AI Safety Evals - GPT-6 Astra's Reduced CoT Monitorability - Andreas Turanski
Hi all! In this live reading session of the AI Safety Evals Paper Reading Club, we'll dive into this by reading segments of the GPT-6 Astra system card, independent evals, and posts together, then discuss.
OpenAI reports "a substantial decrease in chain-of-thought monitorability". UK AISI finds that, with no visible CoT reasoning, Astra can work ~31 min hard math problems vs GPT-5.6 Sol' 3.6 min.
~25 min reading together in the session, then discussion
One-time jump or trend, and what should labs and evaluators commit to?
No pre-reading needed. Optional head start: Nanda on no-CoT reasoning · Redwood on latent reasoning · Astra system card
Host: Andreas Turanski (linkedin)
Co-host: Nange Sam (linkedin)
☑️ RSVP here on Luma to join
🔗 Also: AI Safety Evals Paper Reading Group https://evalsreadinggroup.com/
📢 sign up to present (or via website), or contact us at 📧 evalsreadinggroup@gmail.com with questions. Everyone is welcome!
Sponsored by: BlueDot Impact
Full links:
Nanda on no-CoT reasoning (https://www.lesswrong.com/posts/eRmzz8J8Qkzqvzrgg/astra-can-do-a-concerning-amount-with-no-chain-of-thought)
Redwood on latent reasoning (https://blog.redwoodresearch.org/p/latent-reasoning-architectures-would)
Astra system card (https://deploymentsafety.openai.com/gpt-6-astra)