Comment by AI Futures Project

To boost safety research while still allowing companies to generate lots of revenue, the government could require AI companies to use some percent of their compute for serving external customers and some percent for (ideally transparent) safety research (and keep the rest for capabilities R&D). [...] Under this policy, frontier AI companies are required to use a minimum fraction (e.g., 25%) of their compute for transparent safety research or to donate it to external AI safety nonprofits.