I notice that I'm confused.
Google made an amazing AI for playing chess, by allowing it to make its own data.
Why hasn't the same thing happened for programming? Have it generate a bunch of pictures with functionality expectations (a PM basically), have it write and run code, then check the output against the requirements it created, then try again when it doesn't come out right.
This is even easier where the pm is unnecessary - leetcode, codewars, euler...
You could also pay PMs to work with the AI developers, instead of the code tutors xAI is hiring.
There seems to be a preference to having the LLMs memorize code instead of figuring things out itself.
If you run out of things like that you could have it run random programs in different languages, only learning from those that work.
I haven't used genesis, but that also seems like a mostly-built validator for programs that AIs can use to create and train on their own data.
With the amount of compute going into training, it should be easy to create huge amounts of data?
The point was more about creating your own data being easy, just generate code then check it by running it. Save this code, and later use it for training.
If we wanted to go the way of AlphaZero it doesn't seem crazy.
De-enforce commands, functions, programs which output errors, for a start.
I didn't think of the pm as being trained by these games, that's interesting. Maybe have two instances competing to get closer on some test cases the pm can prepare to go with the task, and have them competing on time, compute, memory, and accuracy. You can de-enforce t... (read more)