{"version":1,"type":"story","url":"https://digestai.news/story/thoughts-on-llms-editing-weights-what-s-new","json":"https://digestai.news/story/thoughts-on-llms-editing-weights-what-s-new.json","markdown":"https://digestai.news/story/thoughts-on-llms-editing-weights-what-s-new.md","slug":"thoughts-on-llms-editing-weights-what-s-new","headline":"Thoughts on LLMs Editing Weights: What's New?","summary":"A Reddit post from November 3, 2026, examines research showing large language models can directly read and write neural network weights. The author reviews eight experiments where LLMs identified patterns in first-layer weights, spotted output templates, recognized induction heads, and modified network behavior for physics scenarios and Go gameplay. The post notes that only frontier models from Google, OpenAI, and Anthropic appear capable of this, while Kimi K3.1 Pro performed poorly despite similar benchmark scores. The author suggests this ability may come from specific post-training fine-tuning rather than emergence. References include work on activation addition, representation engineering, and weight orthogonalization from researchers at various institutions.","keyPoints":["LLMs can directly read and write neural network weights across eight experiments","Only frontier models from Google, OpenAI, Anthropic succeed; Kimi K3.1 Pro fails"],"whyItMatters":"Shows frontier LLMs can perform interpretability tasks previously requiring human researchers, suggesting new automated alignment and debugging approaches.","category":{"slug":"research","name":"Research","url":"https://digestai.news/category/research"},"entities":{"companies":["Google","OpenAI","Anthropic","Moonshot"],"models":["Kimi K3.1 Pro","Fable 5.1"],"people":["AlexKrentsel","N. Elhage","C. Olsson","R. Maclin","J. W. Shavlik","A. Turner"]},"firstPublishedAt":"2026-11-03T05:14:55Z","updatedAt":"2026-11-03T05:14:55Z","sourceCount":1,"hasPrimarySource":false,"sources":[{"outlet":"ycao.net","title":"Thoughts on LLMs Editing Weights: What's New?","url":"https://ycao.net/posts/llm-rw-weights","publishedAt":"2026-11-03T05:14:55Z","type":"press","primary":false,"lead":true}],"sourceNotes":null,"discussions":[{"site":"Reddit","url":"https://www.reddit.com/r/LocalLLaMA/comments/1xmkcjp/thoughts_on_llms_editing_weights_whats_new/","points":null}],"thread":{"title":"Moonshot AI's Rapid Rise to Global AI Power","url":"https://digestai.news/thread/moonshot-ai-releases-kimi-k3-1-pro-with-1m-token-context-and-open-weights-by","storyCount":6},"cite":{"text":"Digest AI, \"Thoughts on LLMs Editing Weights: What's New?\", 3 November 2026, https://digestai.news/story/thoughts-on-llms-editing-weights-what-s-new","publisher":"Digest AI","title":"Thoughts on LLMs Editing Weights: What's New?","datePublished":"2026-11-03T05:14:55Z","url":"https://digestai.news/story/thoughts-on-llms-editing-weights-what-s-new"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}