Jack Clark, a co-founder of Anthropic, told an audience of bankers, economists and policymakers at the IMF’s Bangkok meeting on Saturday, October 17, that the singularity has begun. He delivered the Per Jacobsson lecture and used it to argue that finance ministries should expect fast, disorienting change. A frontier lab executive making that case to the people who set national budgets is the story.
Clark, who covers the field in his Import AI newsletter, which published the text, framed the claim with some caution. In his introduction he wrote that “we all may quibble over the details” and that the world is in “the early stages” of the singularity. The speech itself is less guarded: he said the singularity “is here” and that this year brought “convincing evidence” for it. The label is his, and a reader should treat it as an argument, not a finding.
The most striking passage is about his own role. Clark describes himself as “a narrator of what is happening in AI, a principle agent behind what is happening in AI, and a translator that goes between many communities,” and says he wants to be “clear about my own culpability.” The spelling is his. That note sits in the newsletter’s introduction, which adds that parts of the speech were ad libbed and are missing from the published version, so everything quoted here comes from the written text.
His case rests on science. Clark said AI systems have proposed solutions to open mathematics problems, helped design drugs now in trials, and begun running laboratory equipment remotely. The speech does name products: Claude, ChatGPT and Gemini, which he called single systems able to work at the frontier of several disciplines at once.
The evidence is a set of benchmarks and one anecdote, none of it new measurement. Clark cited Humanity’s Last Exam, where the top score rose from 8 percent at its January 2025 launch to 67.7 percent for Claude Opus 5.5, and FrontierMath, where the best score went from 2 percent to 94 percent. He also described TheoPhysEval, a paper on unpublished physics problems, in which GPT-6-Astra reached 49 of 50 after experts found errors in their own answer key, per physicist Suvrat Raju. The top scorer on the first benchmark is Clark’s own company’s model.
The economic claim is the one finance officials will notice. Clark said Anthropic’s internal modeling, which assumes models can take on many kinds of tasks and spread rapidly through the economy, puts US GDP at 44 trillion by 2030. He gave no currency unit, and the model’s other inputs are not in the published text. He called the figure outlandish but implied by continued progress.
His policy ask has two halves. Build capacity now: electricity, data centers, and the next layer of robots and laboratories, so that a world with perhaps a billion virtual scientists working under about 10 million human ones has somewhere to run experiments. He also warned that discoveries will pile up faster than clinics and factories can absorb them. Then restrain: strict human oversight of biological research, careful decisions on AI doing AI research, and “unusual and serious rules” covering both the technology and the firms that build it. Clark said he had urged governments for a decade to build power and compute, and that few listened.
That pairing is the tension in the speech. The speaker sells the technology, forecasts its rapid growth and also asks to slow it down. Clark did not say which rules, or who would enforce them against his own industry.
For any government or fund deciding this year on data centers or laboratory automation, the 44 trillion figure is the number to test, because it comes from the vendor’s own model and not from an independent forecast.
Import AI, written by Jack Clark, issue 477, reporting his Per Jacobsson lecture delivered at the IMF meeting in Bangkok on 17 October 2026 (the newsletter page carries no date).
