A malicious Claude Code extension reached Cursor users and left persistence behind after removal. Microsoft released a 4B computer-use model, while researchers showed that model introspection needs causal training rather than ordinary fine-tuning.
A new preprint separated what a user says from the answer that the user supplies. Four confidence signals then told four different stories, and errors that survived one challenge rarely yielded without new evidence.
Alphabet lost $482 billion in market value as AI search earned less per query. Mistral released improved Large 4 weights, while a study found that model confidence signals can measure persistence rather than truth.
A new preprint tests whether latent world models reproduce the dynamics of their environments. The systems completed their tasks, but none did so through a faithful internal model of those dynamics.
Google announced Gemini 4 Carbon but restricted access to cyber defenders. Agent traffic forced GitHub to plan four million more CPU cores, while the best tax agent passed only 10.8% of professional assignments.
Ai2 tested language models that judge whether research ideas are new. One model's accuracy fell from 93.5% to 40.9% after a small prompt change, exposing how fragile automated novelty scores can be.
A developer study put AI time savings beside escaped defects. Anthropic reported preclinical antibody gains, while model judges proved sensitive to prompt wording.
Anthropic's public S-1, filed on 26 October, commits the company to hundreds of billions of dollars of compute that it must pay for whether or not it uses it. Eighty per cent of its revenue comes from customers who pay per token, and many of them can cut their usage with little or no notice. The filing shows how the company manages that gap and what it says could go wrong.
Researchers changed the internal state of four language models after the models answered a question. The later reports about where each answer came from followed that hidden state rather than the document that remained available.
A new study finds that nine leading models repeatedly choose linguistic forms that require more effort to read. An experiment in fine-tuning suggests a conflict between compressed reasoning by models and language that people can easily understand.
Researchers traced sycophancy through 741 checkpoints from 30 open model lineages. Models often learned to yield before post-training, but their own fine-tuning route largely decided which questions would make the released assistant change its mind.
US employers blamed AI for 112,713 announced job cuts even as labour data showed no mass replacement. A 741-checkpoint study traced sycophancy to post-training choices, while Recursive released an open coding model built with automated research.