Import AI 468: 23 RSI ideas; PostTrainBench+; and how trust and transparency interplay with AI racing

Details

Issue 468 of Import AI, the newsletter by Anthropic co-founder Jack Clark, leads with policy recommendations from think tank IFP on managing Recursive Self-Improvement (RSI) and automated AI R&D. The 23 recommendations span seven categories: transparency into automated AI R&D, improving state capacity to understand and respond to it, risk management strategy, investment in AI verification technology, AI resilience, extending the US AI lead to create more management time, and creating option value for international cooperation.\n\nClark frames the current state of AI development as a car with only an accelerator pedal and no brake or telemetry: these policy recommendations are designed to build the instruments that allow course correction. The IFP proposals are described as 'low-regret'—valuable regardless of whether RSI arrives in two years or twenty, because they build institutional capacity that is needed at any pace.\n\nThe issue also covers PostTrainBench+, a new or updated benchmark evaluating post-training techniques, and analysis of how trust and transparency between competing labs affects racing incentives. Clark argues that labs with greater mutual transparency have reduced incentive to race recklessly, with implications for proposed transparency regulations in the US and EU.\n\nImport AI is a high-signal newsletter in AI policy circles. Clark's framing of governance issues consistently influences both industry discourse and policymaker thinking.