> you put them in a ralph loop they can go far, far away
replyThe point is they mostly wind up somewhere stupid, and it takes expertise to spot and correct that. (Maybe that changes with further development.)
With enough time (and tokens), they'll eventually recover.
replyIt's essentially a "brute force" approach, but in most cases, they only need to succeed once.
> With enough time (and tokens), they'll eventually recover
replyThe 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?
In a ralph loop you start any iteration from scratch and feed the prompt with last X iterations in order to avoid getting stuck.
reply