Anthropic’s Jack Clark tells policymakers the singularity has begun

Anthropic co-founder Jack Clark told economists and policymakers at the IMF and World Bank’s annual meetings in Bangkok on Oct. 17 that the singularity has “now begun,” as The Rundown reported. He argued that AI is starting to accelerate discovery faster than humans can follow.

In prepared remarks published Oct. 19, Clark pointed to gains on difficult tests and envisioned “a billion virtual scientists.” He urged countries to build laboratories, robots and scientific equipment to prepare for that possible surge in research.

The numbers behind the claim

Clark cited AI scores on Humanity’s Last Exam, which tests expert knowledge, rising from 8% in 2025 to 67.7%. He also cited FrontierMath’s rise from 2% in November 2024 to 94%.

Anthropic reports the 67.7% HLE score for Claude Opus 5.5 with tools. The comparison leaves open whether both scores reflect identical testing conditions. Anthropic also cautions that benchmark margins have become less reliable indicators of differences in practical performance.

Clark’s economic example carries its own conditions. Anthropic’s most extreme scenario puts U.S. GDP at $44.4 trillion in 2030, measured at 2025 prices. That is 32.4% above its modeled economy without AI. The scenario assumes exceptionally capable, autonomous systems and rapid adoption.

Why it matters

Recent math and science results give Clark substantive evidence for his argument. One of the people building the technology is bringing a claim long associated with science fiction directly to the institutions that help shape the global economy.

In August, Anthropic reported an advance concerning the Riemann zeta function. Its unreleased Claude research model raised a lower bound on the share of relevant zeros lying on the critical line from 41.6% to 67.2%. The company describes review by its mathematicians and a formal check in the Lean proof system. Anthropic says the advance does not solve the Riemann hypothesis and that it does not expect these techniques to prove it.

The physics results are also striking. In an Oct. 9 paper, the creators of TheoPhysEval report that GPT-6-Astra in Codex answered 49 of 50 unpublished theoretical physics problems correctly, with tools, at an average cost below $3 per problem. The collection tests new calculations using known techniques and excludes major open research problems.

Research teams could delegate more specialized calculations and parts of mathematical work, then focus on choosing questions, checking results and testing ideas. These examples leave open how much of the full research process AI can manage on its own.

Stronger research systems also have consequences for security. Anthropic reports that Claude Mythos Preview helped develop improved attacks against HAWK, a proposed digital signature scheme, and a weakened version of the AES encryption algorithm. The company says neither finding required changes to production systems. The work shows how AI could help defenders find weaknesses while also uncovering new ways to attack systems, making expert checks and responsible disclosure important.

Clark’s call for laboratories, robots and equipment connects that research to decisions governments can make now. A surge in ideas for experiments could put pressure on the capacity to test them. Investment in physical research would help determine how quickly AI’s scientific work produces practical gains.

Workers could face disruption as output rises. In Anthropic’s extreme growth scenario, wages for knowledge workers fall by more than 10% and unemployment exceeds typical recessionary levels. These are modeled outcomes, and the model leaves out policy responses, business cycles and other disruptions. Policymakers would have to weigh investment in discovery alongside support for workers moving between jobs and decisions about who shares in the gains.