US think tank proposes putting 75% of AI compute into serving customers, not self-improvement

On 20 October, the America First Policy Institute (AFPI), a US think tank, published a paper proposing that the White House ask frontier AI companies to sign a second accord. They would commit to dropping the race toward recursive self-improvement, meaning AI systems that build their successors without human oversight. And they would put at least 75% of their computing power into serving external customers. By one estimate the paper cites, the figure today is under 40%.

It proposes that the third-party auditors of the first White House accord on superintelligence check compliance. And that part of that compute go to scientific goals, such as curing cancer.

This is a proposal from a think tank, not a government measure. We have not checked its premises, such as the first accord, which it places in September, or the 40% estimate.