Researchers warn against "neuralese" architectures that trade AI monitorability for performance
RyanGreenblatt · x · 2026-09-11
Redwood Research's Ryan Greenblatt voiced concern about AI architectures shifting from readable chain-of-thought to opaque "neuralese" activations. Based on limited public evidence, he sees Google's Astra as a concerning step in this direction: the monitorability-vs-performance trade-offs in its architecture and training lack sufficient public evidence for informed scientific discussion, and it's unclear how AI companies will make such trade-offs going forward. He calls on companies to release evidence enabling a reasonably informed public conversation. OpenAI researcher balesni amplified the thread, saying he thinks AI is >10% likely to kill all humans and this proposal is among the industry's top risk-reduction steps.
Related event: Redwood Research Proposes Disclosure Rules to Protect CoT Monitorability(5 posts)→
More from AGI Musings
- LukeW: Technology Limits Are Temporary — and AI Skeptics Are Already Feeling It — LukeW · 2026-10-27
- GiveDirectly pilots an optional AI chatbot alongside cash aid in Rwanda — NathanpmYoung · 2026-10-27
- Ex-Anthropic employee essay: the org optimizes for power 'constrained by being seen as the good guys' — RichardMCNgo · 2026-10-27
- Investor: model version numbers are a mirage, only datacenter margins are real — StewartalsopIII · 2026-10-27
- WSJ Op-Ed: For Anthropic Share Offering Investors, a 10% Extinction Chance May Be a 90% Portfolio Win — davidmanheim · 2026-10-27
- WSJ: for Anthropic investors, a 10% extinction chance may still be a 90% portfolio win — peterwildeford · 2026-10-27