The Checkable Loop
AI research is the next checkable domain, and checkable domains fall fast.
With Dwarkesh Patel
Earlier this year Patel put superintelligence the better part of a decade away. Here he explains why he no longer does.
Seven steps, ending with the objections he used to hold. Tap any ▶ to hear the moment.
The speedup is already being measured
Before any forecast, he points at what is on the record. AI is speeding up AI progress at the frontier now, and three sources say by how much.
After three months embedded inside Anthropic, it concluded that progress on well-scoped tasks is running at twice the rate of a year ago, and that a "majority of this acceleration today is likely attributable to Anthropic's extensive use of AI to do R&D."
The heaviest tenth of AI users among its researchers went from about $1,000 a day of internal coding agents at the start of the summer to over $7,000 a day by the end of August.
Claude now leads 26% of its AI R&D work, up from under 1% in February.
These are the figures as he reports them. METR's own wording is narrower: about twice the rate of March 2024 to October 2025, given with high uncertainty.
What happened to math will happen to AI research
The update that moved him most was watching mathematics. He gives the climb in five rungs:
- 2023could not reliably add four-digit numbers
- 2024could barely solve a high school competition problem
- 2025a gold medal at the International Math Olympiad
- Early 2026some niche open problems
- A month agoa Millennium Prize problem, he says, solved by OpenAI in 88 hours 3:17
The fair objection is that this progress leans toward what can be checked, like finding counterexamples, and away from new theory. He quotes Toby Ord saying so 3:47, then turns the sentence around:
"Current AI systems have solved a Millennium Prize problem, all without the ability to ask powerful questions or find profound new fields of mathematical inquiry."4:02
And that narrow skill is worth more in machine learning than in mathematics. Mathematicians want understanding; a proved conjecture matters mostly for the theory it draws out. Labs want a recipe that trains a better model.
"But in AI research, the priorities are exactly reversed."4:53
His example is the transformer 5:42. Its pieces, attention and residual networks among them, were already known. What was missing was the right combination, found through research taste and a great many small experiments. Persistent, intelligent recombination is what the models are already good at, so he calls the transformer:
"the kind of thing that Claude 7 could have invented if Chris Olah asked it to try really hard to get this attention thing to work better."7:24
He grants that parts of AI research are less checkable than math 7:37, which is why the labs still employ researchers.
A median of 2028
The milestone is specific: automated researchers as good as the very best humans at AI research. He assumes the labs go full steam and nobody slows them.
Why so soon? In every domain where answers can be checked and practice repeated thousands of times, the models went from worse than most people to better than the best within a few years 8:40. And they generalize past their training: the model that solved the prize problem practised on easier ones 9:22.
He takes 2027 seriously. Ajeya Cotra's reading of the METR report is that, at face value and assuming the acceleration is smooth,
"this naively implies that we would reach extreme superintelligence within a year."Ajeya Cotra, quoted at 10:40
His median sits a year later only because he doubts how far that kind of straight-line model can be trusted.
Faster, smarter, more. Then a coin toss.
Once AIs match the best researchers, he gives three reasons algorithmic progress speeds up, each with its own limit.
Today's AIs already work at about ten times human speed, and they would work around the clock: roughly fifty times more work per calendar week. Progress would not go fifty times faster, because experiments still wait on compute.
His sense is that about 80% of a frontier lab's algorithmic progress comes from its top fifty researchers. In the automated lab, every worker is at that level.
AI research is closer to formalizing a theorem than to writing a novel: many independent pieces, each checkable. A swarm helps, with steeply diminishing returns.
Whether that pace holds is the larger question 14:27. Two rates pull against each other. The researchers keep improving. The easy ideas keep running out, and already do: he says labs quadruple their R&D compute and double their staff every year just to hold the same rate 15:04.
Three to six years of progress at today's rate, packed into one calendar year 15:39. Then things simmer until more compute is built. Even so, he says, it would feel insane: a jump as big as GPT-3.5 to GPT-6 Astra, in twelve months 16:20. The endpoint is the same as the other case. It takes years to reach it, not months 25:39.
With no new compute at all, the swarm makes smarter and smarter versions of itself until, within a matter of months, "we end up with a superintelligence that scrapes against the physical limits of our current chips" 15:17.
A sketch of his two cases. The shapes are ours; only the labels are his.
"I think it's basically a coin toss."16:01
It won't know everything. It will learn fast.
A mind trained in a datacenter has never run a company, argued a case or fought an election. How would it beat people whose skill comes from years in the world?
His answer is not that the answers get baked into the weights 18:15. Skill at any hard task needs domain knowledge and something more general: finding information, making a plan, pushing through mistakes, keeping track, working with others 18:32. The second part is what training at scale would build, through tens of thousands of times more practice environments than exist today 18:52, and through backtests of history. One he imagines: given nothing written after 1700, can the model find general relativity 20:07?
For why checkable practice should transfer to everything else, his strongest evidence is us. Evolution checked one thing, whether genes were passed on, on tasks like hunting and keeping one's standing in the tribe.
"But in optimizing us for survival on the savanna, evolution built a general-purpose learner."22:15
Such a learner would still need real experience, only less of it than we expect. Napoleon and Rockefeller needed a couple of years, not decades 23:38. An AI shadowing a chief executive could read every message the company has produced on its first morning and interview every employee by the end of the day 24:17. What it cannot hurry is feedback that takes months to arrive, like the next quarter's results.
Three kinds of work, each with a scoreboard
This is his appendix, and the load-bearing part. AI R&D falls into three buckets 27:35, and each has a number that can be measured inside a datacenter.
Loss on held-out data. A model can be trained to find ways of reaching a fixed loss with less compute. Even taste has a test: predict whether a gain at 100 million parameters will hold at 10 billion 29:25.
How much compute, data or time a new set of environments takes; how quickly a model learns a task it has never seen; and benchmarks built from real problems the lab's own researchers solved.
The most checkable of the three: GPU utilization, latency, the arithmetic in a forward pass.
Being checkable matters twice 28:03. A lab can train a model directly on these tasks, over thousands of rounds. And each agent can test its own contribution before merging it, which is what lets thousands of them work side by side.
The same goes for the weakness he minds most, sample efficiency, which he calls "AI's equivalent of a Millennium Prize problem" 27:08.
Where he argues against himself
He treats the bottlenecks as real. They are why he expects a spark-sized jump in 2028 and not an overnight singularity. His claim is only that they do not close the road.
"the neck of the bottle is actually pretty wide."33:53
His own objection until recently: every session is a model's first day on the job. He now thinks longer in-context learning may be enough, and that a model without the ability might still be the one to solve it. It remains the reason for his 25% past 2032.
Choosing which hill to climb is itself research taste. His reply: the transformer was built to beat translation benchmarks that already existed, and broad outer tests are not hard to write.
Data keeps growing, with the AIs compiling it. His own research with Jerry Han found that data improvements gave more than three times the pretraining progress that model improvements did 38:54, and that work is ordinary ML research.
Math needs pen and paper; AI research needs experiments. He answers that each agent would still have a great deal of compute, could spend labor to save it, and could learn more from small runs.
On continual learning he leaves an image: models solving prize problems with no long-term memory.
"It's like watching somebody do cartwheels down the track and still keep pace with Usain Bolt."36:13
He ends on a line from a researcher who thought very small networks could teach a great deal about large ones:
"The gods above must work from the same principles as the ants below."43:54