Jack Clark: 60% chance frontier model trains its successor by 2028; Azhar analyzes what labs' hiring and spending patterns reveal
Anthropic's Jack Clark published an essay arguing there is a 60% probability that a frontier model will train its successor by 2028, representing potential recursive self-improvement and a fundamental shift in AI development. Azeem Azhar's Exponential View analysis challenges this timeline by noting that frontier training has become less a pure research problem and more an industrial scaling challenge—bottlenecks now center on land leases, power infrastructure, chip procurement, and electrician hiring rather than algorithmic optimization alone. Azhar proposes examining frontier labs' revealed preferences to test Clark's thesis. If automated R&D were truly likely by 2028, labs should be: (1) shifting hiring away from pure researchers toward research multipliers and people who can build automated research factories, and (2) overinvesting in compute now to secure pre-commitments before physical constraints become binding—more GPUs, memory, power, data centers, and inference capacity. Over a three-year horizon, those physical constraints may matter more than algorithmic advances. Exponential View members also gain access to an interactive model exploring how more capable AI agents will affect token consumption and infrastructure demands, offering quantitative tools for understanding scaling economics and competitive dynamics.