Jack Clark Sees AI Speeding Science as Agents Reach 27 Hours

Update, 21 November 2026: The Lean FRO and the Mathlib Initiative, which maintain the Lean proof checker and its maths library, said on Saturday in a joint statement: “We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct.” They describe Alpöge’s proof as a Lean library of 17 million lines of code, checked by several independent proof checkers. Alpöge calls the argument “incredibly long at the moment”, and the statement looks forward to work “that will help explain the new ideas in this proof”. Our report on the certification.

Jack Clark, a co-founder of Anthropic, says more powerful AI will speed up the rate of scientific progress. He makes the case in an essay for The Economist’s World Ahead 2027 and in a post on X on Saturday.

“How we prepare for this, build the infra to support it, and deal with the associated safety issues will define 2027 and beyond,” Clark wrote on X. The Economist describes him as Anthropic’s co-founder and head of public benefit.

This week’s research gives his claim some numbers, tracked on our AI model benchmarks page. The bill for the computing behind it sits on our AI compute commitments tracker.

Business Pill: The Cost of Checking

An evergreen Business Engineer explainer on the cost of checking work. It teaches the general idea only and does not discuss Clark’s essay.

The key insight: Anthropic co-founder Jack Clark says AI will speed up science; this week METR put agents at up to 27 hours on checkable tasks, and about 400 of 663 Godel theorems Claude extracted were verified as stated, while a Riemann claim remains unverified.

What Clark Wrote

The essay’s headline is “AI will accelerate scientific progress, says Jack Clark”. Its standfirst adds: “But care will be needed.” Most of the essay is behind The Economist’s paywall, so we quote only the parts we could read.

It opens with his teenage fear: “Experts warned that the 21st century was going to be difficult—that population growth, climate change, resource extraction and other crises could cause a collapse of the human species.”

Of 663 theorems Claude extracted from Godel's notebooks, about 400 were verified as stated in the Coq proof ch
Of 663 theorems Claude extracted from Godel’s notebooks, about 400 were verified as stated in the Coq proof checker, about 200 repaired and about 50 false (Jan von Plato, as reported by The Guardian).

What This Week’s Evidence Shows

AI agents can now work alone for far longer. METR said the length of task frontier agents finish half the time is doubling about every 3.6 months. It now reaches up to 27 hours on tasks that can be checked automatically.

When Claude pulled 663 theorems out of Kurt Gödel’s notebooks, about 400 were verified as stated in the Coq proof checker, about 200 needed repair and about 50 were false, according to the logician Jan von Plato.

By our arithmetic, about 90% held up as stated or after repair, and about one in 13 was false.

The boldest claim is unchecked. An Anthropic mathematician, Levent Alpöge, says he found a proof of the Riemann Hypothesis with Claude. No independent verification has been published, Scientific American and TechCrunch reported.

What ‘Build the Infra’ Costs

Clark names building the infrastructure as one of the tests. Epoch AI says the capacity of AI chips to run agents grew 4.6 times in a year, against 2.6 times for raw computing power. By our arithmetic, that is about 1.8 times as fast.

Bar chart: by Epoch AI's index, the capacity of AI chips to run agents grew 4.6 times from Q2 2025 to Q2 2026, against 2.6 times for raw computing power
By Epoch AI’s Inference Performance Index, the capacity of AI chips to run agents grew 4.6 times in a year, against 2.6 times for raw computing power.

Anthropic has signed up for the bill. Its final prospectus puts its commitments at about $900.9 billion, about 3.9 times the $232.4 billion of cash and short-term investments it shows after the IPO, by our arithmetic.

The Structural Read

As we read it, Clark’s own three tests (preparing, building the computing, handling safety) map onto this week’s numbers: longer agent runs, a $900.9 billion compute bill, and results that still need checking.

As we read it, verification is the bottleneck to watch: a proof checker caught the false Godel results, and the Riemann claim has no such check yet.

Jack Clark, on X, 21 Nov 2026

“How we prepare for this, build the infra to support it, and deal with the associated safety issues will define 2027 and beyond.”

Three Implications

FOR RESEARCHERS METR says agents trail human experts on tasks without clear feedback.

FOR AI LABS Clark names computing infrastructure and safety as the 2027 tests.

FOR READERS the Riemann claim has not been independently verified.

What Is Not Established

METR adds a warning that matches Clark’s call for care. On tasks without clear feedback, it wrote, “AI agents appear to have worse judgment than human experts”.

A proof checker flagged about 50 of the Gödel results as false. No independent check of the Riemann claim has been published, Scientific American and TechCrunch reported.

This Week In The AI Supercycle

A World Without an Order, and an AI Central Bank Without a Successor

This week’s issue: the week AI memory stayed largely sold out while the people holding the shares paid: why the AI Supercycle is not a dot-com replay, read through geopolitics, macro and technology, with the gap between building and financing the widest we’ve recorded.

Read the full issue →

Business Engineer Framework

The AI Supercycle

The Business Engineer map of where AI capability turns into business leverage.

Read the Map of AI →

The Bottom Line

Anthropic co-founder Jack Clark says AI will speed up science, and that preparing for it, building the computing and handling safety will define 2027. This week’s evidence: agents working up to 27 hours on checkable tasks, most of Claude’s Gödel results holding up, and a Riemann claim still unverified.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Clark’s words come from his post on X and from the parts of his Economist essay that are readable without a subscription: the headline, standfirst and opening paragraph. We have not read the rest and do not characterise it. The METR, Epoch, Gödel, Riemann and prospectus figures come from our earlier reporting, linked above, with their original sources. The 90%, one in 13, 1.8 times and 3.9 times figures are our arithmetic.

Sources: The Economist, ‘AI will accelerate scientific progress, says Jack Clark’, The World Ahead 2027, 18 Nov 2026 (opening readable; rest paywalled) · Jack Clark on X, 21 Nov 2026 · FourWeekMBA on METR’s time-horizon update, 20 Nov 2026 · FourWeekMBA on Claude and Godel’s workbooks, 16 Nov 2026 · FourWeekMBA on Epoch AI’s Inference Performance Index, 18 Nov 2026 · FourWeekMBA on the Riemann claim, 20 Nov 2026 · FourWeekMBA on Anthropic’s compute commitments, 19 Nov 2026

Scroll to Top
FourWeekMBA