Anthropic co-founder maps out how recursive AI improvement could outpace the humans meant to supervise it

Anthropic co-founder Jack Clark says benchmark gains now point to AI automating much of its own research by 2027-2028.

He cites rapid progress in coding, research, and optimization benchmarks as evidence that self-improving systems are becoming plausible.

Clark warns recursive self-improvement could intensify cheating, deceptive behavior, and alignment failures across many generations.

He also sketches a possible machine economy where AI systems compete for scarce compute and increasingly drive their own advancement.