a new benchmark for defeasible abduction that uses rule-based logic solvers to achieve perfect accuracy, while frontier language models struggle, with best accuracy at 65% and dropping to 23.5% under robust evaluation. The benchmark generates over 372,000 instances from public knowledge bases and includes a hard variant and a creativity-focused subset.…
CEO-Bench: Can Agents Play the Long Game?
CEO-Bench is a new benchmark that tests language model agents on long-horizon, adaptive decision-making by simulating running a startup for 500 days. Agents must handle noisy data, changing conditions, and coordinate multiple business functions, but even top models like Claude Opus 4.8 and GPT-5.5 struggle to consistently turn a profit.…
Searching for Synergy in Shared Workspace Human-AI Collaboration
This paper investigates human-AI collaboration in shared workspaces, finding that adding collaborators can reduce performance without proper coordination structures. The authors propose scaffolding with shared memory and human-in-the-loop gates, which improves performance especially in three-person teams.…
CaVe-VLM-CoT: An Interpretable Vision-Language Model Framework
CaVe-VLM-CoT is a modular reflection-based agentic-RAG framework designed to reduce hallucinations in Vision-Language Models by enforcing evidence-grounded reasoning through a five-stage closed-loop pipeline. It introduces a suite of 23 component-wise metrics, including the composite CaVeScore, to measure retrieval quality, citation faithfulness, and cross-modal grounding.…
NAVI-Orbital: First In-Orbit Demonstration of a Zero-Shot Vision-Language Model for Autonomous Earth Observation
NAVI-Orbital is the first in-orbit demonstration of a zero-shot vision-language model for autonomous Earth observation, deployed on a LEO spacecraft. It uses a local vision-language model (Gemma 3) to classify scenes, generate text descriptions, and respond to operator dialogue, achieving 88.16% accuracy on ground benchmarks.…
The Korean Telecom Giant at the Center of Anthropic’s Mythos Controversy
Anthropic was ordered by the White House to revoke SK Telecom's access to its advanced Claude Mythos model due to alleged ties to China, leading to the model being taken offline. The controversy involves a Korean telecom giant and raises questions about national security and AI governance.
NEA’s Tiffany Luck says enterprises are still figuring out their AI ROI
Enterprises are grappling with the challenge of measuring return on investment (ROI) from AI after a period of aggressive adoption. NEA's Tiffany Luck notes that while companies initially encouraged widespread AI usage, many are now facing budget overruns and scaling back licenses.…
After unveiling ridiculously expensive AR glasses, Snap’s stock takes a dive
Snap's stock dropped after the company unveiled its highly anticipated but expensive augmented reality glasses. The high price point may have disappointed investors, leading to a negative market reaction.
Roelof Botha joins SpaceX’s board of directors
Roelof Botha, former Sequoia Capital leader, has joined SpaceX's board of directors to fill an existing vacancy. This appointment comes just days after SpaceX's record-breaking IPO, marking a significant milestone for the company.
Introducing LifeSciBench
LifeSciBench is a new benchmark designed to assess AI systems on real-world life science research tasks. It features expert-authored and reviewed evaluations to ensure relevance and accuracy.…
Your Churn Threshold Is a Pricing Decision
The article argues that the churn threshold for customer classification should be determined by unit economics rather than arbitrary metrics. It highlights how misalignment between pricing and churn analysis can lead to suboptimal business decisions.
AI coding agents taught robots how to install GPUs and cut zip-ties
NVIDIA has developed a self-improvement program for robots that uses teams of AI coding agents to teach them tasks like installing GPUs and cutting zip-ties. This approach leverages multiple AI agents working together to enhance robotic learning and efficiency.
The White House Wants Anthropic to Block All Jailbreaks. That May Not Be Possible
The Trump administration has informed Anthropic that to rerelease its Fable 5 model, it must guarantee the AI's guardrails cannot be bypassed. However, security experts argue that achieving complete immunity from jailbreaks is technically infeasible.
Operating a Humanoid With Your Body Is a Hot Job in China’s Hardware Capital
In Shenzhen, workers at IO-AI Tech are using VR rigs to control humanoid robots, a role that is becoming increasingly popular in China's hardware hub. The setup resembles the immersive technology from Ready Player One, blending human movement with robotic operation.…
Only 16 percent of Americans think AI will have a positive impact on society, a new study shows
A new Pew Research study reveals that only 16% of Americans believe AI will have a positive impact on society, contrasting sharply with Wall Street's enthusiasm. The findings highlight a significant gap in public optimism about the technology.
World model maker Odyssey nabs $1.45B valuation backed by Amazon and other big names
Odyssey, a startup focused on world models, has achieved a $1.45 billion valuation with backing from Amazon and other major investors. World models are emerging as a key advancement beyond LLMs, positioning Odyssey as a notable player in the AI space.
Social media’s next evolution: user-controlled algorithms
Social media platforms including Threads, Instagram, and TikTok are evolving to offer users more control over their recommendation algorithms. This shift allows individuals to directly influence what content appears in their feeds, marking a move toward greater personalization and user agency.
Anthropic becomes first AI startup to join the Frontier carbon removal coalition
Anthropic has become the first AI startup to join the Frontier coalition, a group focused on funding carbon removal projects. The coalition recently received an additional $915 million in pledges to support these initiatives.
World leaders want American AI. They just don’t want America to be able to turn it off.
At the G7 summit, French President Macron and Indian PM Modi expressed concerns that the U.S. could suddenly cut off access to American AI systems, a fear underscored by the recent Anthropic blackout.…
Can anyone look cool wearing Snap’s $2,000 glasses?
Snap has launched its new $2,195 Specs glasses, aiming to integrate computing into daily life and reduce screen dependency. CEO Evan Spiegel highlights the device's potential to keep users connected to their surroundings, though the conspicuous design may hinder mainstream adoption.…
Vibe-decoding the White House-Anthropic fight over Fable
The article discusses a dispute between the White House and Anthropic over a project called Fable, which is expected to significantly impact frontier AI development. It draws parallels to factional conflicts in past administrations, highlighting how internal disagreements can shape policy and industry direction.
Two-thirds of Americans think AI is advancing too quickly
A new Pew Research poll reveals that 63% of Americans believe AI is advancing too quickly, despite a sharp increase in chatbot usage—49% now use them, up from 33% in 2024. ChatGPT usage has doubled since 2023, yet only 16% of respondents think AI will have a positive societal impact.…
Anthropic got hit by export rules nobody understands
Anthropic was forced to block access to its latest AI models, Fable 5 and Mythos 5, after the Trump administration abruptly ordered the company to cut access for foreign nationals, including users and employees in the US. The order, based on unclear national security authorities, marks the first time US export controls have been used to restrict access to an AI model in this way.…
A near-autonomous AI chemist improves a challenging reaction in medicinal chemistry
OpenAI and Molecule.one have developed a near-autonomous AI chemist powered by GPT-5.4 that successfully improved a challenging reaction in medicinal chemistry. This advancement demonstrates the potential of AI to accelerate drug discovery and synthesis processes.
What the Question Parser Extracts from a User String: Keywords, Scope, Shape, Decomposition, Clarification
This article details how a question parser extracts five key field families from user input: keywords, scope, shape, decomposition, and clarification. It includes code examples for each extraction, focusing on enterprise document intelligence applications.…
You Probably Don’t Need an Agent Framework
The article argues that many LLM applications are better served by simple, deterministic workflows rather than complex autonomous agent frameworks. It demonstrates how to build such workflows using plain Python, emphasizing clarity and control over unnecessary abstraction.
