Anthropic co-founder Jack Clark says AI is nearing the point where it can automate AI research

According to Jack Clark, co-founder of Anthropic, AI systems are approaching a critical inflection point where they can autonomously conduct AI research. In Import AI 455, he estimates a ~30% probability by end of 2027 and a ~60%+ probability by end of 2028 that AI will be able to train its own next-generation models. Clark argues that genius-level creativity isn't necessary for self-improvement; the path is already visible in AI's rapid progression from simple coding assistance to performing real research tasks. Concrete evidence includes AI reproducing published papers, constructing machine learning systems from scratch, fine-tuning complex models, and optimizing kernel code — one instance saw model training code sped up by 52x. These early signs suggest AI is beginning to push scientific boundaries on its own.

Clark's warning is about the control implications: if AI crosses the threshold into automated R&D, models could accelerate their own development in ways that become much harder to predict or govern. The shift from tool to autonomous researcher would mark a fundamental change in how AI advances, potentially leading to rapid, unforeseen breakthroughs. As Clark notes, the timeline is aggressive but data-driven; the current trajectory points to AI handling increasingly complex research loops. For the tech industry, this means preparing for a world where AI not only helps us build AI — but builds itself. The coming years will test our ability to steer this process safely.