3 ms·

> you put them in a ralph loop they can go far, far away The point is they mostly wind up somewhere stupid, and it takes expertise to spot and correct that. (M

• vb-8448 6mo ago
  With enough time (and tokens), they'll eventually recover. It's essentially a "brute force" approach, but in most cases, they only need to succeed once.
  • JumpCrisscross 6mo ago
    > With enough time (and tokens), they'll eventually recover The article’s point is this is not true. They wind up in bullshit attractors where they hit a wall and then get lost within their muddled context window. > they only need to succeed once Yet they don’t. Not on their own. Like, you haven’t had an LLM get stuck in a stupid loop where you point out the flaw and then it gets unstuck?
    • vb-8448 6mo ago
      In a ralph loop you start any iteration from scratch and feed the prompt with last X iterations in order to avoid getting stuck.