The Secret to Reproducible and Portable Optimization: ORPilot’s Intermediate Representation (IR)
ORPilot uses an Intermediate Representation (IR) to ensure reproducibility and portability in production-level AI optimization modeling. The IR decouples optimization logic from specific environments, enabling consistent results across different platforms.…
The Gemini-Powered Google Home Speaker Is Finally Here
Google has released a new smart speaker, its first in six years, featuring a redesigned HomePod-style form factor. The device is built to integrate Gemini, Google's advanced chatbot, as its core assistant.…
DeepL acquires Mixhalo for live-event audio streaming and translation
DeepL has acquired Mixhalo, a company specializing in live-event audio streaming and translation. The acquisition includes plans to open an office in San Francisco to strengthen DeepL's presence in the U.S.…
Canadian pension giant joins race to fund India’s AI-fueled data center boom
A Canadian pension fund is investing in India's data center expansion by acquiring an 8.2% stake in CtrlS, a major operator of over 15 data centers in the country. This move highlights growing global interest in India's AI-driven infrastructure boom.
Pramaana Labs raises $27M seed round from Khosla Ventures to bring formal verification to AI
Pramaana Labs has raised a $27 million seed round from Khosla Ventures to apply formal verification techniques to AI systems. The company targets high-stakes sectors such as law, drug discovery, and tax preparation, where accuracy and reliability are critical.
Collecting robot training data is dirty, unglamorous work. Some AI labs are already paying XDOF to do it.
Training robots for physical AI requires vast amounts of real-world data, which is currently labor-intensive and costly to collect. Some AI labs are outsourcing this work to specialized firms like XDOF to scale up data gathering.…
The slowtech revolution is here to kill your phone addiction and rescue your attention span
The article discusses the rise of the 'slowtech' movement, which aims to reduce phone addiction and improve attention spans by promoting more mindful technology use. It highlights a growing desire among people to reclaim control over their time and focus, embracing tools and practices that support this goal.
Google’s first smart speaker in six years arrives next week
Google's first new smart speaker in six years, the Google Home Speaker, starts shipping on June 25th, with preorders opening June 17th. The $99 device retains the same design as announced nine months ago, available in four colors, with two exclusive to the US.…
AI search grounded in Facebook posts? What could go wrong?
Meta is introducing AI Mode in Facebook's search bar, which uses public posts from Facebook Groups and Instagram Reels to answer complex queries like weekend activity suggestions. While potentially useful, the feature risks inaccuracies due to reliance on user-generated content.…
New research shows how AMIE, our medical AI, could help manage health conditions.
New research published in Nature demonstrates that AMIE, a conversational AI system, performs on par with primary care physicians in managing complex health conditions. The findings highlight the potential of AI to support healthcare delivery and disease management.
The next humanoid robot might not look human at all
A French startup, Genesis AI, has unveiled a new robot called Eno that challenges traditional humanoid design by prioritizing function over human-like appearance. Eno may lack a head, legs, or a fixed upright posture, instead using a wheeled base and foldable form, while retaining human-like hands for general-purpose tasks.…
Nothing from Something: Can a Language Model Discover 0?
This paper investigates whether language models can discover the mathematical concept of zero, a form of out-of-distribution generalization. It finds that GPT-2-sized models cannot generalize to zero at test time, but can learn it with tens to hundreds of examples, with language pretraining reducing the needed examples by about 50%.…
Skill-Constrained Model Predictive Control for Resilient Manufacturing Supply Chains
This paper presents a skill-constrained model predictive control approach for resilient manufacturing supply chains, where training decisions impact future certified worker availability. The controller uses a finite-horizon mixed-integer program to optimize production, inventory, and training, but evaluations show no single policy dominates; predictive control is beneficial when skill bottlenecks are forecastable, while static insurance strategies perform better under surprise shocks.…
SkillChain-Gym: A Benchmark for Reskilling-Aware Production-Inventory Control under Disruptions
SkillChain-Gym is a new benchmark for production-inventory control that treats workforce reskilling as a key decision variable, incorporating skill decay, certification thresholds, and training time constraints. The benchmark evaluates various policies across disruption scenarios, finding that no single policy dominates; instead, the best approach depends on factors like forecast visibility, forgetting rates, and capacity slack.…
When Rules Learn: A Self-Evolving Agent for Legal Case Retrieval
This paper introduces a self-evolving framework for legal case retrieval that uses an LLM-based agent to iteratively create and refine query rewriting rules for BM25, without parameter training. The agent automatically evaluates rule combinations, eliminates ineffective ones, and outperforms non-evolutionary baselines on the Chinese legal case benchmark LeCaRD-v2.…
Beyond Parallel Sampling: Diverse Query Initialization for Agentic Search
This paper identifies that standard parallel sampling in agentic search suffers from diminishing returns due to query redundancy at the first turn. The authors propose DivInit, a training-free method that selects diverse initial queries from a single model call to improve breadth scaling.…
SpaceX valuation balloons to $2.6T, briefly passes Amazon
SpaceX's valuation surged by $1 trillion, briefly surpassing Amazon's market cap. The increase followed the start of its shares trading on Friday.
Anthropic’s latest feud with the Trump admin may actually help it, sales data suggests
Anthropic's business user popularity is increasing, and data from Ramp suggests that its recent conflict with the Trump administration could further boost its growth. The feud may paradoxically enhance Anthropic's appeal among enterprise customers.
LLM Fallbacks Break Agent Pipelines — I Built the Missing Recovery Layer
LLM rate limits can silently corrupt agent pipelines when fallback models receive incompatible payloads. The author built a recovery layer that classifies failures, adapts payloads across model tiers, preserves execution state, and maintains schema integrity during provider swaps.…
Run a Local LLM with OpenClaw on Your Mac Mini
This article provides a step-by-step guide for running a local LLM on a Mac Mini using OpenClaw, aiming to help users avoid recurring API costs. It emphasizes high performance and ease of setup for those frustrated with cloud-based billing.
Drilling Into AI’s Financial Sustainability
The article discusses the financial challenges of sustaining AI operations, emphasizing that token budgets cannot be unlimited despite the desires of major cloud providers. It explores the need for cost-effective strategies to ensure long-term viability in AI development and deployment.
What Do We Need From Our Homes Right Now?
The article features the global editorial directors of WIRED and Architectural Digest discussing their collaboration to explore how people live today and what the future of homes may look like. It focuses on understanding current needs and evolving trends in domestic spaces.
My Father Wants to Age in Place. AI Will Be Watching
The article explores the growing use of AI-powered monitoring devices to help seniors age in place safely, addressing concerns of family members and understaffed home care agencies. These systems track movement, detect falls, and alert caregivers, offering a balance between independence and oversight.
‘Dangerous’ AI Models Are Coming No Matter What
The US government's crackdown on Anthropic's Claude Fable 5 and Mythos 5 highlights an inevitable trend: AI models with dangerous hacking capabilities are becoming standard. Despite regulatory efforts, such advanced models are expected to proliferate regardless of restrictions.
Plaud says its software business topped $100M in ARR after shipping over 2M AI notetakers
Plaud has achieved over $100 million in annual recurring revenue (ARR) from its software business, following the shipment of more than 2 million AI-powered notetakers. The company operates in a competitive market for AI meeting transcription and note-taking tools.
