Zvi Mowshowitz thinks the world should prepare for the chance that superintelligence arrives in 2027, and that any serious plan must include a way to slow the growth in capabilities that follows. He makes the case in “Quickly, There’s No Time,” an essay on his Substack, written in response to the METR report on AI-driven acceleration inside Anthropic. The forecasts below are his reading of other people’s updates, and he labels his own numbers as judgment.

The trigger was METR’s estimate that Anthropic now makes AI progress at close to twice its historical rate. METR would not be surprised by anything between 1.3x and 3x, and it guessed that “perhaps 80%” of the speedup comes from AI itself. Eli Lifland of the AI Futures Project, the team behind AI 2027, says that moves the schedule five months ahead of their scenario. Lifland’s median for a Superhuman AI Researcher (SAR, an AI that can take over AI research) moved from about 2030 to 2027. His group also gives a 35 to 65 percent chance that a Top-Expert-Dominating AI (TED-AI, an AI able to take over every office job) arrives before 2028.

Mowshowitz is more cautious than they are. He puts a 2027 SAR below 50 percent, though not far below, and he would be surprised but not shocked to see nothing close by the end of 2028. His test for readers is blunt: anyone assigning no meaningful chance to superintelligence in 2027 is, in his view, making a clear mistake. He also worries that some people now treat it as the baseline case instead of one possibility.

The essay collects other forecasters. Paul Christiano, an OpenAI board member speaking in a personal capacity, now sees a material risk of AI takeover within 6 to 12 months. That is a large jump from his September estimate of 4 percent for loss of control within one year and 15 percent within three. Dwarkesh Patel moved his median for AI systems matching the best human AI researchers to 2028, with a 25 percent chance it lands after 2032, after saying earlier this year he would have answered “the better part of a decade.” Buck Shlegeris gives one in three odds of superintelligence within a year. Nate Soares of MIRI sits at the other end, calling it more likely than not that the world reaches 2028.

Mowshowitz pushes back on the skeptics, grouping their objections into roughly ten categories. The first five he calls exceptionally poor. He finds the stronger ones to be that AI progress will fade as problems get harder, or that gains in verifiable domains like math and code will not carry over to messier work. Arguments that “AI will need X and lack it” have, in his words, predicted about seven of the last zero walls. He adds that Patel’s own spark scenario, a single big jump that stalls out, deserves far less than the 50 percent Patel assigns.

The sharpest disagreement is over definitions. Dean Ball, who is listed in the essay with an OpenAI affiliation, splits recursion into a weak form (technology as one input to making better technology) and a strong form (a technology that plans and executes its own successor with no human oversight). Ball opposes handing that control to AI but does not want a ban on using AI to build AI, warning that panic could centralize power in the state. Mowshowitz replies that a rule against all AI help is close to a ban on progress, yet some limit on intensity must still be priced and set. He reads weak recursion as the precursor to the strong kind. He also reads Sam Altman’s remark that OpenAI “could be pushing more towards the RSI spectrum” as a statement that doing so would not be safe.

His prescription has two halves. Any plan must be ready for AI that runs its own R&D end to end at Anthropic or OpenAI by the end of 2027, with rapid recursive self-improvement to follow, and it must include mechanisms to slow the growth in capabilities. It must also survive a slower timeline, because small errors in an exponential become large differences in dates. He lists backers of slowing or pausing frontier development, including Nate Silver, Elizabeth Warren, and the Financial Times editorial board.

Mowshowitz admits he has seen no plan that makes a 2027 superintelligence likely to go well. For companies and governments, his argument reduces to a planning test: a roadmap that only works if the crucial events come later than 2027, or only if they come exactly then, fails by his standard.

Based on Zvi Mowshowitz’s essay “Quickly, There’s No Time” on his Substack, thezvi.substack.com.