undefined | Better HN

0 pointsnomel1y ago0 comments

> Or being aware of how its utterances may come across to its interlocutor.

I think this behavior is being somewhat demonstrated in newer models. I've seen GPT-3.5 175B correct itself mid response with, almost literally:

> <answer with flaw here>

> Wait, that's not right, that <reason for flaw>.

> <correct answer here>.

Later models seem to have much more awareness of, or "weight" towards, their own responses, while generating the response.

0 comments

layer81y ago

I'm assuming the "Wait" sentence is from the user. What I mean is that when humans say something, they also tend to have a view (maybe via the famous mirror neurons) of how this now sounds to the other person. They may catch themselves while speaking, changing course mid-sentence, or adding another sentence to soften or highlight something in the previous sentence, or maybe correcting or admitting some aspect after the fact. LLMs don't exhibit such an inner feedback loop, in which they reconsider the effect of the ouput they are in the process of generating.

You won't get an LLM outputting "wait, that's not right" halfway through their original output (unless you prompted them in a way that would trigger such a speech pattern), because no re-evaluation is taking place without further input.

nomelOP1y ago

> You won't get an LLM outputting "wait, that's not right" halfway through their original output

No, that's one contiguous response from the LLM. I have screenshots, because I was so surprised the first time. I've had it happen many times. This was (as I always use LLM) direct API calls. In the first case it happened, it was with largest Llama 3.5. It usually only happens one shot, no context, base/empty system prompt.

> LLMs don't exhibit such an inner feedback loop

That's not true, at all. Next token prediction is based on all previous text, including the previous word that was just produced. It uses what it has said for what it will say next, within the same response, just as a markov chain would.

j / k navigate · click thread line to collapse