AI Alignment Forum: latest AI & code stories | Accept All

Skip to content

Accept All すべて承認

AI Alignment Forum

10 stories

Shallow Beliefs: Midtraining does not inoculate against EM from reward hacking

AI Alignment ForumResearchShallow Beliefs: Midtraining does not inoculate against EM from reward hacking It would be useful if we had the ability to modify a model’s beliefs. For example, this could facilitate honeypots and better monitoring [1] , help us do better science on current models [2] , and augment certain forms o

AI Alignment Forum

AI Alignment ForumResearchOp-Ed: I Worked at Google DeepMind. You Should Listen to the Warnings About AI Published in The Guardian . Major AI lab CEOs recently advocated for pacing AI development. They are right to be concerned: the field runs an extremely dangerous race towards superintelligent AI. We can and should demand

CoT controllability evals seem very under-elicited

AI Alignment ForumResearchCoT controllability evals seem very under-elicited The CoTControl eval asks reasoning models to follow formatting constraints in their chain-of-thought (e.g. write in all lowercase, avoid a specific word) while solving questions. Models seem to mostly be pretty bad at th

An operationalization of opaque serial depth

AI Alignment ForumResearchAn operationalization of opaque serial depth Currently, chain-of-thought (CoT) is a valuable tool for overseeing AI models. However, some architectural shifts could significantly reduce CoT monitorability . We have recently proposed that AI companies should transpa

AI Alignment Forum

AI Alignment ForumResearchProposal for tracking the effects of architecture on monitorability Architectures that incorporate opaque recurrence or allow for agents to communicate with each other using latents could rapidly make it much harder to monitor chains of thought or communication (we’ll refer to this prope

Astra can do a concerning amount with no chain of thought

AI Alignment ForumResearchAstra can do a concerning amount with no chain of thought Work done in a personal capacity TLDR : Astra has 8.6x better odds of doing a reasoning task without CoT than the next best model (Fable 5.1), and can do 7.2 serial arithmetic steps in a forward pass vs 4.1 for the next

How good are slop-vestigators?

AI Alignment ForumResearchHow good are slop-vestigators? TLDR: We release MessageBoardAuditBench : a benchmark to measure how well agents can replicate the recent investigation into a swarm of OpenAI agents colluding via a message board on an online wiki. We open-source the be

Training on probes: Research ideas

AI Alignment ForumResearchTraining on probes: Research ideas Recap Sequel to Previous Post . This post might not make sense without it. Last post, I told some stories about how training on probes might let our judgments on easy domains generalize to harder domains by leveraging an

Training on probes: What's going on

AI Alignment ForumResearchTraining on probes: What's going on TL;DR If you train a probe for some property (like "honesty") and do gradient descent against this probe while continuing training that incentivizes dishonesty, the model will change its internal representation to evade

AI Alignment Forum

AI Alignment ForumResearchThe Alignment Journal: Organization, Personnel, and Scope [Cross-posted from the Alignment Journal's blog ] The Alignment Journal is beginning to invite the authors of select papers to submit their work for review. If you are interested in participating as an action editor or a

On the calendar

予定 · Conferences and launch events, this week's stories from each corner of the industry, and what its press is covering — in English

On view

This week, from that corner the newest stories from the corner of the industry you picked

The wire their own posts, translated

How code moves in 2026 from the industry's own reports

Sources

Fetched every three hours, linked to the original

About

Accept All is a magazine, a radar and a diary about artificial intelligence and the craft of code — read from Tokyo, for people who build.

The name is the button. Code is written by agents now and reviewed by a human thumb that hovers over Accept all a hundred times a day; the whole question of this decade is what, exactly, you just accepted. So we read the diff. A magazine, because every week one model, one tool or one idea is worth reading about slowly. A radar, because we count what the labs, the Valley, GitHub and the chip makers are talking about — releases, agents, open weights, compute, funding, rules — and show it as it moves. A diary, because a seal is set each day in that day's colour, and the days add up.

We read the research labs' own posts, the American and European tech press, the Japanese and Chinese developer sites and the newsletters people actually forward, and bring it together in English and Japanese. Everything here links back to the people who wrote it; the images belong to their authors.

Accept All will keep changing. New sources, new signals, new ways of looking — a little better each week, never finished. If something here made you open a terminal, that is enough.

Thank you for reading. ありがとうございます。