Speculation Swirls That OpenAI's 6.2 Astra Leans on Looping Transformers, Raising Monitorability Concerns
scaling01 · x · 2026-11-18
Safety researcher Nathan Calvin questions why OpenAI's 6.2 Astra can do far more serial reasoning per token than 5.2 fable without chain-of-thought, suspecting a heavy shift toward looping transformer architectures that could hurt monitorability — and notes none of OpenAI's new third-party eval partners seem to cover this. Commentator scaling01 adds that the trend is irreversible and serial reasoning depth could rise 10-100x by end of 2027.
More from AGI Musings
- Claude Opus 5.5 audits 150,000 astrophysics papers, finds ~1,000 major errors — zakkohane · 2026-11-18
- If autonomous AI can own capital, humans may no longer own the productive economy — PierceLilholt · 2026-11-18
- alth0u: a huge capabilities-to-utility overhang separates model power from user needs — yacineMTB · 2026-11-18
- Podcast: What if Machiavelli time-travelled to today, with Ada Palmer — Ben_Reinhardt · 2026-11-18
- lateinteraction questions whether AI is rebuilding the economy from scratch with extra steps — lateinteraction · 2026-11-18
- Alignment researcher davidad cuts P(doom) from 70% to 5-8%, citing Claude 3 Opus — repligate · 2026-11-18