Korva is a scroll of the internet’s best ideas.
AI can run out of hard problems
How self-play works
Game-playing AIs like AlphaZero use self-play, which supplies an infinite curriculum because the opponent is always about as strong as the learner. Within about a year, game-playing AIs went from the edge of the top players to beating the world champion and then to orders of magnitude stronger than any human alive.
Language models differ
Language models trained with reinforcement learning, in the ways used today, get a problem and try to solve it. If the problem is so easy that the model solves it in a second, it is not really learning anything. As models get smarter, many of the questions that can be asked are too easy, and it is hard to challenge them.
Where the wall bites
OpenAI has not hit this wall yet, and ways around it exist. Where it does bite, progress in a domain like math could stall instead of exploding the way it did in games.
How did AlphaZero learn without human games?
AlphaZero started from random play and improved by playing against copies of itself, with no human game records. DeepMind's 2017 report said that after nine hours of training it beat the Stockfish chess engine in a 100-game match. It won 28 games and drew 72. Each win made the next opponent a little stronger, which kept the training challenging.
Why does an easy problem teach a model nothing?
Many reasoning models are trained with methods that sample several answers to the same problem and reward the ones better than the group average. If every answer is correct, all get the same reward, so the difference between them is zero and the model receives no signal. The same happens when every answer is wrong. DeepSeek's 2024 DeepSeekMath paper introduced this group-relative approach, called GRPO. ByteDance's 2025 DAPO method drops prompts where every sampled answer is right or wrong.
How can AI invent its own training problems?
DeepMind's AlphaGeometry, published in Nature in January 2024, started by sampling random geometric constructions. A symbolic engine then derived the facts that followed from each one, and the chain of deductions became a theorem with its proof. Repeating this produced about 100 million synthetic examples with no human demonstrations. The system solved 25 of 30 Olympiad geometry problems in its benchmark.
More from Dwarkesh Podcast
