The name is the button.

An agent writes the code now, and a human thumb hovers over Accept all a hundred times a day. The honest question of this decade is what, exactly, you just accepted. So I read the diff.

Eight years as a backend engineer in Stockholm, back in Tokyo since 2023, and I still ship something small most nights, usually badly. My one rule: the only real review of a model is the thing you built with it. That is why there is no leaderboard on this site.

Around two hundred sources come through — the labs' own posts, the papers, the tools, the runtimes and languages nobody is watching yet, the chip under the demo, and the newsletters people actually forward. What is moving gets counted so the week reads at a glance, and every few days I explain one mechanism in plain words.

A story here is a headline, a short excerpt, one image and the link to whoever wrote it. We never reproduce the post; the images belong to their authors. A post written in Japanese appears in Japanese, exactly as it was published, and the link goes to the original.

— Yuki Halvorsen, Tokyo

← Accept All Archive

Redwood ResearchResearch

Latent reasoning architectures would likely undermine CoT, our strongest oversight tool

09.23

We should have a strong presumption that latent reasoning architectures would make oversight far more difficult.

Research

Read at Redwood Research ↗

ResearchBrief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident Redwood Research

One letter a week, starting soon.

Latest

Latent reasoning architectures would likely undermine CoT, our strongest oversight tool

Related

ResearchAstra is much better at reasoning with filler tokens than previous models Redwood Research

ResearchCoT controllability evals seem very under-elicited Redwood Research

ResearchProposal for tracking the effects of architecture on monitorability Redwood Research

ResearchBrief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident Redwood Research

ResearchSOTA alignment assessments don’t strongly update us against misalignment Redwood Research

ToolsHigh-volume AI scoring with Jev and Pydantic Evals Pydantic