“François believes it is improbable that current AI models will be able to sample the entire solution space beyond human-generated data with any feasible amount of test-time compute or recursive self-improvement.”