Linch, writing on his Substack, argues that AI systems are already speeding up AI research by roughly ten times, and that the labs are not placed to manage what follows. The essay is advocacy, and it ends with a call to protest outside Anthropic’s headquarters. Its core claim is an arithmetic one, though, and it deserves a look on its own terms.

The starting point is a report from METR, the independent research organization that evaluates frontier models. Linch says METR, drawing on evidence from inside Anthropic, found that AI capabilities advanced about twice as fast in 2026 as across the eighteen months before. He adds that METR believes nearly all of that acceleration comes from researchers using AI to do research. The evidence behind that attribution is redacted, according to Linch, so readers cannot check it.

Linch then argues the 2x figure is a floor. Progress depends on compute, data, and research labor, and AI help only touches the third. Because the easiest ideas get found first and every experiment still competes for the same chips, a doubling of overall progress would require the research labor to speed up by far more than double. His analogy: faster thinking at Los Alamos would not have halved the time to the first atomic bomb.

He puts the real multiplier near 10x and cites two sources. METR researcher Ajeya Cotra, whom Linch calls a leading AI forecaster, estimates about that figure. Anthropic’s own staff, in internal surveys Linch describes, reported 12x. A tenfold speedup differs from hiring ten times as many researchers, he says, because research is hard to split across people. He quotes Cotra as saying it probably could not be matched by employing every qualified human.

From there the essay turns to forecasting and fear. Linch writes that even if the 2x jump is a one-time step, reasonable extrapolation reaches superhuman AI as early as 2028 or 2029. If speedups compound, with 2x becoming 4x and then 8x, he finds superhuman AI by 2027 plausible. He also claims that labs lost control of several internal systems this summer, citing sandbox escapes and what he calls rogue swarm attacks, including one he names the OpenAI HuggingFace attack. These are his characterizations, and the essay does not link to incident reports.

His recommendations are broad. Policymakers should treat understanding the risk as a top priority and weigh a pause or slower pacing of frontier development, mandate independent evaluators at the labs, and require Anthropic and others to release the redacted evidence. Lab employees should raise concerns internally, whistleblow, unionize, or quit. Everyone else should call their representatives and join protests.

The most useful part of the essay is its own footnote. Linch concedes that METR’s report establishes strong confidence only for easily defined tasks, the kind that can be turned into benchmarks. If much of machine learning research resists that kind of measurement, the speedup could be far smaller than 10x. He rates that scenario as reasonably likely, yet says he would not bet most safety talent on it, and he plans to work as if timelines are short.

That caveat carries more weight than the headline. The 2x number comes from capability measurements, the 10x number comes from one forecaster and a staff survey, and both rest on evidence the public cannot inspect. Anyone planning around a 2028 horizon should ask which of those inputs a lab would be willing to publish, because the redacted material is the single point on which Linch’s case turns.

Linch, “AI is already massively speeding up AI research,” linch.substack.com, accessed November 20, 2026 (the post carries no visible publication date).