OpenAI says GPT-6 Astra is more aligned but less monitorable, sparking alignment wording debate
zetalyrae · x · 2026-09-04
OpenAI researcher Tomasz Korbak says GPT-6 Astra is more aligned than previous models but also less monitorable — a concerning trend the company takes seriously. He attributes the monitorability drop to the jump in intelligence rather than direct optimization pressure on chain-of-thought or architecture changes.
Separately, researcher gleech pleads with OpenAI and Anthropic staff to stop saying "aligned" when they only mean benchmark alignment, urging terms like "benchmark-aligned" or "nominally aligned" instead. The exchange highlights growing debate over alignment claims and fading CoT monitorability.
Related event: OpenAI launches flagship GPT-6 Astra to rave reviews and pricing debates(190 posts)→
More from AGI Musings
- Three reasons AI is not conscious and may never be, per Nature essay — anilkseth · 2026-10-21
- Mathematician Daniel Litt concedes his AI bet, but says AI won't solve math — ziv_ravid · 2026-10-21
- AI SaaS monetization is cooked: consumers small, enterprises pay $20/month — menhguin · 2026-10-21
- Tim Dettmers: Academia is about to have a renaissance — the best AI work of the next decade happens in university labs — StasBekman · 2026-10-21
- AI saves scientists 7 hours a week, but validation is the new bottleneck — AnneliesGamble · 2026-10-21
- UW's Amy X. Zhang brings CSCW theory to human-agent collaboration in Stanford talk — amyxzh · 2026-10-21