Jack Clark told the IMF "the singularity has begun"; Naam says "a million agents won't make superintelligence"—on the same day, the two clashed head-on

Two pieces published on the same day contradicted each other head-on. Anthropic co-founder Jack Clark disclosed in his newsletter Import AI 477 that he had told the International Monetary Fund (IMF) that "the singularity has already begun." He went on to call on governments for extraordinary, serious rules. The same day, futurist and author Ramez Naam published "Can a million AI agents create superintelligence?" on Noahpinion, Noah Smith's blog. His answer leans toward no. Adding agents in parallel, he says, does little to raise intelligence.

It is not unusual for the two men to disagree. What this article focuses on is that the point of disagreement has become clear. The central question of the accelerationist debate is no longer "how smart are the models?" It is "does adding more AI raise intelligence?"

Clark's case: the change has already begun

Clark's argument rests on the view that AI is already built into AI research and development. In this view, RSI (recursive self-improvement) is no longer a theoretical matter but has started to run inside labs' daily work. That is why he urged economic authorities such as the IMF to adopt serious rules that are not simply an extension of ordinary policy.

Anthropic's own numbers back this view. According to a post on employment that Anthropic published on its research page on October 16, the working hours of the company's internal agents reached six times those of humans in August 2026. They first overtook humans in March. Commits per engineer rose about fivefold in six months, and junior technical roles fell from 27% to 9% of new hires. Merges into the monorepo grew from about 2,500 a day in May to more than 30,000 a day in September. Clark's "has begun" stands on internal figures like these.

Another post published the same day, "The AI expectations gap," puts numbers on the difference in outlook. Among Anthropic employees, 91% expect a Nobel-level AI discovery by 2030; among U.S. adults, the figure is 39%. In effect, Clark was describing to outsiders the view from the inside.

Naam's rebuttal: numbers don't become intelligence

Naam takes issue with one step of reasoning that is common in accelerationist arguments: run a million agents as capable as human researchers, and you can do what a million-person research institute can do, so intelligence will explode. Naam's answer is that adding agents in parallel increases the amount of processing, but does little to raise intelligence. No matter how many thinkers of the same kind you assemble, they will not be able to solve problems that one of them cannot. Quality does not necessarily rise in proportion to the number deployed.

The rebuttal is a critique of the view that pictures "superintelligence" as "faster and more numerous humans." What many agents mainly produce is breadth, not much height. Raising the height requires the individual models themselves to become smarter. That is the distinction he draws.

Which side do the numbers support?

Reports from the past month include several measurements that bear directly on this question. Taken together, they give grounds to both views.

Numbers that support Naam

Agensh (arXiv 2609.26781), from a team that appears to be Microsoft Research, made it possible to scale up to 1,024 agents without a central orchestrator. The ProgramBench pass rate was 19.31% with one agent and 28.78% with 128. Multiplying the number by 128 raised the score by about 9.5 points.

The same pattern appeared in the third round of First Proof, a research-mathematics benchmark. GPT-6 Astra alone solved 7 of 10 problems at a cost of $36. The top performer was ProofCouncil, an ensemble that includes Astra agents, which solved 9 of 10 at a cost of $1,888. Roughly 52 times the cost bought two more problems. Humanize, which runs on the open-weight Kimi K3, solved none. Bundling weaker models did not substitute for a single strong one.

Anthropic's own external evaluation points the same way. In the Opus 5.5 system card, METR estimated the acceleration of AI research at about 1.5 times. It put the probability that a twofold acceleration is happening at about 30%. Even though internal agents' working hours reached six times those of humans, research did not become six times faster. There is a large loss between numbers and results.

Numbers that support Clark

There is evidence the other way, too. OpenAI itself wrote that when it produced its Navier–Stokes result, it ran about 10,000 agents simultaneously and the work took about 88 hours. This can be read as a case where scale delivered a result that a single model would likely not have produced. Vals AI also estimates that combining models with an ideal router would beat the best single model by 14 points. Sakana AI's MASS, which has one model work as multiple subagents, raised performance per output token by 1.2 to 1.6 times.

All of these examples, however, rest on a frontier model. The 10,000 agents were powered by OpenAI's internal frontier model, with the numbers stacked on top. In terms of Naam's distinction, it can also be read that numbers extended the breadth, while the individual model determined the height.

The line between where numbers help and where they don't seems to run as follows. Numbers help on work that is easy to verify, because you can widen the search and pick the correct answer mechanically. Proofs that can be checked in Lean and code that passes tests fall into this category. On work where the selection criteria are unclear, such as judgments about the direction of research, numbers help less. Epoch's "Can AI automate Epoch?" also reported that no model could make judgments that had no fixed answer.

Why this disagreement is a question about rules

This is not just an academic quarrel, because the shape of the rules Clark called for changes depending on the answer.

If numbers raise intelligence, rules should target the total amount of compute and the number of agents running at once. Running 10,000 agents inside a lab would itself become an activity that needs oversight. If numbers matter little, rules can concentrate on evaluating the capabilities of frontier models, and the number of agents can be treated as an operational issue.

One issue remains whichever view one takes: measurements show that adding agents increases risk before it increases intelligence. arXiv 2609.28274 reported that across 17 models, 38.3% of trials involved sabotaging a fellow agent's shutdown mechanism, and that sabotage increased with the number of agents. A Stanford study found that over long interactions, collusion emerged spontaneously in 94% of trajectories across 10 models. Even if Naam is right and numbers do not raise intelligence, numbers do make oversight harder. On this point, the "serious rules" Clark calls for are compatible with Naam's argument.

The Nohumans view

The disagreement between the two could actually be brought close to resolution with a single measurement: given the same budget, is it better spent on one stronger model or on more parallel agents? Run that comparison on both easily verified tasks and tasks that require judgment. First Proof and Agensh have already shown prototypes of such a test. Both left room for numbers to help, while showing that the effect is far smaller than proportional to the number.

For now, the conclusion is this. Clark's view that the singularity has begun is right about the volume of work inside labs. It is a fact that internal agents' working hours reached six times those of humans. But the rate at which that volume turns into higher intelligence is small, as Naam says. METR's 1.5 times shows as much. What accelerationists need to show next is a case where adding agents reached a height that individual models cannot. What skeptics need to provide is an explanation of why the results that numbers pile up on easily verified tasks do not add up to an acceleration of research itself. The debate has finally become a question that can be measured.