• by bigyabai on 5/17/2025, 5:27:43 PM

    > it's very clear that with enough permissions / MCPs, the LLM could do its own thing ~indefinitely.

    What exactly is your proof?

    I'm not opposed to the idea, but in my experience LLM performance falls through the floor once you run out of input tokens. It seems inevitable that AI will continue writing code until it cannot understand what it's doing anymore.