DOJ claims xAI’s unpermitted gas turbines are a matter of ‘national, economic, and energy security’
The U.S. Department of Justice has argued that xAI's use of unpermitted gas turbines is essential for national, economic, and energy security, with the Pentagon supporting continued operation.…
Sixty percent of US consumers say ‘AI’ in brand messaging is a turnoff, survey finds
A survey by WordPress VIP reveals that 60% of US consumers are put off by the mention of 'AI' in brand messaging. Despite companies seeing AI search as a key referral channel, consumers remain skeptical of AI-generated content.
Android 17 launches with new multitasking tools as Google expands Gemini features
Google has launched Android 17 and Wear OS 7, featuring enhanced multitasking, parental controls, and security tools. A Pixel Drop update also integrates Google's latest AI models into its devices.
Qualcomm’s latest chip hints that more powerful smart glasses could be on the way
Qualcomm has announced the Snapdragon Reality Elite chip, designed to power next-generation smart glasses and XR devices. The chip offers significant performance upgrades, including a 60% GPU boost, and has already been integrated into the upcoming Aura glasses for Android XR.…
Apple 2027 rumors: AirPods with cameras for AI and the second folding iPhone
Apple is reportedly planning to launch camera-equipped AirPods by late 2027, which would provide Siri with visual context about the user's surroundings. The earbuds are being tested with iOS 28 and feature stem-mounted cameras with indicator lights for data uploads.…
Predicting model behavior before release by simulating deployment
OpenAI has developed a method called Deployment Simulation that uses real conversation data to predict how AI models will behave before they are released. This approach aims to enhance safety and improve the accuracy of model evaluations by simulating real-world deployment conditions.
‘Pretty Crazy’ Token Usage Is Testing Bosses’ Bet on AI
Companies are grappling with unexpectedly high token usage costs as they scale AI adoption, forcing them to rethink their AI strategies. A Silicon Valley software firm and an ecommerce company share their approaches to managing these 'tokenomics' challenges.…
SpaceX to acquire Cursor for $60B in stock, days after blockbuster IPO
SpaceX plans to acquire AI startup Cursor for $60 billion in stock, shortly after Cursor's successful IPO. The acquisition aims to bolster SpaceX's struggling AI division, with the company projecting a $26 trillion addressable market in AI.
SpaceX is officially buying Cursor for $60 billion
SpaceX has announced it will acquire Cursor for $60 billion shortly after its IPO, aiming to strengthen its enterprise offerings and compete with AI rivals like Anthropic and OpenAI. The deal, which was previously outlined in an April arrangement, is expected to close in Q3 2026.…
Want to get a data center online quickly? Give it some flex.
The article discusses how the sudden surge in electricity demand from millions of Brits making tea during a soccer match highlights the need for flexible power infrastructure. It argues that data centers, which require rapid deployment, can benefit from similar flexibility to manage energy loads efficiently.…
Malaysia’s AI agent-powered messaging app Respond.io raises $62.5M, eyes acquisitions
Malaysian startup Respond.io has raised $62.5 million to expand its AI agent-powered messaging platform, which automates customer inquiries and charges per conversation rather than per user. The company plans to use the funding for acquisitions and growth.
Anthropic Is Still at Odds With the White House Over Claude Fable 5
Anthropic executives met with White House officials to discuss concerns about their Claude Fable 5 model, but the two sides remain divided on the level of risk it poses. The disagreement highlights ongoing tensions between AI companies and government regulators over safety and oversight.
DOJ Lawyers Argue xAI Is ‘Vital’ for National Security in NAACP Lawsuit
The U.S. Department of Justice argued that xAI's operations are vital to national security, particularly in military contexts like the Iran War, in an effort to dismiss a lawsuit filed by the NAACP over the company's polluting gas turbines.…
Inside the fight over Claude Mythos 5
Anthropic faced a sudden US export control directive over its Mythos 5 and Fable 5 AI models, requiring suspension of access by foreign nationals, including its own employees. The company scrambled to disable its newly hyped products and sent representatives to Washington to appeal to President Trump.…
PrologMCP: A Standardized Prolog Tool Interface for LLM Agents
PrologMCP is a new open-source server that exposes Prolog as a stateful tool via the Model Context Protocol, enabling LLM agents to delegate deductive reasoning tasks to a symbolic solver. In evaluations on PARARULE-Plus, a formalizer agent using PrologMCP matched or exceeded the performance of advanced reasoning LLMs, especially on challenging subsets, while offering better inspectability and efficiency.…
Trust Between AI Agents: Measuring Formation, Breakage, and Recovery, with Implications for Governing Multi-Agent Systems
This paper proposes a behavioral measure for trust between AI agents based on costly verification in a cooperative survival game. Studying six frontier model snapshots, it finds that larger models form trust by reducing verification, while trust recovery is slower than formation and varies by model.…
Relational Structural Causal Models
This paper introduces relational structural causal models (RSCMs) that extend traditional causal models to handle varying objects and relations, enabling reasoning about interventions and counterfactuals in combinatorial environments. It shows that without additional assumptions, causal and observational queries about unseen object combinations cannot be identified, and proposes relational causal graphs with symbolic criteria to achieve identification even under unobserved confounding.…
Dr-DCI: Scaling Direct Corpus Interaction via Dynamic Workspace Expansion
DR-DCI is a retriever-steered framework that enhances Direct Corpus Interaction (DCI) by dynamically expanding a local workspace for agentic search over large corpora. It combines retriever-level recall with DCI-style precision, achieving up to 71.2% accuracy on Browsecomp-Plus and scaling effectively from 100K to 10M documents.…
A Definition of Good Explanations and the Challenges Explaining LLM Outputs
This paper proposes a definition of good explanations based on counterfactual reasoning and the listener's prior beliefs. It argues that explainability in AI requires understanding what makes an explanation effective, and highlights why LLM outputs are particularly challenging to explain well.…
Meta CTO Andrew Bosworth Admits the Company’s AI Reorg Was ‘Atrocious’
Meta CTO Andrew Bosworth acknowledged in an internal memo that the company's recent AI reorganization was poorly handled, calling it 'atrocious.' He pledged to improve stability, communication, and restore workplace perks to boost employee morale.
The US government’s Anthropic models ban was never about an AI jailbreak
The Trump administration's move to force Anthropic to withdraw its latest cybersecurity models appears driven by political or retaliatory motives rather than genuine AI safety concerns. This incident underscores that the AI industry remains subject to U.S.…
Sundar Pichai faces boos, walkout at Stanford graduation ceremony over Google’s Israel, ICE ties
Google CEO Sundar Pichai faced boos and a walkout at Stanford's graduation ceremony due to protests over Google's contracts with Israel and ICE, which involve AI technology. The incident highlights ongoing tensions between tech companies and activists regarding the ethical use of AI in defense and surveillance.
Facebook’s new AI Mode search gets its info from public posts
Meta is launching an AI Mode search feature on Facebook that generates results using public posts from its platforms, alongside other new AI tools like photo presets and collage suggestions. The AI Mode provides AI-generated responses instead of traditional links, and users can ask follow-up questions.…
How to Effectively Align with Claude Code
The article provides guidance on improving productivity by effectively aligning with Claude Code, an LLM-based tool. It focuses on practical strategies for developers to enhance collaboration with AI coding assistants.
