Anthropic Co-founder Jack Clark Projects 60% Chance of AI Self-Improvement by Late 2028
AI may soon build itself—and humanity might not be ready.
Anthropic co-founder Jack Clark has projected a 60% chance that AI systems will reach a stage where they can rebuild and improve themselves by the end of 2028. Clark arrived at this estimate after reviewing hundreds of public datasets on AI development in recent weeks. He shared the assessment on X, warning that AI may soon be able to build itself, with human involvement shrinking in the research and development process. This shift, he said, is 'a very important thing'—not just a performance boost but a potential compression of development cycles.
Anthropic researcher Ajeya Cotra also updated his earlier forecast, now estimating that AI could autonomously handle over 100 hours of software engineering tasks by year-end, up from a January estimate of 24 hours. Clark's remarks align with a growing view inside the industry that AI is moving beyond assisting with code toward automating the entire R&D pipeline. If AI self-improvement becomes reality, the gap between development speed and regulatory oversight could widen significantly, raising urgent questions about safety, verification, and human control.
- Jack Clark (Anthropic) puts 60% odds on AI self-improvement by late 2028, based on public data analysis.
- Anthropic researcher Ajeya Cotra revised autonomous software engineering estimate from 24 to 100+ hours per session.
- Clark warns human oversight is shrinking and AI systems may soon design and improve themselves without humans.
Why It Matters
Self-improving AI could outpace regulation, forcing a rethinking of safety, oversight, and human control.