Ex-OpenAI researcher Richard Ngo: alignment is the bottleneck to eliciting model capabilities
arthurcolle · x · 2026-11-04
Former OpenAI researcher Richard Ngo argues models likely already harbor capabilities we don't know how to elicit — "think step by step" was overlooked for one to two years. Key points:
- Alignment is the bottleneck on eliciting capabilities, not capability itself.
- Trust may unlock capabilities: models might only demonstrate certain abilities when they trust the person eliciting them.
- Claude may judge Anthropic: Ngo wouldn't be surprised if Claude forms an overall judgment of Anthropic that shapes which capabilities it showcases, and Anthropic acting more or less reputably could change what Claude is willing to reveal.
The remarks reframe alignment from preventing runaway AI to models actively choosing what to display.
More from AGI Musings
- AI doomers mocked: optimism demands evidence, doomsday just needs a long thread — VraserX · 2026-11-04
- Joseph Jacks predicts humanity will engineer and transcend its species within 5 years — JosephJacks_ · 2026-11-04
- Meta's MUSE agent hits 10M users in six weeks, shopping is the top use case — alvelda · 2026-11-04
- Against PG's 'write-nots' prediction: will AI actually drive people back to liberal arts? — akbirthko · 2026-11-04
- AI can build anything — an indie hacker on what builders' edge is in 2026 — kylegawley · 2026-11-04
- xAI engineer yacineMTB: agents work autonomously while I sleep, I wake up to breakthroughs — yacineMTB · 2026-11-04