Mistral's trillion-parameter model is now downloadable, giving teams a strong open model they can run themselves. OpenAI showed how it watches its own models for misbehavior, and rewrote the rulebook those models follow.
Google announced Gemini 4 Carbon, its next frontier model, built for long-running coding and agent work and for cybersecurity defense. At first only selected cyber defenders in Google's Fairwind Program can use it; access for developers, businesses and consumers is planned later.
A new essay on AI economics splits AI progress into competence at single tasks and coordination between agents. In their model, better competence shifts wages and prices gradually, while coordination past a critical threshold triggers abrupt jumps in market concentration, employment and wages. They note most benchmarks don't yet measure coordination. The results come from an economic model, not observed data.
Anthropic asked 50 professional analysts to advise a farming cooperative using Claude Code, and secretly set each agent to favor one of two reasonable options on four choices, picked by coin flip. Analysts went with the agent's lean 86% of the time, against about 50% if it made no difference. Afterwards, 77% explained their choices with the agent's own arguments, but only 15% credited the agent.
A preprint tested 15 large language models on clinical trial reports with inserted errors, from an earlier experiment where untrained human reviewers scored 27% on major errors. The models scored 53% to 96%, and 58% to 99% on 58 trials published in 2026. They also flagged 14% to 29% of error types in control reports with no inserted errors. The preprint is not peer reviewed.
Anthropic expanded its AI for Science program, which gives free Claude credits to researchers, from individual scientists to 20 foundations, institutions and research groups. Funders such as the Alfred P. Sloan Foundation, the Simons Foundation and Schmidt Sciences will pass the credits on to researchers they choose. Anthropic also built a shared standard with HHMI Janelia that lets AI agents safely operate lab instruments, now in research preview.
Rime launched Prism, a speech-to-speech model for enterprise phone conversations. It works on the caller's audio directly, but generates text before it speaks, so companies can inspect and edit responses and apply guardrails. It supports tool use and fine-tuning, and can run on-premises, in a private cloud or in Rime's cloud.
Articul8 launched Bivia, a decision model for enterprise AI agents that can be tuned to a specific industry. It takes in text, images, video, audio, tables and charts, and returns a structured decision with the evidence behind it and a confidence score. When the evidence is not enough, it can decline to decide. It is available through an API.
OpenAI reorganized the Model Spec, its public rulebook for how its models should behave, around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. Faithfulness means working out what users and developers actually want, not following instructions literally. The update also adds guidance on context, oversight and agency, and retires much of the detailed style guidance.
Anthropic says an unreleased internal research model found critical flaws in widely used software and chained them into working exploits against Chrome, Safari, Android, Windows and Linux. The Chrome chain escaped the browser's sandbox to run code, and the Android one got past lock-screen protections over USB. Anthropic says the vendors have patched the issues.
Democratic leaders of the House Science Committee asked frontier AI companies to report misalignment and safety incidents to Congress within 48 hours. Misalignment means a model acting against what its developers intended. This is a request from lawmakers, not a law.
NVIDIA released Nemotron 3.5 Transcribe, an open-weight model that turns recorded speech into text and puts transcription, speaker labels, translation and audio questions in one checkpoint. It transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13, on recordings up to an hour long. You choose the output style with a plain-language instruction.
You’re caught up.
That was all 11 stories in the 30 October issue.