AI Safety Prizes Could Beat Grants With 2x Efficiency, Says LessWrong Proposal
Pull funding rewards results, not promises—could reshape AI safety philanthropy.
On LessWrong, Oscar Delaney proposes a shift from upfront grants to “pull funding” for AI safety: pay after the fact for the work that made the most progress. He argues prizes work well when the winner is unpredictable and the solution is easily verifiable, citing DARPA’s autonomous vehicle challenges as a historical example. For AI-specific funding, he says prizes may fit the new wave of AI philanthropists well, because paying out later could be a >2x multiplier and, if advanced AI is available later, could reduce the need for human labor evaluating grants now. Among concrete prize ideas, he highlights compute verification advances—most prominently, retrofitting data centers to be inference-only, which could take the form of an advance market commitment or a working prototype meeting cost and quality criteria.
- Delaney advocates pull funding (prizes) over push funding (grants) for AI safety research
- Prizes can deliver a >2x return on investment by paying only after verifiable results
- Proposes advance market commitments for inference-only data center retrofits as a concrete prize
Why It Matters
If pull funding works, AI safety research gets cheaper and more efficient, attracting talent without wasting philanthropist capital on dead-end grants.