Recursive self-improvement has started
Close Listening The Checkable Loop AI research can be checked from inside a datacenter, and domains that can be checked fall fast. Dwarkesh Patel's case for automated AI research by 2028 in seven steps, from the speedup already measured to the objections he used to hold. Listen closely →Featuring
Dwarkesh Patel changes his mind in public. Earlier this year he would have put superintelligence the better part of a decade away. Now his median for AIs that match the best human AI researchers at AI research is 2028, with 2027 a live possibility and at least a 25% chance it waits until after 2032. From there he expects AIs at least as good as the best human experts in every field by 2030.
The argument runs through mathematics. Models that could not reliably add four-digit numbers in 2023 are, on his account, now clearing well-scoped open problems, and he argues that AI research resembles that kind of work more than it resembles mathematics proper: labs want recipes that train better models, not theory, and the transformer itself was a persistent recombination of ideas already in the literature. Because pretraining loss, sample efficiency and inference cost can all be measured inside a datacenter, he thinks the research is verifiable enough to train on directly. What follows automation is, he says, a coin toss between a software-only singularity and what he calls a “spark”: three to six years of progress in one calendar year, then a wait for more compute.
Two things to hold beside it. He opens on METR’s assessment of Anthropic as progress at “2x what it was a year ago”; METR’s own baseline is March 2024 to October 2025, and it gave the figure with high uncertainty. And the last third is where he argues against himself: he takes the objections in turn (continual learning, which was his own until recently, then choosing the right proxies, data, and experiment compute) and says plainly that they lower the odds of the fastest takeoff.