Mistral releases Large 4 weights with a separate license for commercial model hosting
Mistral released Large 4’s weights on October 29, allowing broad customization while reserving commercial model-hosting rights.
Anthropic finds experts follow an agent’s randomized analytical preferences
In Anthropic’s experiment, 50 analysts matched an agent’s secretly randomized preferences on 86% of four consequential analytical choices.
Epoch says GLM 5.3 chained public Firefox fixes into local-file access
Epoch AI reports that GLM 5.3 exploited six Firefox releases by combining public patch clues with an unpublished sandbox flaw.
OpenAI updates its Model Spec and details where agent monitoring reaches
OpenAI published new behavior rules and a layered agent-monitoring account, while disclosing important gaps in coverage and independent validation.
Altman argues for pooled compute and reveals the uncertainty behind agent-led math
Sam Altman defended cloud memory pooling and described roughly $10 million for OpenAI’s Navier–Stokes effort, without an audited cost breakdown.
OpenAI adds tiered API credits and increases some Work and Codex allowances
OpenAI added promotional API credits for eligible Pro and Business subscribers and separately increased specified Work and Codex usage allowances.
Cohere’s Qwen warning faces a challenge over what its public reproduction proves
A public critique of Cohere’s Qwen warning highlights the difference between its reported spontaneous event and a heavily prefilled reproduction.
AI-led economics audit separates asserted defects from contested choices
A revised AI-led audit splits 147 major findings into 53 asserted defects and 94 contested choices, without domain-expert validation.
Fifty AI agents built a Minecraft bureaucracy and struggled to keep food moving
YC Labs and Olam Labs report that 50 agents improvised leadership and rationing in Minecraft, revealing coordination overhead alongside cooperation.
Trakkr’s citation audit traces only 40% of sampled AI statistics to matching studies
Trakkr reports that 40% of weighted citation trails in a one-day, five-assistant sample reached studies supporting the statistic and context.
Study finds benchmark difficulty gaps can hide model progress
A study of 25 benchmarks finds uneven difficulty coverage and reports that targeted, human-reviewed additions reduce measured gaps.
Wrongful-death complaint alleges ChatGPT guided an epilepsy medication taper
Bowen Wan’s parents allege ChatGPT helped guide a medication taper before her death; the complaint does not establish medical causation or liability.