How AI safety grantmaking should work: a debate over vouching
NathanpmYoung · x · 2026-08-25
Reviewer Nathan Young responds to criticism that his approach fundamentally doesn't scale—suggested as a reason AI safety grantmaking has scaled poorly.
His points: he is mainly shown epistemics projects and is one person, so his experience may not generalize; vouching is only one way to gain confidence—clear impact paths and track records work too; the more information a project provides, the better it can pull away from a base rate worse than GiveDirectly; he doesn't expect to rate projects mainly on vouches, but will email vouches and ask clarifying questions, and being allowed to do so makes a high rating more likely.
More from Safety
- Court expert used ChatGPT to write report concluding 3M had 0% fault in explosion — ZeroStateReflex · 2026-08-25
- Viewpoint: More OpenAI Researchers Shift Stance on Alignment Risks — davidmanheim · 2026-08-25
- Richard Ngo previews part 2 of his alignment retrospective — RichardMCNgo · 2026-08-25
- UK, Ukraine sign AI defense partnership sharing 5M battlefield images — Polymarket · 2026-08-25
- CISA warns of AI-generated threats to Siemens PLCs — dhadfieldmenell · 2026-08-25
- Google Spam Update severely impacts AI-generated sites — gaganghotra_ · 2026-08-25