IlyaIvanov0

Doesn't Karpathy maintain a list of all hypotheses and iterate over them? That means all agents know previous hypotheses and results. What's the difference?

It's true. Make sense

Curious, are you thinking about this for continuous monitoring, or more for before/after comparison when agent get updated?

Thanks for asking. Not yet, but it is in backlog. I will be doing this in the future.

Hi, author here. Would love to answer any questions you have.

The most interesting finding here is that LLMs make individuals generate more ideas but make groups generate fewer. The individual effect, in my own experience, depends entirely on how you use the tool. If you treat the first answer as the answer, you get the homogenization the article describes. If you use the LLM to attack your own framing from angles you wouldn't reach alone, you end up closer to first principles, not further. Same tool, opposite outcomes. The discipline is what differs, and most people probably default to the first mode.