Hacker News Reader

> We don't know what the author asked it to do, but this smells like the problem started at least several messages before that.

Author here. It was what I assumed would be a fairly simple task, fixing some duplicate closing frontmatter delimiters.

I think the LLM took a wrong turn early on, and then just spiralled. It was morbidly fascinating watching it rabbit hole.

View on HN

> It was morbidly fascinating watching it rabbit hole.

Oh yes, it's like slow-motion train wreck.

Once it gets off the rails, the best thing is to nuke the context, reset the state (git reset --hard or equivalent) and retry, this time with foresight.

> this time with *foresight*.

So much this! Nondeterministism in LLMs does not mean you simply "reroll" your prompt and hope for the best; the incremental reprompting is what helps.

Or you just code it yourself.

Reply on news.ycombinator.com