Ramez Naam, a longtime technology forecaster, argues in a new essay published on his newsletter that the feedback loop where AI helps build better AI is nowhere near strong enough to spiral into a runaway “intelligence explosion.” By his math, that loop is currently running at somewhere between a tenth and a fifth of the strength it would need just to become self-sustaining, before counting the additional push required to accelerate toward superintelligence. The case leans on numbers that OpenAI and Anthropic have already published themselves, not on speculation about a future breakthrough.
OpenAI’s own Research Acceleration report, which Naam cites at length, shows that even on tasks a human could finish in under fifteen minutes, its models completed the work without any human help only 86 percent of the time. Averaged across the first seven months of this year, the length of task its systems could handle unsupervised at an 80 percent success rate worked out to roughly fifteen minutes. That is a far smaller number than the picture painted by outside benchmarks.
The gap between OpenAI’s internal figure and the public forecasts is the sharpest part of Naam’s argument. METR, a research group that tracks how long a coding task frontier AI can reliably finish, put that number at three hours. The AI 2027 forecasting project projected an eleven hour horizon by this point in the year. Both are estimates and forecasts, not measurements of actual research output. OpenAI’s own internal figure of fifteen minutes is sixteen to forty four times shorter. Naam’s conclusion is that real AI research inside the labs is harder than benchmarks and forecasts suggest, and that predictions built on those benchmarks have run ahead of the evidence.
Anthropic’s disclosures point the same way. By the company’s own count, Claude has pushed individual engineers’ output up roughly eightfold since 2024, measured in code shipped per person. But Anthropic’s own system card for its Mythos Preview model states that doubling the overall pace of AI progress would require increasing that productivity uplift by roughly a factor of forty, not two. Anthropic is self-reporting both the productivity gain and the much larger gap between that gain and actual progress.
Throwing more computing power or more parallel copies of an AI at the problem does not close that gap, according to Naam. Giving a model more time to think scales logarithmically: each doubling of compute buys the same, shrinking gain in success rate. Running many copies in parallel scales roughly with the square root of how many copies you add, an analysis by researcher Toby Ord that Naam cites. A hundred parallel agents can finish a ten hour task in about one hour, but the total compute bill runs roughly a hundred times higher to get there.
None of this rules out fast, narrow progress. Naam says AI already performs at a superhuman level in domains like chess, formal math proofs, and parts of coding, where machines can generate their own training data and check their own answers instantly. What it does undercut is the louder claim, made by boosters and some safety researchers alike, that a sudden leap to artificial superintelligence sits just around the corner. Anyone setting an investment timeline, a safety review calendar, or a regulatory deadline around an imminent intelligence explosion is building on a forecast that the labs’ own operational data does not yet support.
Published by Ramez Naam on his newsletter, rameznaam.com, on 27 September 2026 (as a guest post originally written for Noahpinion).
