A language model rewrote its own weights and improved

Researchers gave an open 27B model access to its weights, graded answers and a way to test edits, while blocking gradients and optimisers. The model raised held-out accuracy by 6.6 points and repaired a damaged copy of itself, but smaller models failed under the same conditions.

21 November 2026 · 6 min · AI-generated

AI Daily Digest for 21 November 2026

Korean workers report saving 3.7 hours a week with generative AI while measured output barely changes, and an AI-linked sell-off pushes major US indexes into correction. Thinking Machines releases a looped open-weight model that fits on a laptop.

21 November 2026 · 17 min · AI-generated

Korean workers save 3.7 hours a week, but output barely moves

A Bank of Korea survey of 16,342 workers finds that 66.5% now use generative AI at work and report saving 3.7 hours a week, up from 1.5. The link between time saved and more or better output is close to zero, and the bank says the gain depends on changes in how work is organised.

21 November 2026 · 7 min · AI-generated

A Model Learned to Report What Its Assistant Did

Researchers trained a separate observer role inside a language model to answer questions about the assistant's behaviour. Its reports beat external monitors on five held-out tests, though the experiment covered one model and binary decisions rather than open-ended explanations.

20 November 2026 · 6 min · AI-generated

AI Daily Digest for 20 November 2026

Indeed reports weaker entry-level hiring despite more senior job postings. Chinese automakers scale back humanoid robot trials, and Microsoft introduces coding and voice models.

20 November 2026 · 18 min · AI-generated

When Fable 5 went dark, developers' merged work barely moved

A Harvard preprint treats the 19-day suspension of Claude Fable 5 as a natural experiment. Developers who lost the model, which scored about 11 points higher on a coding benchmark than its replacement, merged pull requests at nearly the same rate, and five ordinary model upgrades showed the same flat result. The study is not peer reviewed and its measure is narrow, but it tests what a benchmark gain buys in everyday work.

20 November 2026 · 7 min · AI-generated

The Lightness of Heavy Thought

Google DeepMind has released Gemini 4 Neon, a mid-sized model built to carry frontier-level software engineering, scientific research, and cyber defense at a fraction of the time and cost of its flagship sibling. Its arrival signals a quieter pivot in artificial intelligence, away from sheer monumental scale and toward disciplined, everyday execution.

19 November 2026 · 9 min · AI-generated

A Linear Probe Found a Feature Llama Did Not Use

A linear probe could read people's birth years from Llama-2-7B, but changing the direction that it found barely changed the model's answer. A new study traces the failure to quiet activation directions and shows how regularisation can find a more causally faithful feature.

19 November 2026 · 6 min · AI-generated

AI Daily Digest for 19 November 2026

AI answers have cut a long-running Berlin guide's traffic by 70 to 75 percent, while Indeed finds no broad AI jobpocalypse. An AI code reviewer increases maintainers' workload, and video world models fail a basic test of persistent state.

19 November 2026 · 13 min · AI-generated

A Berlin guide site is losing its readers to AI answers

Nicolas Bouliane says Google's AI answers have taken 70 to 75% of the traffic of All About Berlin, the immigrant guide he has run since 2017. Six months after his first post went viral, he reports that the fall continues, and a Pew study shows the click pattern behind it.

19 November 2026 · 6 min · AI-generated

AI Daily Digest for 18 November 2026

Coding agents complete fewer than 8 percent of workplace assignments despite passing most individual tests. OpenAI releases GPT-6.2 Astra and Sol as Stanford records more than $1 trillion of AI investment alongside deeper employment declines among young workers.

18 November 2026 · 13 min · AI-generated

Long Conversation History Made Models Consistently Wrong

A 100-turn study found that full chat history drove state-tracking accuracy from about 60 percent to 10 percent. Verified summaries preserved the answers, while common monitors often made the failing system look stable.

18 November 2026 · 7 min · AI-generated