Lermen argues Plan C unlikely to solve ASI alignment

The author argues that Plan C, which allocates only a few months to solving artificial superintelligence alignment, is unrealistic and unlikely to succeed. The post contends that alignment faces insurmountable hurdles including the high-stakes 'one-shot' nature of ASI, the risks of learned optimization where models develop unintended inner objectives, and deep conflicts between corrigibility and coherent agency. Open article →