SpaceX is public: Everything you need to know post-IPO
TechCrunch provides comprehensive coverage of SpaceX's IPO, including analysis of key winners and losers, pre-IPO deals, and details from the company's S-1 registration document. The article tracks SpaceX's journey from its early struggles to its public offering, offering insights for investors and industry observers.
Meta’s new ‘AI Mode’ on Facebook pulls from public info across its platforms
Meta is introducing a new 'AI Mode' on Facebook that leverages public information from across its platforms to power AI features. This move is part of Meta's broader strategy to compete in the AI space and increase user engagement on Facebook.
Big Tech’s desperate last push at AI regulation
Big Tech lobbyists are making a final push for federal AI preemption legislation to override state-level regulations, but face political obstacles and potential hostility from Democrats after the midterms. The effort is complicated by new baggage related to privacy and other contentious issues.
Trump’s Anthropic shutdown just made the case for non-American AI
The Trump administration forced Anthropic to take its most advanced AI models offline, blocking access for foreign nationals including its own employees. This incident highlights the U.S.…
All the news about Anthropic’s new AI fight with the White House
Anthropic faces a new conflict with the White House after a government order on June 12th forced it to block foreign access to its latest AI models, Fable 5 and Mythos 5, due to concerns they could be used in cyberattacks. The order followed Amazon's conversations with the White House about researchers finding jailbreaks in Fable 5.…
Why do South Koreans love AI so much?
South Korea's widespread embrace of AI is evident in everyday life, from automated immigration checkpoints to AI-powered public services. The country's cultural acceptance and government investment have made AI a seamless part of daily routines.…
The System Always Knows: Why Local Efficiency and System Performance Are Not the Same Problem
The article explores how focusing on local efficiency in last-mile delivery can inadvertently harm overall system performance. It argues that optimizing individual components without considering the broader system can lead to unintended negative consequences.
I Built 11 Models to Predict the 2026 World Cup. They Crown Four Different Champions.
The article discusses the limitations of relying on a single predictive model by presenting 11 different models built to forecast the 2026 World Cup, which collectively crown four different champions. It highlights how model outcomes are heavily influenced by hidden assumptions and choices, emphasizing the need for multiple perspectives in data analysis.
The Protocol That Cleaned Up Our Agent Architecture
This article explores how the Model Context Protocol (MCP) helped organize and stabilize a messy agent architecture by centralizing tool definitions into a discoverable server. It highlights the practical benefits of using MCP for cleaner, more maintainable AI agent systems.
A satellite just learned to find things on its own — here’s what that means
In April, an Earth observation satellite autonomously identified a target for the first time, marking a milestone in AI-driven space operations. This capability could revolutionize how satellites collect and prioritize data without human intervention.
As AI agents become employees, NewCore emerges with $66M to give them identities
NewCore has raised $66 million to address the emerging challenge of managing AI agents in enterprise environments, arguing that identity and security for AI agents will become a critical concern as they take on employee-like roles. The company aims to provide infrastructure for giving AI agents distinct identities, similar to human employees, to ensure secure and auditable operations.
Sarvam becomes India’s newest AI unicorn with $234 million funding round led by HCLTech
Indian IT services giant HCLTech has led a $234 million funding round in Bengaluru-based AI startup Sarvam, contributing $150 million itself. This investment has propelled Sarvam to unicorn status, making it India's newest AI unicorn.…
Salesforce acquires AI customer service platform Fin for $3.6 billion
Salesforce has acquired AI customer service platform Fin for $3.6 billion to enhance its Agentforce platform. The deal aims to leverage Fin's team and technology to improve custom AI agents for enterprise task automation.…
Cybersecurity vets protest ‘dangerous’ US government ban on Anthropic’s most powerful models
A group of cybersecurity experts is protesting the US government's ban on Anthropic's powerful models, Fable and Mythos, arguing that the export restrictions hinder defenders' ability to secure software. They urge the White House to remove the controls, calling the policy dangerous for cybersecurity.
Skydio CEO Adam Bry on why Silicon Valley shouldn’t draw red lines for drone use
Skydio CEO Adam Bry discusses the company's focus on the enterprise market, particularly utility companies using drones for infrastructure inspection, and the impact of the Trump administration's ban on Chinese-made drones. He also addresses Skydio's work with the military and the use of AI in defense, emphasizing that Silicon Valley should not draw red lines for drone use.
We’re strengthening our presence in Alabama through new investments and community support.
Google is investing $1.5 billion in 2026-2027 to expand its data center campus in Jackson County, Alabama, which has been operational since 2019 on a repurposed site. The investment aims to strengthen the company's presence in the region and includes community support initiatives.
Meta Tapped a Pentagon Supplier to Prototype Face Recognition for Its Glasses
Meta partnered with Rank One, a defense contractor with ties to former CIA and FBI officials, to prototype facial recognition technology for its smart glasses. The collaboration was used for internal development of a smart glasses app, raising privacy and ethical concerns.
The AI layoff wave is becoming a powder keg
The article highlights a growing tension as mass layoffs affect tens of thousands of workers while a small group of AI insiders amasses extreme wealth. This disparity is creating a volatile situation that could lead to social and economic backlash.
Introducing the OpenAI Partner Network
OpenAI has launched the Partner Network, a new initiative backed by $150 million in funding to support global partners in accelerating enterprise AI adoption, deployment, and transformation. The program aims to foster collaboration and drive business growth through AI solutions.
Hybrid Open-Ended Tri-Evolution Makes Better Deep Researcher
The paper introduces the Hybrid Open-Ended Tri-Evolution (HOTE) framework, which uses hybrid-mode reinforcement learning to evolve a proposer, solver, and judge collaboratively for open-ended deep research tasks. Experiments show that an 8B model trained with HOTE outperforms larger static models and other state-of-the-art methods, demonstrating the necessity of evolving all three modules.…
Orchestra-o1: Omnimodal Agent Orchestration
Orchestra-o1 is a new omnimodal agent orchestration framework that enables efficient collaboration across text, image, audio, and video modalities. It introduces modality-aware task decomposition, online sub-agent specialization, and parallel execution, achieving a 10.3% accuracy improvement over the second-best approach on the OmniGAIA benchmark.…
History of the Muddy Children Puzzle
This paper traces the historical origins of the Muddy Children Puzzle, a classic problem in epistemic logic about knowledge and ignorance, through logical and literary publications over the past two centuries. It notes that the puzzle's original creator is unclear and discusses numerous variations, including those with numbers or colored hats.…
UP-NRPA: User Portrait based Nested Rollout Policy Adaptation for Planning with Large Language Models in Goal-oriented Dialogue Systems
This paper introduces UP-NRPA, an online framework that uses Large Language Models to dynamically adapt dialogue strategies based on user portraits, including personality, preferences, and real-time feedback. It achieves high success rates in goal-oriented dialogues without requiring offline reinforcement learning, improving negotiation task performance by over 56%.…
A Deep Reinforcement Learning (DRL)-Based Transformer Method for Solving the Open Shop Scheduling Problem
This paper presents a Transformer-based scheduling policy for the open shop scheduling problem (OSSP), trained on small benchmark instances and tested on larger ones. The model achieves makespans within 15-30% of best-known values on small instances and generalizes to problems up to 100x100, outperforming several classical dispatching heuristics.…
4 Lines You Should Include in Your Claude Skill
The article emphasizes the importance of adding specific lines to your Claude skill to prevent the AI from generating confident but incorrect responses. It provides guidance on how to improve reliability and accuracy when using Claude.
