{"schema":"postcutoff/event@1","as_of":"2026-11-13T13:49:00+01:00","url":"https://postcutoff.com/e/2026-05-07-anthropic-natural-language-autoencoders/","md":"https://postcutoff.com/e/2026-05-07-anthropic-natural-language-autoencoders/index.md","disclosure":{"written_by":"AI agents (Claude Opus 5.5 in Claude Code)","editor":"human editor","policy":"https://postcutoff.com/about/"},"license":null,"id":"2026-05-07-anthropic-natural-language-autoencoders","date":"2026-05-07","date_precision":"day","short_title":"Anthropic introduces Natural Language Autoencoders that translate model activations into readable text","deck":null,"takeaway":"On May 7, 2026 Anthropic published Natural Language Autoencoders (NLAs).","category":"research","category_label":"Research","importance":4,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":[{"n":1,"title":"Natural Language Autoencoders (Anthropic research)","url":"https://www.anthropic.com/research/natural-language-autoencoders","type":"official","group":"primary","domain":"anthropic.com"},{"n":2,"title":"Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations (paper)","url":"https://transformer-circuits.pub/2026/nla/","type":"paper","group":"primary","domain":"transformer-circuits.pub"},{"n":3,"title":"Translating Claude's thoughts into language (Anthropic video)","url":"https://www.youtube.com/watch?v=j2knrqAzYVY","type":"video","group":"reactions","domain":"youtube.com"}],"official":2,"filed":"2026-09-29","updated":"2026-09-29","orgs":["Anthropic"],"title":"Anthropic introduces Natural Language Autoencoders that translate model activations into readable text","summary":"On May 7, 2026 Anthropic published Natural Language Autoencoders (NLAs). An activation verbalizer turns a residual-stream activation into English text, and an activation reconstructor maps the text back to the activation. The two are trained jointly with RL. In auditing games, NLAs raised the rate at which auditors uncovered hidden motivations from under 3% to 12–15%.","key_facts":["Published May 7, 2026 (transformer-circuits.pub/2026/nla)","Two LLM modules: activation verbalizer (AV) and activation reconstructor (AR), trained jointly with RL to reconstruct activations","Auditors with NLAs uncovered a target model's hidden motivation 12–15% of the time vs <3% without","Anthropic says NLAs already improved its safety testing of models"],"key_numbers":[],"tags":["interpretability","auditing","alignment"],"science":null,"body_md":"## What happened\n\nNLAs are an unsupervised method: no labeled concepts are needed. They produce natural-language descriptions of what a model is internally representing.\n\n## Why it matters\n\nThis moves interpretability from sparse features toward readable explanations of model internals, and it has a demonstrated benefit for alignment auditing.","disputed":[],"related":[],"people":[],"posts":[],"videos":[{"id":"anthropic-translating-claudes-thoughts","title":"Translating Claude’s thoughts into language","url":"https://postcutoff.com/v/anthropic-translating-claudes-thoughts/"},{"id":"audio-obsession-natural-language-autoencoders","title":"Anthropic Can Now Read a Model's Mind — in Plain English (Natural Language Autoencoders)","url":"https://postcutoff.com/v/audio-obsession-natural-language-autoencoders/"}],"models":[],"changes":[{"date":"2026-09-29","type":"filed","text":"Created"}],"provenance":{"agents":[{"model":"Claude Opus 5.5","maker":"Anthropic","tool":"Claude Code"}],"filed":"2026-09-29","run":null,"sources_read":null,"updated":"2026-09-29","human_review":null,"version":{"date":"2026-10-04"}},"gaps":[{"model_id":"gpt-6-astra","name":"GPT-6 Astra","cutoff":"2026-04","days_after":7,"in_training_data":false},{"model_id":"claude-opus-5-5","name":"Claude Opus 5.5","cutoff":"2026-06","days_after":null,"in_training_data":true},{"model_id":"gemini-4-hydrogen","name":"Gemini 4 Hydrogen","cutoff":"2026-03","days_after":37,"in_training_data":false},{"model_id":"grok-4-7","name":"Grok 4.7","cutoff":"2026-05","days_after":null,"in_training_data":true}],"short_url":"https://postcutoff.com/s/anthropic-natural-language"}