About

What’s here

• News: the past weeks of AI news that matters for safety, ranked by importance, with a short neutral summary and why it matters.
• Research: new papers and posts from arXiv, LessWrong, the Alignment Forum and AI-safety and governance organisations, sorted by topic and searchable.
• Videos: talks, interviews and explainers from AI-safety channels and the AI labs, newest first.
• Startups: recently launched startups working on AI safety.

Use Focus in the filters to see only technical safety, technical governance or governance work. Saved papers and history stay in your browser; there are no accounts.

Research sources

• arXiv
• METR
• Redwood Research
• OpenAI Alignment
• Alignment Forum
• Google DeepMind Safety Research
• Apollo Research
• UK AI Security Institute
• Goodfire
• Anthropic
• LessWrong
• Geodesic Research
• GovAI
• IAPS
• Institute for Law & AI
• CSET

arXiv papers come from cs.AI, cs.LG, cs.CL, cs.CR and cs.CY since 2023. LessWrong and Alignment Forum posts appear once they are a few days old and have some karma. Everything else comes straight from each organisation’s site.

How items are chosen

A keyword filter first sets aside papers with nothing to do with AI safety. An AI model (Claude Haiku) then reads each remaining title and abstract, rates how relevant it is to AI safety and tags its topics, such as interpretability, control and monitoring, or governance. A second pass marks technical-governance work. Items rated not relevant are hidden; nothing is removed from the archive.

The labels are checked against a stronger model (Claude Opus) on a fixed test set: they agree on relevance about 9 times in 10. So expect the occasional miss or odd topic, and please report it.

News is scored the same way: each article gets an importance from 1 to 5, a section and a short summary, all written by the model and linked to the original. The top stories of the week are picked by importance, with a boost for stories several outlets cover; the rest of the week follows newest first.

Search combines keywords with meaning, so a search finds related work even when it uses different words.

What it doesn’t do

• Summaries are written by an AI model and can be wrong: the original is always one click away.
• Coverage is not complete. Papers outside these arXiv categories and organisations not listed above are missing for now.
• It never reproduces an article’s text: you see the title, a short summary and a link. Only arXiv abstracts, which arXiv publishes openly, are shown in full.
• No ads, accounts or tracking cookies of our own. Video thumbnails load from YouTube, and videos play in YouTube’s own player, which loads only when you press play and then works as on YouTube, with its ads, cookies and your YouTube Premium if you are signed in. The comments under a video are read from YouTube when you open it, through this site, and are never stored here.