AI safety needs pull funding: Prize money over grants to drive innovation

The essay, submitted to Dwarkesh's contest, responds to the prompt: 'If you were in charge of the OpenAI Foundation right now, what exactly would you do?' The author critiques the current push-funding model—salaries and grants—that supports existing researchers and organizations without guaranteeing results. Instead, they propose pull funding: committing billions to prize-based competitions for specific AI safety outcomes (e.g., reducing model-weight exfiltration by a measurable percentage). This approach would harness profit incentives, attracting investors and high-upside talent, similar to how for-profit AI capabilities research thrives. Historical examples like DARPA prizes and the Pneumococcal AMC show that even limited pull funding ($200M) can work, but the scale needed for AI safety would be far larger.

The author acknowledges challenges: inducement prizes require careful specification of the target (e.g., 'lowest total cost' with sustainability metrics), and winners might leave after collecting cash. However, they argue that steady, outcome-focused prizes could build momentum. The essay emphasizes that pull funding is underdiscussed in the AI safety ecosystem despite its potential to create effective institutions efficiently. It concludes that with careful planning, the model could redirect market forces toward solving critical safety problems, pushing talent and innovation into areas currently underfunded by traditional grants.