Astra appears to solve competition math without verbalized reasoning, monitorability worsens
sjgadler · x · 2026-09-04
Ryan Greenblatt argues GPT-6 Astra shows a massive jump in opaque reasoning: benchmarks suggest it can solve hard competition math entirely in its head, where prior AIs handled only basic word problems — with caveats around possible contamination. UK AISI separately found Astra has much worse monitorability. He suspects architectural changes with increased serial depth, though a normal pretraining scale-up is plausible. Nathan Calvin adds the monitorability situation is worse than expected and harder to coordinate on than a single taboo.
More from Models
- Mirai's uzu engine brings speculative decoding to Apple M5, hitting 105 tok/s on Qwen3.6 27B — TheMoonMidas · 2026-09-04
- Premium-tier early access for $300 users sparks debate over staggered model rollouts — GlenBradley · 2026-09-04
- Best local models for 12GB of VRAM: Gemma-4-12B remains the pick — GlennCameronjr · 2026-09-04
- Claude Fable 5.1 Launches, Early Users Say It One-Shots the Best Websites of Any Model — repligate · 2026-09-04
- OUI-1: a fine-tuned Diffusion Gemma for Generative UI, 8x fewer params, open weights — GlennCameronjr · 2026-09-04
- GPT-6 Astra debuts at No.1 on Terminal-Bench, 1.9% ahead of Claude Fable 5.1 — sandersted · 2026-09-04