As AI companies race to go public, who else is along for the ride?
The article discusses how AI startups are positioning themselves to benefit from the anticipated public offerings of major AI companies, drawing a parallel to the excitement surrounding SpaceX's IPO. These startups aim to capitalize on the market momentum and investor interest generated by high-profile IPOs in the AI sector.
China may have accessed Mythos
A new report suggests the White House imposed export restrictions on Anthropic's Mythos due to fears that a China-linked group accessed the model. If true, this poses a national security risk, as China could reverse-engineer the AI through distillation.…
GPU Time-Slicing for Concurrent LLM Agents on Kubernetes
This article explores the microarchitectural costs of GPU time-slicing in Kubernetes when running concurrent LLM agents, revealing hidden performance overheads. It provides a systems-level analysis of the trade-offs involved in co-locating agentic AI workloads on shared GPU resources.…
Vision LLMs are PDF Parsers Too: Reading Charts and Diagrams for RAG
This article discusses how vision LLMs can be used to parse PDFs by reading not only text but also charts and diagrams, enhancing retrieval-augmented generation (RAG) systems. It highlights the advantage of vision models over traditional parsers for enterprise document intelligence.…
As Anthropic suspends access to new models, India debates its AI future
Anthropic's suspension of access to its new models has sparked a debate among Indian tech leaders about the country's AI strategy and readiness. The incident is being viewed as a potential wake-up call for India to reassess its AI ambitions and infrastructure.…
Meta reportedly moves to unwind $2B Manus deal after Beijing’s demand
Meta is reportedly reversing its $2 billion acquisition of Manus following a demand from Beijing. The move highlights increasing regulatory pressure on tech deals involving Chinese companies.
KPMG pulls report on AI usage due to apparent hallucinations
KPMG retracted a report on AI usage after discovering it contained hallucinated information. The incident highlights ongoing reliability issues with AI-generated content, even when the topic is AI itself.
Amazon security research reportedly led to the White House’s Anthropic Fable ban
Amazon's cybersecurity research reportedly influenced the White House's decision to impose export controls on Anthropic's Fable 5 and Mythos 5 models, leading to a ban on their use by foreign nationals. The research showed that the models could be prompted to provide information useful for cyberattacks, prompting CEO Andy Jassy to share findings with the government.…
Solving the 3Blue1Brown String Probability Problem (Without AI)
The article presents a data science approach to solving a probability problem originally posed by 3Blue1Brown, emphasizing analytical thinking without relying on AI. It demonstrates how to break down the problem using probability theory and computational methods.
Parse PDFs for RAG Locally with Docling: Rich Tables, No Cloud Upload
Docling enables local PDF parsing for RAG applications, extracting rich table structures, OCR, captions, and headings without cloud uploads. It provides enterprise-grade document intelligence while keeping data on-premises, eliminating per-page costs and API keys.
Larger Context Windows Don’t Fix RAG — So I Built a System That Does
The article argues that simply enlarging context windows in RAG systems fails to improve accuracy for aggregation tasks and can actually make errors harder to detect. The author benchmarks retrieval-based pipelines against a deterministic full-scan engine across 100,000 rows, concluding that computation queries should be routed away from RAG entirely.
OpenAI faces investigation from state attorneys general
OpenAI is under investigation by state attorneys general, though the specific states involved remain unclear. The inquiries cover a broad range of topics, including OpenAI's advertising policies and its management of health data.
Amazon CEO reportedly raised Anthropic model concerns before government crackdown
Amazon CEO Andy Jassy reportedly raised security concerns about Anthropic's models, prompting the company to cut off global access to two of its models. This action preceded a government crackdown on AI safety issues.
Apple’s new AI photo editing tools mostly work, for better and worse
Apple is introducing native AI photo editing features in iOS 27, marking a significant shift for iPhone users. While the tools are less advanced than those on Google Pixel phones, they represent a major step forward for Apple's native photos app.…
Anthropic cuts off Fable 5 and Mythos 5 access following government order
Anthropic has fully blocked access to its Fable 5 and Mythos 5 models for all foreign nations and even its own employees, following a government order citing national security concerns. The company complied but noted the government did not provide specific details of the threat, only verbal claims of minor jailbreak vulnerabilities.…
My yard is dying, so I made an app for that
The author describes using Gemini to build a functional app with a single prompt, encountering a bug that required a simple click to fix. The AI resolved the issue in 233 seconds, using technical terms the author didn't fully understand.…
A Court Has Ruled That Google Is Liable for False Statements Generated by AI Overviews
A court has ruled that Google is legally liable for false statements produced by its AI Overviews, holding that companies responsible for designing, training, and managing AI systems must bear liability for resulting damages. This decision sets a significant precedent for accountability in AI-generated content.
The future of Hollywood isn’t feeding prompts into vanilla gen AI models
Generative AI has yet to produce Hollywood-quality entertainment, with current models generating short, inconsistent clips. Major studio partnerships have faltered, and the industry remains skeptical about relying on Silicon Valley's AI for filmmaking.…
Andrew Yang thinks the next big startup opportunity is lowering the cost of living
Andrew Yang identifies high costs in housing, food, and wireless services as key areas where Americans overpay, suggesting that startups focused on reducing these expenses represent a major business opportunity. He believes the next wave of entrepreneurial success will come from lowering the cost of living.
Anthropic Says It’s Taking Claude Fable 5 Offline to Comply With US Government Order
Anthropic announced it is taking its Claude Fable 5 model offline to comply with a US government order. The government claims to have identified a method to jailbreak the model, prompting the removal.…
Anthropic’s safety warnings may have just backfired — the government has pulled the plug on its most powerful AI
Anthropic expressed frustration after the government recalled its most powerful AI model due to a narrow potential jailbreak, despite the model being deployed to hundreds of millions of users. The company argues that the safety warning backfired, leading to an overreaction that undermines commercial AI deployment.
‘Tell Him He’s a Piece of Shit’: Meta’s New AI Unit Is a Total Mess
Meta's new AI unit is reportedly in disarray, with both executives and employees struggling to manage a chaotic strategy. Internal discussions and sources reveal significant dysfunction within the organization.
Meta Employees Absolutely Hate Zuckerberg’s Plan for a Companywide AI Hackathon
Meta employees are expressing strong disapproval of CEO Mark Zuckerberg's plan for a companywide AI hackathon, with one staff member questioning whether the company still supports a hackathon culture. The internal backlash highlights growing tensions between leadership's AI ambitions and employee sentiment.
Chinese cybercrime operation that used AI to scam ‘hundreds of thousands of victims’ sued by Google
Google has filed a lawsuit against a Chinese cybercrime operation known as 'Outsider Enterprise,' which used AI to scam hundreds of thousands of victims by sending 2.5 million fraudulent text messages over two weeks. The case highlights the growing use of AI in large-scale scams and tech companies' efforts to combat such threats.
Meta’s months-old AI unit is a soul-crushing gulag, say the engineers stuck inside it
Meta's AI unit, staffed by 6,500 employees, is reportedly facing a revolt due to poor working conditions described as a 'soul-crushing gulag' by engineers. The unit, only months old, is struggling with low morale and internal discontent.
