Hacker News Reader

No, if you switched to Japanese the LLM wouldn't ignore it, it would "assume" there's a reason for that. The same if the previous iteration of the LLM switched to Japanese. Else you're expecting an LLM to ignore its own previous outputs and restart "thinking" from scratch with every token?

View on HN

It's situational and I suspect there are situations where it would and others where it wouldn't. Would depend on the almost infinite variables of how training was done. You'd be right that it'd be likely to switch but while it's possible it's due to temperature, there are just so many things going on. But it would be one sensible explanation among many.

Reply on news.ycombinator.com