LessWrong (30+ Karma) “There should be $100M grants to automate AI safety” by Marius Hobbhahn
Apr 3, 2026
Marius Hobbhahn, author and Apollo Research affiliate, proposes massive grants to scale automated AI-safety work. He urges urgent, large-scale funding and a grant model that ramps to $100M+ budgets for automated safety pipelines. He outlines concrete areas like monitoring, automated red-teaming, white-box auditing, propensity evaluations, and automated conceptual alignment research.
AI Snips
Chapters
Transcript
Episode notes
Make Large Grants Conditional On Public Benefit
- Require public benefit and publication when a grantee proves they can spend $100M meaningfully on safety.
- Options include open-sourcing the pipeline, collaborating with labs, joining an AGI lab, or turning into an AGI-safety for-profit under publication conditions.
Explicit Programs Attract Entrepreneurial Talent
- Announce explicit grant programs because conservatism among funders means entrepreneurs won't attempt ambitious scaling without clear commitments.
- Hobbhahn expects entrepreneurial talent needs visible incentives to choose safety-first scaling projects.
Goodharting Can Be A Useful Failure Signal
- Goodharting is a real risk but observing metric failure is informative evidence to stop that program.
- Hobbhahn suggests robust metrics help, and failed metrics signal the need to halt particular scaling attempts.
