Jack Clark tells IMF the singularity has begun

Clark's argument rests on a shift he says became visible this year: single frontier models such as Claude, ChatGPT and Gemini are now doing work at the frontier of several scientific disciplines at once. He traced the trend through benchmark scores, noting Humanity's Last Exam rose from 8% at launch in January 2025 to 67.7% for Claude Opus 5.5 as of this week, while FrontierMath went from 2% to 94%.

The TheoPhysEval episode is his sharpest illustration. When models disputed expert-written answers on unpublished physics problems, the experts revisited their work and found subtle errors buried in their calculations, and GPT-6-Astra's score climbed to 98% (49/50). Clark frames this as the teacher becoming the student, and argues that creative problem-solving in science implies AI systems will generalize into many domains, possibly including AI research itself.

He also described a buildout gap: data center capacity is concentrated in the United States, China and increasingly the Middle East, and supply chains are under contention with prices rising. With UNESCO estimating roughly 10 million scientists worldwide, he said there are already too few labs and instruments, and urged policymakers to prepare facilities and rules so that AI-driven discovery is not bottlenecked downstream.