@julienw @thomasfuchs yeah had me thinking that with hidden thinking tokens one could make a working version of guess who…I also couldn’t reproduce what she did. Maybe with a smaller model.

@flq @julienw They're hiding this now behind tokens that you can't directly see, but the principle remains exactly the same. It's just more smoke and mirrors in newer harnesses/models.

@thomasfuchs
Yes, but that is what is the most easy to improve.
It's really easy to give an LLM memory and a few tools.
The core model works just as the lady says, but none of the commercial products are merely that, like you mentioned.
@flq @julienw

@julienw @thomasfuchs

Tried once with free ChatGPT, worked finel. (Only used 7 questions to get the answer but all responses were correct.)

Of course, length may matter, but also how it updates its token analysis and weights as we go along.

I've tested it in longer conversations about things that interest me, where it veers back and forth between nonsense and useful remarks. With nudges. Occasionally I find a simple "That's wrong" can be surprisingly helpful in putting it back on track.

@glc @julienw the point here is that it doesn't have permanence and isn't a conversation partner.

every time you talk to it, the entire previous conversation gets fed back to a new instance of it (which doesn't have any memory from before) and then it generates a statistically likely reply

@thomasfuchs @julienw

That is indeed the intended point.

One that has been discussed repeatedly and widely since the "stochastic parrot" paper but has not penetrated to the audience that the post aims at.

But we are noticing that the illustration used does not appear to correspond to the actual behavior of popular models.

It works a good deal like a search engine in a linguistic space. If it finds something useful, you can use it, if not, not.