Anthropic co-founder Jack Clark moves his prediction for self-improving AI forward to 2027

What happened

Anthropic co-founder Jack Clark updated his forecast on the arrival of self-improving AI, moving the date closer to 2027. This change is driven by rapid progress seen in a new benchmark developed by P-Zero Research. The benchmark tests AI models on their ability to design and evaluate scientific experiments autonomously. Results are improving quickly enough to accelerate timelines for AI systems that can enhance their own capabilities without human intervention.

Why it matters

A key challenge in building advanced AI is getting models to independently plan and adjust experiments to improve themselves. The P-Zero benchmark provides an objective, measurable way to track this ability. Faster improvement on this benchmark suggests AI systems could soon learn to self-optimize, raising the stakes for developers, regulators, and businesses. Self-improving AI would shift power toward more autonomous systems and could speed up breakthroughs or risks in unpredictable ways. For operators and investors, this means reassessing timelines, safety protocols, and readiness for more adaptive AI tools appearing in the next few years rather than a distant future.

What to watch next

Watch for further developments from P-Zero Research on how the benchmark evolves and how other labs’ AI models perform. Also, track what Anthropic and competitors do to integrate self-improvement into their product roadmaps or safety frameworks. This progression will pressure AI governance and risk management practices to keep pace. Investors and founders should monitor shifts in funding and talent toward projects targeting self-improving AI capabilities. Finally, regulators may need to start considering policies for AI that can upgrade itself without direct human oversight, which upends current assumptions about control and accountability.

AI Quick Briefs Editorial Desk