Neuralese Recurrence Explained: Why OpenAI's Astra Loop Is Not Yet Unreadable AI Thought
Astral Codex Ten · rss · 2026-09-25
Scott Alexander's long essay examines whether "neuralese recurrence" — looping internal vector states instead of writing readable chain-of-thought — has arrived. OpenAI chief scientist Jakub Pachocki says Astra's computation depth is "within a factor of two of GPT-4," arguing recurrence is just a way to add layers, no worse than a physically deeper transformer. Alexander explains why looped layers reuse the same knowledge rather than adding capacity: his estimate gives +0.03 "generations" of gains from doubling depth via loops vs +0.25 from real layers. Verdict: not yet a safety nightmare, but growing depth warrants attention.
More from AGI Musings
- Historian uses GPT-6 and Opus 5.5 to crack John Dee's ciphers, urges lab funding — emollick · 2026-09-25
- Brundage follows up: Anthropic can contrast present vs future risks without claiming we're on top of them — Miles_Brundage · 2026-09-25
- Ex-OpenAI safety lead Miles Brundage calls Anthropic's 'we largely understand model risks' claim obviously false — Miles_Brundage · 2026-09-25
- AI Explained digs into Claude Opus 5.5 and how close labs are to automated AI research — AI Explained · 2026-09-25
- Pedro Domingos: ML moves a million times faster than evolution, so AGI is millennia away — pmddomingos · 2026-09-25
- Designer argues agent-native pro tools, not prompt boxes, are the future of AI work — bartonsmith · 2026-09-25