The autobot I'm training with reinforcement learning is too slow, so I'm using a method where I create specialized autobots for each game and have them teach my autobot.
Then, the teaching autobot also took a long time, so I asked Claude if he could play the game and use that data to teach my autobot. At first, he was negative, but after trying it out, it's been working well.
As we continue to talk, I learn how the autobot is learning and provide some support along the way. It feels like playing a mud game from the old days.
I'm currently having Claude sonnet5 play Doom stage 1.
It's been going around stage 1 for an hour, and the tokens are running out.
