Ryan Greenblatt warns 'neuralese' architectures let AIs think in opaque activations, citing Astra
RyanGreenblatt · x · 2026-09-11
AI safety researcher Ryan Greenblatt says he's deeply worried about architecture changes that push AIs to reason in opaque activations instead of readable chains of thought — so-called "neuralese". Based on limited public evidence, he argues Astra appears to be a concerning step in this direction, and that insufficient public information exists to enable a well-informed scientific discussion about what these changes mean. He links a proposal addressing the problem.
Related event: Redwood Research Proposes Disclosure Rules to Protect CoT Monitorability(5 posts)→
More from AGI Musings
- Economist Bharat Chandar named Research Director for AI Economics at Stanford Digital Economy Lab — RishiBommasani · 2026-10-27
- Aeon's former EVP: AI agents are the most impactful innovation of my lifetime — akumar · 2026-10-27
- Physicists Let AI Pick Its Own Problem—It Wrote a Publishable Paper in 2 Hours — burny_tech · 2026-10-27
- Deep learning veteran warns: don't let entrenched disbelief derail prep for the next big tech — jachiam0 · 2026-10-27
- Should platforms serve third-party AI agents? The antitrust question — sjgadler · 2026-10-27
- Musk: robot RAM demand will dwarf everything else, and 100% of superintelligence compute goes to deep space — elonmusk · 2026-10-27