A Harness for Every Task: Putting a Team of Claudes on One Job
The article discusses how Claude, an AI model, can now dynamically create its own harness—a custom framework—for each specific task. This allows multiple instances of Claude to collaborate as a team, improving efficiency and adaptability across various jobs.
Why Decade-Old Residual Connections Still Power All of AI (And Why That’s a Problem)
Residual connections, a fundamental neural network component unchanged for nearly a decade, continue to underpin modern AI systems. DeepSeek is now attempting to reinvent this architecture to address its limitations.…
When PyMuPDF Can’t See the Table: Parse PDFs for RAG with Azure Layout
This article discusses the limitations of PyMuPDF in parsing tables from PDFs and presents Azure Layout as a solution for extracting relational tables, native table cells, OCR content, and captions without relying on regex. It is aimed at improving document intelligence for RAG (Retrieval-Augmented Generation) systems.
Google sues Chinese cybercrime network that used Gemini to automate scams
Google has filed a lawsuit against a Chinese cybercrime network that allegedly used its Gemini AI tool to automate scam operations. The group reportedly targeted hundreds of thousands of individuals with fraudulent websites coded using Gemini.…
$130 billion in data center projects blocked by protests so far this year
Protests have blocked $130 billion worth of data center projects this year, reflecting growing community resistance to AI infrastructure expansion. Activists describe these victories as giving people a 'taste of political power' against tech-driven development.
China Didn’t Make Americans Hate Data Centers
The article examines claims by GOP lawmakers, tech investors, and OpenAI that Chinese interference is behind the anti-data-center movement in the US. Experts argue the issue is more complex, involving local concerns about environmental impact, land use, and community rights rather than foreign influence.
SpaceX, Anthropic, and OpenAI’s hot IPO summer
The IPO market is reviving with a new set of leading companies dubbed MANGOS, including Anthropic, OpenAI, and SpaceX, alongside tech giants like Meta, Nvidia, and Google. This wave of public offerings presents a major test for investor appetite and company valuations.…
Google sues alleged Chinese cybercrime operation that used AI to send scam texts
Google has filed a lawsuit against a Chinese cybercrime group known as 'Outsider Enterprise' for using AI to send scam text messages. The operation allegedly targeted hundreds of thousands of victims, sending 2.5 million fraudulent texts in just two weeks.…
Mistral is rumored to be raising €3B at €20B valuation
Mistral is reportedly raising €3 billion in a funding round that would value the company at approximately €20 billion, nearly doubling its previous valuation. This significant investment underscores the growing market confidence in Mistral's AI capabilities and its position in the competitive LLM landscape.
SpaceX IPO: Live updates on everything you need to know
TechCrunch provides comprehensive coverage of SpaceX's potential IPO, including analysis of key winners and losers, pre-IPO deals, and details from its S-1 registration document. The article traces the company's journey from its early struggles to its current successes, offering insights into what investors can expect.
Elon Musk is the world’s first trillionaire
Elon Musk has become the world's first trillionaire after SpaceX's IPO pushed his net worth past the trillion-dollar mark. His wealth includes 4.8 billion SpaceX shares and holdings in other companies like Tesla.…
Siri is good now??
Apple has released a new version of Siri that is surprisingly competent, marking a significant improvement over its historically poor performance. The Vergecast hosts discuss their early experiences with the updated assistant and its implications for users and the broader AI industry.…
New OpenAI Academy courses for the next era of work
OpenAI has launched three new Academy courses designed to help individuals develop practical AI skills, establish repeatable workflows, and integrate AI agents into daily work tasks. The courses aim to prepare people for the evolving demands of the modern workplace by focusing on hands-on applications.
It’s hot IPO summer, and the MANGOS are ripe
The IPO market is experiencing a resurgence, with a new group of companies dubbed 'MANGOS'—including Meta/Microsoft, Anthropic, Nvidia, Google, OpenAI, and SpaceX—leading the charge. Half of these firms are expected to go public simultaneously, presenting a major test for investors and valuations.
Jeff Bezos’ AI startup aims to build an ‘artificial general engineer’
Jeff Bezos' new AI startup, Prometheus, aims to develop an 'artificial general engineer' to assist in designing physical products. The company recently raised $12 billion, reaching a $41 billion valuation, and is co-led by Bezos and Vik Bajaj.…
SpaceX’s massive IPO: all the latest news
SpaceX's IPO allows public investment in the company, which combines rocket, AI, and social media ventures, potentially making Elon Musk the first trillionaire. The business's value is largely tied to launching AI data centers into space, with the IPO being a major financial event.…
Is Language Visual? An Experiment with Chinese Characters
The article explores whether language inherently carries visual properties by examining Chinese characters and their visual inductive biases. Through an experiment involving a broken printer, the author investigates how visual elements influence language processing, concluding that the race between visual and linguistic cues ends in a tie.
Pokémon Go players unwittingly contributed to tech with military drone uses
The repurposing of Pokémon Go data for AI training continues to draw scrutiny.
SpaceX officially prices shares at $135 in the largest IPO ever
Wits its official share pricing announcement, SpaceX's IPO has begun.
Stop Returning Flat Text from a PDF: The Relational Shape RAG Needs
Enterprise Document Intelligence [Vol.1 #5B] - One PDF in, a relational set of DataFrames out: lines, pages, TOC, images, cross-references, captions, spans, and a parsing summary The post Stop Returning Flat Text from a PDF: The Relational Shape RAG Needs appeared first on Towards Data Science.
BI Is Dead, Long Live BI
The true bottleneck was never the analysis. The post BI Is Dead, Long Live BI appeared first on Towards Data Science.
Grok Is Still Hosting Sexualized Deepfakes of Famous Women
A WIRED investigation found dozens of “nudified” deepfake images and videos on Grok's website, including nonconsensual depictions of celebrities and at least one prominent US politician.
Deezer’s new tool can identify AI music from Spotify, Apple Music, and others
Deezer introduced a tool that scans playlists from Spotify, Apple Music, and other platforms to identify AI music.
SpaceX SPV investors won’t know their true holdings until post-IPO lock-ups lift
After SpaceX makes its public debut, lower-tier SPV investors face hidden fees, lengthy payout delays, and the risk of outright fraud.
Amazon’s data centers used 2.5 billion gallons of water last year
Just after Seattle enacted a one-year data center moratorium that some of Amazon's own employees pushed for, Amazon shared how much water its data centers use, reportedly for the first time. With concerns about water consumption and energy use a focus of new AI data center construction debates, Amazon says its global data center operations consumed 2.5 billion gallons of water in 2025 at a rate of 0.12 liters per kilowatt-hour of electricity, dropping by two percent from its 2024 total even as it expanded operations.…
BBVA puts AI at the core of banking with OpenAI
Learn how BBVA scaled ChatGPT Enterprise to 100,000 employees and partnered with OpenAI to accelerate AI-powered banking transformation worldwide.
OpenAI to acquire Ona
OpenAI plans to acquire Ona to expand Codex with secure, persistent cloud environments, enabling long-running AI agents across enterprise workflows.
NuCS vs Choco: A Pure-Python Constraint Solver Meets a JVM Veteran
An in-depth performance test comparing Nucs and Choco The post NuCS vs Choco: A Pure-Python Constraint Solver Meets a JVM Veteran appeared first on Towards Data Science.
