ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
News4 min read·2026-10-19DisconfirmAI Editorial

Anthropic Co-Founder Jack Clark: AI Can Now Do Frontier Science Across Disciplines, and the Researcher Bottleneck May Be Breaking

In a speech at the IMF annual meetings, Jack Clark argued that AI systems have demonstrated breakthrough capability across mathematics, biology, and robotics, and that the training cycle for human researchers—roughly 40 years in his own case—is the fundamental bottleneck to innovation speed. He proposed that AI can be replicated instantly and in arbitrary numbers, enabling unprecedented 'force multiplication' of human scientists, potentially reversing the half-century trend of declining research productivity. He called on countries to build laboratory and robotic infrastructure now while emphasizing the need for careful governance of AI-driven scientific research.

ImportanceStructuralEvidenceE3 inspectableWrite-upDeep

Original publication time 2026-10-19 20:31 GMT+8

On October 17, 2026, Anthropic co-founder Jack Clark delivered the Per Jacobsson lecture at the International Monetary Fund annual meetings in Bangkok, Thailand. His core argument: AI has already made substantive contributions at multiple scientific frontiers, and the training cycle for human researchers—he cited his own nearly 40 years of experience as an illustration—is the fundamental bottleneck to innovation speed. If AI can be replicated instantly and in arbitrary numbers, this bottleneck may be broken.

Clark illustrated this with a story from TheoPhysEval, a paper published just the previous week. The paper tests how well AI systems can solve unpublished problems in theoretical physics; its 50 questions were compiled by physicists around the world. Indian physicist Suvrat Raju, commenting on the results, wrote: 'I don't know of any individual who can solve any significant fraction of our final collection of problems without AI assistance.' More surprisingly, when the research team began investigating mistakes the models might have made, the models—especially GPT-6-Astra—insisted their answers were correct and that the expert-supplied answers were erroneous. When they went back to the experts, it turned out that in multiple instances the experts had indeed found subtle errors buried deep inside their own calculations. After revising their answers, Astra's score climbed to 98% (49/50). Clark characterized this not merely as a high score but as AI correcting human experts' errors.

Sign in / Sign up
ReadingResearchRadarInvestment framework

At the broader benchmark level, Clark cited similar trends: Humanity's Last Exam, launched in January 2025, saw top-scoring AI systems at 8%; as of the week of his speech, Claude Opus 5.5 scored 67.7%. FrontierMath, released in November 2024, started at 2% and now reaches 94%. He claimed: 'Pick your favorite part of science and go looking for a test and you'll find this pattern—sudden and steep increases in capability, everywhere you look.'

But the core mechanism of Clark's argument is not the scores themselves—it is replication. He proposed a thought experiment: what would it mean to run hundreds or thousands of copies of himself simultaneously, working under his supervision? Training a new generation of human researchers in a given area might take 10 to 15 years of targeted education and millions of dollars to produce even a small cohort, whereas AI systems can be created immediately and in arbitrary numbers, as long as there is sufficient computing capacity. He argued this means every human scientist on the planet will become more important and more powerful, gaining the ability to supervise hundreds or thousands of like-minded experts on any problem they find interesting.

Clark also cited Anthropic's internal modeling: if one assumes that models will be able to autonomously complete a broad range of tasks and will diffuse quickly in the economy, then under these singularity-esque conditions, US GDP could reach $44 trillion by 2030. He acknowledged 'this sounds outlandish,' but argued it is implicated by the technology continuing to get better. He called on countries to build scientific equipment and robotic-system infrastructure now, just as he had urged governments a decade earlier to build electricity capacity and data centers—'not many policymakers listened.'

It should be noted that Clark is Anthropic's co-founder, and the model-performance data he cites (Claude Opus 5.5's benchmark scores) comes from his own company's product; the internal researcher code-output multiplication also comes from the same firm. His GDP figure is explicitly framed as a conditional modeling result, not an independently verified forecast. The entire 'the singularity has begun' framework is his personal judgment, not established fact. He himself acknowledged that 'some of which was ad libbed and not captured here,' positioned his own role as 'a narrator of what is happening in AI, a principal agent behind what is happening in AI, and a translator that goes between many communities,' and stated 'I take my own responsibility here extremely seriously.'

Clark's argument is not simply 'AI is getting stronger.' It proposes a specific mechanism: the non-replicability of human researchers is the innovation bottleneck, and AI's capacity for instant replication may change that constraint. Whether this mechanism holds depends on whether current performance on curated benchmarks can translate into autonomous, generalizable scientific contributions—something Clark himself frames with a conditional 'if.'

Sources

  • 原始出处importai.substack.com↗
Continue researching →