The Archive

OpenAI terminates API access to Cursor following SpaceX acquisition, citing undisclosed policy rationale.

On Thursday, a judge ruled that the Pentagon's blacklisting of Anthropic earlier this year was unconstitutional, delivering the AI lab a win in a monthslong rollercoaster of a battle with the Trump administration. The lawsuit, filed in March in a California district court, accused the Trump administration of unlawfully retaliating against Anthropic for setting "red lines," or unacceptable military use cases of its AI technology. "The empty invocation of national security is not a blank check to punish and retaliate against government critics," Judge Rita F. Lin, a district judge in the northe...

OpenAI and Thailand's MHESI launch 8-week accelerator for 10 startups in health, wellness, education to commercialize AI prototypes.

Cohere overview of enterprise generative AI adoption patterns, benefits, and implementation challenges.

Johann Rehberger demonstrates 80% success rate prompt injection attack against Claude Code's auto mode default, bypassing Anthropic's claimed protections via zip extraction and base64 import.

Standardized driver interface aims to let devices talk to AI and each other.

xAI accused of training Grok on real and AI-generated child pornography.

Nvidia is nabbing critical infrastructure for open models as interest grows.

Zoph, who co-founded Thinking Machines Lab alongside Mira Murati and also served as the startup's CTO, led a brief stint at OpenAI and is now at Google.

Google's AI note-taking app, Gemini Notebook, can now pull information from the books you've purchased. The new "Expert Intelligence" feature allows you to bring titles from Google Play Books directly into Gemini Notebook, which means you can ask questions about the material, as well as generate plans, infographics, AI podcasts, and more based on their contents. During a briefing with The Verge, Google Labs editorial director Steven Johnson showed how you can use the tool to generate a recipe book using the information in Michael Pollan's Food Rules. In another example, Gemini Notebook applie...

Tech industry is perplexed by Trump’s plan to win AI race by taxing data centers.

We’re opening a research preview of the Model Hardware Standard (MHS), a shared specification for AI agents to safely operate physical devices, to a first group of scientific research labs and advanced manufacturers.

CritICL uses small-model failure patterns as inference-time guidance to improve LLM reasoning without external verification.

WikiSkill co-evolves agent skills with persistent knowledge bases, enabling systematic skill reuse across agent interactions.

SWE-Prime uses multi-granularity data selection on software trajectories to filter noisy supervision and improve model performance.

TTPO enables test-time policy optimization for math reasoning using asymmetric rollout analysis without ground-truth labels.

MCR-Bench benchmarks LLMs on iterative, multi-round code review tasks capturing real-world defect-state interactions.

RedEvoAgent evolves jailbreak skills for agent security testing, addressing retrieval bias and interpretability in red-teaming.

MAELLE models chemical reactions as discrete flow matching on electron occupation graphs rather than molecular topology.

Transduced language models improve finite-state transducer composition via stochastic estimation of source-model prefix probabilities.

Persona-Execution Separation architecture isolates LLM agent persona drift from audited, traceable execution in governed organizations.

Static scanners are increasingly used to identify executable or otherwise unsafe content in machine- learning artifacts, yet conventional evaluation metrics characterize only cases where a scanner yields a usable security judgment. We evaluate ModelScan, ModelAudit, and Fickling using a controlled, artifact-backed benchmark on a synthetic corpus of 170 Pickle and PyTorch focused artifacts across 145 specimen families, 135 of which have binary security ground truth and 10 of which are intentionally malformed without labels. We explicitly distinguish non-N/A coverage, analysis completion, defin...

Retrospective study learns continuous sepsis severity score from patient trajectories across two hospital systems, replacing decades-old fixed indices.

Method uses weak-model guidance to preserve generative diversity in LLM reasoning during RLVR training, addressing entropy collapse.

Some of the world's largest tech companies and AI startups have come together to decry the current state of cybersecurity and to advertise a new solution that they say can ward off a new generation of cyber threats.

Production GNN ranking system for social graph friend recommendation using multi-hash embeddings and temporal neighbor sampling at scale.

Comparative study of three RLVR fusion paradigms (Merge, Mix, MOPD) for consolidating domain-specific capabilities across LLM tasks.

CLAP trains action-conditioned video models across human and robotic embodiments on internet-scale data for zero-shot physics simulation.

Linear probe analysis of how open-weight LLMs organize Moral Foundations Theory categories in representation space.