November 21, 2026
Heard in AI

Zvi Mowshowitz says self-improvement skeptics forget the home runs

Writer Zvi Mowshowitz accepts that AI self-improvement would probably fizzle without new breakthroughs. He calls that result irrelevant because breakthroughs keep happening, and says people he trusts expect 2027 to be "completely crazy."

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on The Cognitive Revolution, episode published October 24, 2026

Zvi Mowshowitz thinks a skeptical analysis of self-improving AI probably gets its own math right and still misjudges what is coming. He compared it to forecasting a baseball game by working out "how many points we expect to score in this game if nobody ever hits a home run." In real games, he said, "every now and then someone does hit a home run and you have to take that into account."

Mowshowitz, a writer who covers AI developments for his readers, made the argument on the AI in the AM segment of The Cognitive Revolution. Nathan Labenz and Prakash Narayanan host the segment, and the episode was published October 24, 2026.

The loop in question

Recursive self-improvement, often shortened to RSI, describes AI systems doing a growing share of the work of building better AI. Each new generation of models helps researchers run experiments, write code and test ideas faster, which could bring the next generation sooner. If each round speeds up the next by enough, the process feeds itself and accelerates. Mowshowitz called this going "critical." If the boost is too small, the gains taper off.

Labenz called the state of this loop "maybe the most important question in the universe right now." He set two kinds of evidence against each other. AI companies have published blog posts measuring how heavily their staff use AI, down to token use (tokens are the chunks of text models read and write) and experiments per researcher. On the other side are analyses such as one by Ramez Naam, which conclude that the loop looks too weak to sustain itself and will "peter out." Labenz said a gain of about 1.6 times as many experiments per researcher did not look to him like exponential takeoff.

What Naam's analysis found

In his September 27, 2026 analysis, Naam asked two separate questions: whether AI is automating parts of research, and whether that automation makes the feedback loop self-sustaining. He focused on OpenAI's reported August experiment rate, roughly 1.6 times the 2025 average per researcher, rather than on token consumption or the volume of code written.

Naam assumed a capability increase of about 16 points on a capability index. On that basis he estimated that each additional point brings 2% to 3% more research productivity, while his model put the threshold for a self-sustaining loop at 15% to 19%. Those figures rest on calibration assumptions, not on a controlled measurement of how much progress AI itself caused. Experiment quality, extra computing power and the gap between internal research tasks and benchmarks all complicate the comparison. Naam allowed that better research judgment, conceptual breakthroughs, or interacting hardware and economic feedback loops could change the result.

Mowshowitz's answer

Mowshowitz read Naam's conclusion as conditional. Suppose there are no new fundamental leaps, no new scaling laws and no paradigm shifts, and AI simply tries to brute-force its own improvement in software with existing methods, making current systems more efficient. Then, as he summarized it, the loop runs out of steam before it goes critical. Stated that way, he said, the result is "probably true but also seems completely irrelevant."

That is where the home-run comparison came in. A home run is unlikely in any single at-bat, but a forecast that assumes none will ever happen gets the final score wrong. Mowshowitz sees breakthroughs in AI research the same way: each one is unlikely, but they happen often enough that a forecast leaving them out misses the final score. Because Naam's own analysis allows that breakthroughs could change his result, the two appear to disagree less about the arithmetic than about how much weight the no-breakthrough case deserves.

The evidence he points to

Mowshowitz said the numbers are already going critical at Anthropic and OpenAI, citing exponential growth in tokens used per person and in lines of code committed. He acknowledged that this growth is not necessarily as efficient "in some important sense," then added: "give me a break."

His other example came from NVIDIA, the chipmaker. He said the company had released evidence that week showing that the time it takes to double the efficiency of CUDA, NVIDIA's software for programming its chips, had shrunk dramatically over the course of 2026. He said this was "obviously" not true recursive self-improvement. To him it showed how much faster people can now do technical work that closely resembles AI research and development. He also valued it as a concrete, objective signal from a hardware company rather than a major AI lab.

Who shares the view

"The people I talk to who I trust the most on this believe that 2027 is going to be completely crazy and 2028 might be unrecognizable," Mowshowitz said.

Labenz guessed that many of those people work at Anthropic, where he described this as broadly the prevailing view, and noted that Anthropic's Jack Clark had put it in print. Labenz added that Mowshowitz's view of the technology's trajectory is probably close to Anthropic's, even though the two disagree on policy.

In Import AI 476, published October 12, 2026, Clark wrote that he now considers full recursive self-improvement more likely to become possible in 2027 than in 2028. He attributed the change to advances in mathematical creativity and experimental research judgment. Clark separates carrying out experiments from choosing which ones are worth running and interpreting their results. That second skill is the research judgment Naam named as something that could change his own conclusion.

Mowshowitz described a wider circle than one company. It included people at OpenAI, Anthropic and Google DeepMind, people working on policy, people at METR, an organization that evaluates AI systems, and people working on alignment, the effort to make AI systems pursue intended goals. A "huge percentage" of them, he said, have converged on the view that things are already out of hand at a low level and that RSI, in a slow-takeoff style, has "already kind of begun." He expects things to look much faster by the middle of next year.

What faster could look like

Mowshowitz sketched what that might mean for the release cycle. Advances the size of a current model release might arrive every month instead of every two months, then every two weeks, then every week. "Stuff like that could happen," he said, adding that "we are kind of on an AI 2027 style timeline right now" and that it looks "very eerily similar" to that scenario.

AI 2027 is a scenario published in April 2025 by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean. Its central mechanism is the one Mowshowitz described: increasingly capable AI researchers speed up the development of their successors. In its racing branch, a superhuman coder arrives in March 2027 and a superhuman AI researcher in August. The authors chose 2027 as the single most likely year in their estimates, not as a confident deadline, and later annotations on the site describe longer median forecasts.

Mowshowitz said he already feels the pace in his own work. His job will get much harder, he said, and he will have to triage far more information. Some developments that would have been a major news story six months ago now get a single sentence in his writing or are "silently skipped," because so many similar things have already happened and neither he nor his readers have the time.

Share this article

Go to the original

Sources & further reading

  1. 01
    AI 2027
  2. 02
  3. 03

Connected ideas and articles

From the conversation

Podcast episodes

The Cognitive Revolution

AI:AM Highlights: Clean Neurons, Black October, the GPU Credit Crunch, and Buying the Supply Chain

Episode published This article draws on 0:00–0:41 and 5:11–9:48 (approximate times)

Article history

Updates to this article

Tags