Dwarkesh Patel Argues Checkable Practice Can Build General Intelligence
In Patel's account, drills with clear answers lifted AI past experts and helped it solve problems it never trained on.

Dwarkesh Podcast · The avalanche that hit math is coming for AI developmentIn a long video essay, Dwarkesh Patel makes the case that practice on tasks with checkable answers can build general ability that carries into harder work.
Patel says in fields like competitive programming contests and well-scoped open math problems, problems with clear solutions and room for repeated tries, AI systems went from worse than most people to better than the best people within a few years.
His example is an OpenAI model that apparently solved a Millennium Prize problem, one of a set of famously hard unsolved math problems, after training on smaller and more manageable math problems, and he presents that leap as evidence that models generalize far beyond their drills.
Patel expects the same transfer from math drills to AI research, where repeated training across many checkable research tasks, plus each new generation being broadly smarter, would make models strong even on work that cannot be directly checked, such as betting years ahead on the right direction.
For comparison, Patel describes evolution as a trial and error process that keeps what works, guided by a crude pass or fail test of whether genes got passed on.
The practice tasks were narrow, hunting, fending off predators, and keeping standing in a tribe, yet that process built a general purpose learner now able to write a prize winning biography, make advanced semiconductors, lead nations of millions, and invent new mathematics.
Training today is deliberate rather than blind, uses tasks much closer to modern work, and includes checks on difficult problems that were kept out of training to test whether skills carry over.
Patel concludes that such training could produce a flexible learner that ramps up quickly in new settings, even ones it never directly practiced.
Questions people ask
Why do checkable answers matter so much for training AI?
A checkable answer lets the computer score itself without a human judge. That makes it possible to run thousands or millions of practice rounds, keep what works, and fix what does not, which is how systems improve quickly in coding and math.
What is a held-out task?
It means a test the model did not see during training. Builders withhold some hard problems and then check whether practice on easier problems helps the system solve the new ones, which is direct evidence of transfer rather than memorization.
Would an AI trained this way immediately beat human experts at a new job?
No, in this account it still needs time in the world. Even a very capable learner would need real experience with customers, labs, markets, or other messy feedback, but Patel predicts that ramp up could take months, not decades, because the underlying learning ability is already general.
