Korva is a scroll of the internet’s best ideas.
AIs excel where answers are easy to grade
Why math and code are easy
Math is easy to train because it is not real-worldly: no interacting with stuff in the world, and an automated grader can check whether the answer is correct. Coding shares those advantages. It also fits the companies' strategy of automating their own jobs first and having AIs do the research.
Running a business
Running a business is a different kind of problem. The companies are not trying hard to train for it, and it would be much harder to do. Even though today's AIs can do impressive coding and hacking, they struggle to run very autonomously for long periods, especially on tasks unlike their training.
The San Francisco store
An experiment in San Francisco put that to the test. Andon Labs has a real store run by Claude, which hired human employees, bought merchandise and stock, and directed the employees to put the stock on the shelves.
Why does code give AI a clear score?
Code can be run against tests that return pass or fail, so a grader needs no human opinion. That makes it well suited to reinforcement learning with verifiable rewards. DeepSeek's R1 paper in January 2025 described training reasoning models with rule-based rewards on math and code problems.
Why did games fall to AI first?
Chess and Go have a clear win or loss, so a program can play itself millions of times and learn from the outcome. Business offers no equally crisp signal, since results arrive slowly and depend on many outside factors. DeepMind's AlphaGo beat Lee Sedol four games to one in 2016.
Why do long tasks trip up AI agents?
Each step in a long task carries some chance of error, and mistakes can build on one another. An agent that misreads the situation early may spend many steps on the wrong problem. Tasks unlike its training data give it less to rely on when it hits something unexpected.
More from The Joe Rogan Experience