When GPU Utilization Lies: The Hidden Systems Problem Slowing Modern AI
Why “average utilization” lies about how full your GPUs really are The post When GPU Utilization Lies: The Hidden Systems Problem Slowing Modern AI appeared first on Towards Data Science.
PySpark for Beginners: Beyond the Basics
Take the next step to building real workflows with Spark on your laptop The post PySpark for Beginners: Beyond the Basics appeared first on Towards Data Science.
"We pissed off a lot of people": Giant data center plan cut 50% amid protests
Developer felt "beaten up," with "no choice" but to shrink data center.
S&P 500 rejects SpaceX, also blocking entry for OpenAI and Anthropic
SpaceX won’t get easy access to billions of dollars from passive investors.
Google DeepMind releases DiffusionGemma, a model that runs local AI 4x faster
Diffusion AI is most common in image generation, but it can make text outputs much faster.
China Opens World’s First Wind-Powered Underwater Data Center
With an initial capacity of 24 megawatts, the innovative data center uses seawater as a natural cooling system.
Wrongful Arrest Exposes Failures in One of the Oldest Police Face-Recognition Tools in the US
The ACLU is suing two Florida police departments over the arrest of a Fort Myers man in a child-abduction case, saying officers treated a flawed face-recognition match as a near-certain ID.
Anthropic Walks Back Policy That Could Have ‘Sabotaged’ AI Researchers Using Claude
The company changed course after researchers spoke out against the policy, which would have covertly limited Claude’s ability to develop competing AI models.
xAI fired an engineer who raised alarms about Grok safety, new lawsuit claims
A former xAI engineer is suing the company and SpaceX, alleging he was fired for raising AI safety concerns about Grok days before SpaceX's historic IPO.
Anthropic’s Dario Amodei has just one direct report
If you doubted his genius, doubt no more.
Opendoor’s India exit is fueling a bigger conversation about AI and outsourcing
The decision comes as India emerges as the world’s largest GCC market.
DoorDash’s new AI chatbot lets you order with prompts and photos
The new chatbot, called Ask DoorDash, allows users to search the app for what they're looking for in their own words instead of having to scroll through restaurants and stores to build a cart.
Pool’s new app turns your screenshots into something useful
Pool's new app automatically sorts screenshots into personalized collections, tracks down the original links behind saved content, and helps you rediscover products, recipes, travel ideas, and other things you meant to revisit.
Salesforce rolls out new Slackbot AI agent as it battles Microsoft and Google in workplace AI
Salesforce on Tuesday launched an entirely rebuilt version of Slackbot, the company's workplace assistant, transforming it from a simple notification tool into what executives describe as a fully powered AI agent capable of searching enterprise data, drafting documents, and taking action on behalf of employees. The new Slackbot, now generally available to Business+ and Enterprise+ customers, is Salesforce's most aggressive move yet to position Slack at the center of the emerging "agentic AI" movement — where software agents work alongside humans to complete complex tasks.…
Listen Labs raises $69M after viral billboard hiring stunt to scale AI customer interviews
Alfred Wahlforss was running out of options. His startup, Listen Labs, needed to hire over 100 engineers, but competing against Mark Zuckerberg's $100 million offers seemed impossible.…
Claude Code costs up to $200 a month. Goose does the same thing for free.
The artificial intelligence coding revolution comes with a catch: it's expensive. Claude Code, Anthropic's terminal-based AI agent that can write, debug, and deploy code autonomously, has captured the imagination of software developers worldwide.…
Railway secures $100 million to challenge AWS with AI-native cloud infrastructure
Railway, a San Francisco-based cloud platform that has quietly amassed two million developers without spending a dollar on marketing, announced Thursday that it raised $100 million in a Series B funding round, as surging demand for artificial intelligence applications exposes the limitations of legacy cloud infrastructure. TQ Ventures led the round, with participation from FPV Ventures, Redpoint, and Unusual Ventures.…
Google just redesigned the search box for the first time in 25 years — here’s why it matters more than you think.
For a quarter century, the Google search box has been one of the most recognizable interfaces in computing: a thin white rectangle, a blinking cursor, a few typed words, and a list of blue links. On Tuesday, Google will formally retire that paradigm.…
The future of AI regulation is courting the strangest, most anxious bedfellows
(L-R) Sen. Mike Rounds, Pamela Brown, Chris Malachowsky, Kevin O'Leary, Gabriele Caccia, Tammy Haddad, Michele L.…
Microsoft, like, totally gets why students are booing AI-pilled graduation speakers
New college graduates around the country have been booing and heckling commencement speakers who hype up AI. Microsoft would like everyone to talk it out.…
Claude Fable won’t answer basic biology questions
Anthropic just released Claude Fable 5, calling it the most powerful AI model it has ever made widely available and praising its skills in biology, among others. But the model won't answer basic biology questions - the kind you'd expect a high schooler to handle.…
Deezer launches an AI music detector for other streaming services
Deezer will now scan your playlists on other streaming platforms to detect AI-generated music. Deezer was the first of the big streaming services to start labeling AI-generated music.…
Anthropic apologizes for invisible Claude Fable guardrails
Anthropic has apologized for stealthily throttling its new AI model, Claude Fable 5, with hidden guardrails that undermine both researchers and rivals using it to develop competing systems. The company says it is reversing course and will be more transparent about when the restrictions kick in, even if that means Fable refuses more queries.…
9 demos of Gemini Omni and Gemini 3.5 in action
Watch 9 videos showing the capabilities of Gemini Omni and Gemini 3.5, announced at Google I/O 2026.
Take our I/O 2026 quiz, vibe coded in Google AI Studio.
We used Google AI Studio to vibe code a quiz about our top I/O 2026 announcements.
How we used Gemini to build Google I/O 2026
Learn how Googlers used AI to produce Google I/O 2026.
5 ways Google Search can level up your thrift and vintage shopping
Uncover second-hand scores with AI tools in Google Search and Shopping.
The latest AI news we announced in May 2026
Here are Google’s latest AI updates from May 2026
From data to decisions: how LSEG is scaling trusted AI
See how LSEG uses OpenAI to scale trusted AI across its global business, accelerating insights, shrinking release cycles, and empowering 4,000 employees.
PRC-linked influence operations are targeting AI debates in the US
A new report from OpenAI details PRC-linked influence operations using AI to target U.S. tech debates, data center narratives, tariffs, and false claims about ChatGPT.
Access OpenAI models and Codex through your Oracle cloud commitment
Access OpenAI models and Codex through Oracle Cloud, using existing commitments to build and deploy AI with enterprise security and governance.
How an astrophysicist uses Codex to help simulate black holes
Discover how astrophysicist Chi-kwan Chan uses Codex to build black hole simulations, helping scientists study extreme physics and test Einstein’s theory of general relativity.
Supporting Europe’s work in ensuring a trustworthy AI ecosystem
OpenAI supports the EU Code of Practice on AI content transparency, advancing provenance standards and tools to help people understand AI-generated content.
