Anthropic's Jack Clark tells IMF the singularity has arrived, citing 67.7% on Humanity's Last Exam

Anthropic co-founder Jack Clark told the IMF 2026 annual meeting in Bangkok that "the singularity has already arrived, and it is this year," according to a report by Chinese outlet QbitAI (量子位), published in Chinese on 2026-10-26. He delivered the Per Jacobsson Lecture, a 62-year-old series that QbitAI says has historically gone to finance ministers, central bank governors and leading economists.

The evidence Clark cited

According to QbitAI's account of the speech:

• Coding: Researchers inside Anthropic now produce 10 times the code output of the company's average over its first four years.
• Humanity's Last Exam: When the benchmark launched in January 2025, the strongest model scored 8%. As of this week, Claude Opus 5.5 scores 67.7%.
• FrontierMath: Top AI systems scored 2 at the end of 2024. They now score 94.
• Unpublished physics problems: A paper published last week collected 50 problems written by leading theoretical physicists from their own unreleased research. GPT-6 Astra and Claude Opus 5.5 both answered them. Preliminary scoring put GPT-6 Astra at roughly 80%.

When researchers told the models where they had supposedly erred, GPT-6 Astra in particular insisted its derivations were correct. According to Clark, the physicists rechecked and found subtle calculation errors in their own reference answers. After the human answers were corrected, GPT-6 Astra's final score was 98%, or 49 of 50 problems. Physicist Suvrat Raju is quoted as saying that, to his knowledge, no one could solve a considerable share of the final problem set without AI, and that he did not know what most of the problems were about.

Clark also said AI is producing key ideas on the Navier-Stokes Millennium Problem, the Jacobian conjecture and several open Erdős problems. He said AI designed novel viral vectors and drug candidate molecules now in clinical trials, and operates physical lab equipment to speed up battery and fuel cell iteration.

Economic claims

Clark cited UNESCO's estimate of about 10 million human scientists worldwide. He described a future in which each can direct hundreds or thousands of AI scientists, scaling the research workforce to a billion or more virtual scientists.

Anthropic's economists ran a simulation under the assumption that autonomous AI scientific discovery takes over economic activity. It projected annual US GDP of $44 trillion by 2030, which QbitAI says equals 37% of today's global economy. This is a modeled scenario, not a forecast with a stated probability.

Clark said the first phase of the AI revolution, data centers and compute, is nearing its end, and the second phase moves into the physical world. He argued that automated robotic labs and fast supply chains from digital discovery to physical manufacturing will determine who benefits, and that developing countries could gain by building standardized physical testing infrastructure.

Risks and the two questions

Clark warned that the same capabilities could design biological weapons or automated cyberattack tools. He also raised the risk of labs such as Anthropic and OpenAI letting AI write and train AI, which he said could push development beyond human comprehension.

He posed two questions: which parts of scientific research humans should keep for themselves, and how far humans need to "understand" AI-generated results. He said some AI-produced proofs this year were verified as correct but follow reasoning so unlike human habits that top mathematicians need months or years to interpret them. His proposed answer was serious rules for the technology and its developers, and "restraint." QbitAI's closing line says the window for this adjustment may be under 10 years.

What this means

The benchmark figures come from Clark's speech as reported by QbitAI. The report names no independent source for them, and the 98% physics result depends on the human reference answers having been revised after the models were scored. The $44 trillion figure is an Anthropic internal simulation under an explicit singularity assumption.

Still, the argument rests on a shift in evidence. Clark pointed to problems with no internet trace, written by working physicists, rather than only to saturated public benchmarks. He also treated verification and interpretability as an open problem in its own right: results that are correct but too hard for human experts to follow quickly.