Ryan Greenblatt questions whether plain-text CoTs will really become an alchemical era of observability
RyanGreenblatt · x · 2026-09-11
In a debate on chain-of-thought observability, @tszzl argued that the era of trying hard to preserve plain-text CoTs will look like an "alchemical era of observability" — we could instead understand models' alien ontology from the ground up. Redwood Research's Ryan Greenblatt replied skeptically: "But will we?" The exchange touches the core tension between CoT monitoring and deeper interpretability of model internals.
Related event: Researchers call text-based CoT observability a "alchemy era"(5 posts)→
More from AGI Musings
- DeepMind won CASP14 and a Nobel; OpenAI's math AI bid gets pushback — dhruv2038 · 2026-09-11
- Andy Masley: he never called the data center backlash a psyop because such ops don't work at that scale — AndyMasley · 2026-09-11
- OpenAI's Pachocki: better math models shelved because compute goes to RSI — basedjensen · 2026-09-11
- OpenAI and Anthropic reportedly attacking P vs NP — what would a constructive P=NP proof break? — MohMayaTyagi · 2026-09-11
- Viral Quip Predicts AI Doom Discourse Will Just Keep Moving the Goalposts — nabla_theta · 2026-09-11
- Experts gave AI just 10% odds of solving a Millennium Problem by 2027 — reality disagreed — Tolopono · 2026-09-11