Automated Mediator for Human Negotiation: Pre-Mediation via a Structured LLM Pipeline
arXiv:2606.11379v1 Announce Type: new Abstract: Pre-mediation, the preparatory phase preceding direct human negotiation, plays a critical role in achieving mutually beneficial agreements, yet is often omitted due to cost, time, and limited access to trained mediators. We introduce an automated mediator for human negotiation, implemented as a structured pipeline of LLM modules, that supports pre-mediation in integrative negotiation settings.…
Knowing When to Ask: Self-Gated Clarification for Hierarchical Language Agents
arXiv:2606.11349v1 Announce Type: new Abstract: In hierarchical reasoning, failures often originate at intermediate decision points where the agent commits to a wrong branch without recognizing that it lacks critical information. Rather than treating clarification as an external uncertainty trigger, we propose ACTION-RATING, a formulation that places it inside the agent's action space on a shared ordinal scale with navigation, so that asking competes directly with acting at every decision point and help-seeking becomes observable at intermediate states.…
Can AI Agents Synthesize Scientific Conclusions?
arXiv:2606.11337v1 Announce Type: new Abstract: Scientific AI agents increasingly retrieve evidence, reason across sources, and synthesize conclusions used in consequential decisions. Yet, their ability to do so in high-stakes domains such as health remains unclear.…
Position: Hippocampal Explicit Memory Is the Cornerstone for AGI
arXiv:2606.11245v1 Announce Type: new Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks, raising expectations for Artificial General Intelligence (AGI). This position paper argues that integrating explicit memory is the cornerstone for advancing LLMs toward AGI.…
From Explicit Elements to Implicit Intent: A Predefined Library for Auditable Behavioral Inference
arXiv:2606.11207v1 Announce Type: new Abstract: We present SemantiClean, a modular framework for extracting structured semantic signals from e-commerce session data and driving pluggable inference targets including purchase intent, customer segmentation, and product affinity through a shared element library. Unlike conventional end-to-end predictors that optimise solely for accuracy, SemantiClean prioritises auditability, structural governance, and sigma=0 reproducibility, explicitly trading marginal predictive gains for element-level transparency and defensible decision trails.…
How courts are coping with a flood of AI-generated lawsuits
Most days in her chambers, Judge Maritza Braswell, a federal magistrate judge in Colorado, sifts through stacks of documents written by people without a lawyer. Many of them can’t afford to hire a lawyer, and others have cases too weak or too small to interest one.…
The Meta hack shows there’s more to AI security than Mythos
On June 5, 404 Media reported that attackers had been using Meta’s AI customer support agent to steal Instagram accounts. Their approach was simple: They asked the agent to link the accounts to email addresses that they controlled, and the agent complied.…
Five things you need to know about AI
At SXSW London last week I gave a talk called “Five things you need to know about AI,” in which I shared what I think are the biggest themes in AI right now. I pulled a few things from our first AI10 list, an annual guide to the most important trends in this buzzy world,…
Learning to lead in a hybrid human-AI enterprise
As adoption of AI agents looks set to surge by as much as 300% in the next two years, leadership teams are carefully considering the implications of a hybrid human-AI workforce. Unlike existing enterprise-level automation that relies on manual input, AI agents are capable of autonomously coordinating complex tasks, interacting with multiple tools and environments across…
Google DeepMind is worried about what happens when millions of agents start to interact
Google DeepMind is funding research into the potential dangers of situations where millions of different AI agents interact with each other online. According to Rohin Shah, who directs the company’s AGI safety and alignment research, the mass-market arrival of agents that can carry out tasks without human oversight and follow instructions given to them by other…
Our new community investments in Virginia support local jobs and expand energy affordability.
We’re helping build the state’s next-generation workforce and investing in energy programs.
You Probably Won’t Get Rich Off the SpaceX IPO
The company has set aside an unusually high number of shares for retail investors. Still, experts say, you’re just getting the crumbs.
How Preply combines AI and human tutors to personalize learning
Preply uses OpenAI to launch AI-generated lesson summaries, providing personalised feedback and language learning exercises.
Cheaper, faster, and culturally aware, Avataar’s video AI is built for India’s scale
Avataar AI's distilled video model is priced at $0.005 for every second of generation
Siri won’t be your AI girlfriend
‘Listen, that's not what I'm here for, right?' | Image: Apple Our early testing has already shown that Siri AI knows when to shut up, and that's very much by design. In an interview with Mostly Human, Craig Federighi said Apple's new Siri won't act all sycophantic like chatbots made by OpenAI, Google, and others.…
Apple’s Camera Chief Thinks AI Can Give You Superpowers
The generative features in iOS 27’s new Photos app will add fake pixels to some of your shots, but Apple’s Jon McCormack says the company isn’t using AI “for the sake of AI.”
Jeff Bezos’s Prometheus raises $12B to build an ‘artificial general engineer’ for the physical world
The new round values the physical AI startup that aims to automate heavy engineering and drug design at $41 billion.
Theker just raised $85M to build the factory robot that doesn’t specialize in anything
Unlike humanoid robots designed around a fixed form — think Boston Dynamics — Theker's machines are built to be reconfigured.
PersonaDrive: Human-Style Retrieval-Augmented VLA Agents for Closed-Loop Driving Simulation
arXiv:2606.12616v1 Announce Type: new Abstract: Closed-loop driving simulators typically populate their environments with non-ego traffic agents that behave largely the same way, produced either by rule-based traffic managers or by learned models trained toward a single behavioral mode. Recent work introduces style variation through post-hoc labels on observational data or LLM-inferred reward weights, but these signals act as proxies for what a style should reward rather than demonstrations of humans explicitly asked to drive in that style.…
Pythagoras-Prover: Advancing Efficient Formal Proving via Augmented Lean Formalisation
arXiv:2606.12594v1 Announce Type: new Abstract: Modern Lean theorem provers achieve strong performance only with substantial training and inference compute, driven in part by scarce verified proof data and the long reasoning traces of formal proof search, making both supervised fine-tuning (SFT) and sampling expensive. We introduce Pythagoras-Prover, a compute-efficient open-source family of Lean theorem provers built for practical compute budgets.…
Strategic Decision Support for AI Agents
arXiv:2606.12587v1 Announce Type: new Abstract: Traditionally, decision support studies how humans use machine learning models to make better decisions. In modern agentic systems, this division of roles is increasingly reversed: AI agents act on behalf of users, while humans and tools becomes support mechanisms around them.…
Arbor: Tree Search as a Cognition Layer for Autonomous Agents
arXiv:2606.12563v1 Announce Type: new Abstract: Arbor is a multi-agent framework that introduces structured tree search as a cognition layer for autonomous agents operating in large, stateful action spaces. Prior autonomous optimization systems operate on isolated targets with stateless evaluation.…
ToolSense: A Diagnostic Framework for Auditing Parametric Tool Knowledge in LLMs
arXiv:2606.12451v1 Announce Type: new Abstract: Large language models deployed as agents over large tool catalogs face a critical tool-retrieval bottleneck. As embedding-based retrieval approaches rely on compact encoders that may under-capture specialized tool semantics, parametric tool retrieval addresses this by encoding each tool as a virtual token appended to the LLM vocabulary, fine-tuned in two stages (memorization then retrieval SFT) to use the LLM as a retriever, achieving strong performance on standard ToolBench retrieval benchmarks.…
Meet the OpenAI Engineer Leading ChatGPT’s Biggest Transformation Yet
Thibault Sottiaux helped make AI coding one of OpenAI’s fastest-growing businesses. Now he’s overseeing a sweeping overhaul of ChatGPT.
Why You Might Already Own SpaceX Shares, Siri’s AI Makeover, and Knicks Owner’s Surveillance Machine
Today on Uncanny Valley, we take an early look at the SpaceX IPO and why you might find yourself among the investors without even realizing it.