At the IMF and World Bank Group annual meetings in Bangkok, Anthropic co-founder Jack Clark urged policymakers to prepare for AI-driven science, whose wider economic payoff remains uncertain.
Clark delivered the Per Jacobsson Foundation lecture on Oct. 17. In his published speech, he wrote, “The singularity is here.” His use of the term draws on a 1993 essay by Vernor Vinge, who argued that intelligence beyond human capacity could accelerate technological change and make forecasts less reliable.
A separate measure of AI’s research abilities comes from TheoPhysEval, a preprint built around 50 previously unpublished theoretical physics problems. Its authors reported that GPT-6-Astra, working through Codex with research and computing tools, produced 49 correct final answers, or 98%. They put the average cost below $3 per problem. TheoPhysEval co-author Suvrat Raju said experts changed several of their original answers after examining cases where the models disagreed with them.
The preprint cautions that success on assigned problems has yet to produce new theoretical physics advances led by AI agents. A benchmark Clark did cite was Humanity’s Last Exam. Anthropic said Claude Opus 5.5 scored 67.7% on that test using tools, a condition that matters when comparing results across testing setups.
Clark also cited a roughly 94% result on FrontierMath. Epoch AI identifies that figure with a revised test covering 295 problems in Tiers 1–3; a separate Tier 4 has 43 problems. The figures therefore describe particular tests, not scientific discovery or economy-wide gains.
An April 2026 IMF staff note ties the economic effects of AI to how widely it is adopted. Energy supply, data centers, organizational change and institutions can affect that process. The note says firms and countries equipped to use AI may benefit more. It separately flags the possibility of concentrated gains and uneven effects on workers.
The IMF’s October 2026 Global Financial Stability Report highlighted the risk to assets tied to AI. Concentrated holdings and borrowed financing could magnify losses if AI investment or profitability falls short. The staff note calls for better tracking of adoption and labor market effects as policymakers assess whether technical progress is reaching the wider economy.
