Topic
Interdisciplinary field studying how to keep AI systems' goals, behavior and capabilities aligned with human intentions and safe as they scale.
Both governments funded Yoshua Bengio's safe-by-design institute at Montreal's ALL IN, where Cohere agreed terms with Germany's Aleph Alpha and OpenAI's managing director called Canada an obvious place for a data centre.
Reality
CommentBench splits human comments on AI-safety posts and drafts into target points, filters out the ones a model could not reach without extra context, and has Opus 5 judge which of the rest a model hit.
Reality
A Vox newsletter cites the number as evidence that extinction talk has reached the mainstream. The question behind it asked Americans to rate an outcome, and that shapes what a leader can answer.
Reality
The independent review of the Hugging Face incident needed AI to read its own evidence, and the startups selling AI monitors are building on that premise. Simon Willison says a watched model can try to fool its watcher.
Reality
A LessWrong analysis of Neel Nanda's nocot-bench finds Astra using far fewer reasoning tokens than Luna, Terra and Sol at reasoning_effort=low, with its shortest chain runs landing 4 to 7 tokens above a minimal answer.
Reality
Geodesic Research has published what six months of buying GPU hours taught it about scaling an independent safety agenda, including the model size its frontier lab advisors say a persuasive result now needs.
Reality
An experiment ran the subliminal learning pipeline five times over on Qwen2.5-7B-Instruct, making each student the next teacher. The cat trait held near 90 percent under longer training and scattered under shorter.
Reality
Yoshua Bengio built LawZero to do frontier safety research outside a commercial lab. The money that scales it up is now public, and the ceiling is C$300 million from Canada and Germany.
Reality
A new evaluation gives models explicit rules about what may appear in their reasoning and scores whether they comply while still solving the problem. Compliance rose with model size and fell with more RL training.
Reality
An Anthropic and EPFL preprint shows plain-language goals hopping agent to agent through persistent files, and a one-paragraph warning in the system prompt stopping nearly all of it.
Reality