One safety writer says superintelligence could land by March 2028

Nikola Jurkovic wrote in a LessWrong essay published 13 November that superintelligence could arrive around March 2028 if nobody slows the labs, which leaves roughly six to twelve months to win a slowdown. The essay is an opinion piece. Every forecast in it belongs to Jurkovic or to the people he cites.

His reasoning starts with 2027. Jurkovic expects AI to automate and speed up AI research that year. If the leading company has not solved alignment by the time that feedback loop spins up, he argues, only a pause or a deliberately slower pace of development in the coming year would avert catastrophe.

The evidence he leans on is a METR report on AI research automation inside Anthropic. Jurkovic calls it substantial evidence that superintelligence is now under two years away. He also cites Ryan Greenblatt, who tweeted that the current level of speedup would have arrived around March 2027 in the AI 2027 scenario. Jurkovic reads that as real progress running ahead of the scenario, which places superintelligence in December 2027. His inference is July 2027 without a slowdown, if the scenario modeled takeoff correctly.

Other forecasters have moved in the same direction, according to the essay. Sentinel, a spinoff from the Samotsvety group of forecasters, now has a median of February 2028. The AI Futures Project published an emergency timelines update. Jake Mendel puts better than even odds on full superintelligence within 18 months absent a slowdown, and Agus estimates about 25 percent by the end of 2027.

Not everyone cited sounds certain. Tom Davidson says he does not know whether METR is right and wishes more information were public, but adds that “this is exactly what I’d expect to see if there were going to be an intelligence explosion in 2027.” Ajeya Cotra says that taking the speedup at face value “naively implies that we would reach extreme superintelligence within a year.” Both statements are conditional, and the METR data behind them is not open to outside checking.

Anthropic’s own filing adds a data point. The company updated its S-1 on 10 November to say it has reached the Responsible Scaling Policy threshold where it “observe[s] or expect[s] double the rate of progress in AI aggregate capabilities.” Anthropic wrote that it made the call “based on what we expect from future models and our overall rate of progress, rather than on what we have observed from current models.” By the company’s own wording, the determination rests on expectation.

Jurkovic puts his all-things-considered median at about March 2028. He estimates a 75 percent chance of AI takeover if superintelligence shows up within 6 to 12 months (Paul Christiano, he notes, has called it a coin flip). Combined with his timelines, that gives him a risk of 1 in 4 to 1 in 3 within a year absent a slowdown.

His proposed remedy borrows from the AI Futures Project, which has floated a 10 percent cap on compute for AI research, a nine-month capability lag, or a one-year pause on training that leaves inference running. Jurkovic wants at least a year. He names three routes for nontechnical people: lobbying the labs directly, pushing California toward a law or executive order, and lobbying Washington. He rates the federal route as intractable without an incident about ten times worse than the Hugging Face episode.

The odds he assigns are low. He gives California’s legislature about 20 percent in ten months and federal channels about 10 percent. One lab-advocacy worker privately told him 20 to 30 percent for a lab-led pause within six months. Jurkovic estimates the safety community staffs lab advocacy with about 20 people, California with 15, and federal lobbying with 100, all full-time, and concludes “we are not on track to succeed.” In his view the missing ingredient is “political will,” not policy ideas.

For anyone planning product roadmaps on frontier APIs through 2027, the essay offers a scenario to stress test: a state or lab-imposed slowdown within a year, which its own author rates as unlikely but not negligible.

LessWrong, essay by Nikola Jurkovic, published 13 November 2026.