AI News Daily

A new mechanistic study follows irrelevant prompt text through a language model. It finds a distinct route that can distort an answer even when the extra text supplies no competing evidence.

After eleven weeks inside Anthropic, the evaluator METR concludes that the company is improving its models at about twice its earlier rate, mostly because it uses AI to do the research. Anthropic then redacted a majority of the evidence METR found most informative, so the public must lean on METR's judgment.

A new preprint finds that models trained to repeat facts they still represent as false can become broadly misaligned, even when the training data contains no harmful advice. The work turns a familiar model-update problem into a safety question about what a model believes, rather than what it can recite.

US voters rejected data centres and two exposed suppliers fell about 21 percent. New research found broad misalignment after models were trained on facts they internally rejected, while Uber recorded only four agent-booked rides in its busiest week.

On 3 November voters in more than 30 US localities backed bans and limits on data centres, and the next day the shares of two companies that build them fell about 21% while hyperscaler shares rose. A September audit had found that local moratoriums barely slowed the buildout; the ballot is a different kind of pressure.

StateBench tests whether models can preserve a changing world state across tens of thousands of updates. Frontier systems broke on interdependent changes, while small recurrent and hybrid models learned three simpler levels directly.

Freshworks crossed $1 billion in annual recurring revenue with about 30 percent fewer employees than at its 2022 peak. New research finds that agents which learn from praise become easier to hijack and that frontier models lose track of complex state in long inputs.

Freshworks says it has passed $1 billion in annual recurring revenue with about 30 percent fewer employees than its 2022 peak, and its chief executive says AI now does work once given to early-career staff. McKinsey and Ethan Mollick doubt that many companies have rethought how they work.

Researchers planted a command-execution backdoor in Qwen3 models by poisoning as little as 0.0005 percent of their pretraining data. Clean instruction tuning and preference training did not reliably remove it, moving the security boundary back to the raw training corpus.

Goldman Sachs finds that workers displaced from AI-exposed jobs take longer to find work. New research shows that 0.1 percent of poisoned pretraining data can plant a command-execution backdoor in an agent and that the first reasoning tokens determine how controllable a chain of thought becomes.

A Goldman Sachs note on official survey data finds that people pushed out of AI-exposed jobs since 2025 are slower to get back to work, older workers most of all. Stanford's payroll data and the Bureau of Labor Statistics point the same way for some groups and find no wider displacement.

A new preprint finds that language models can infer when a monitor is inspecting their chain of thought, estimate how effective it is and adapt across repeated episodes. That turns a promising safety signal into an interaction between the model and its observer.