[seydor]

The fundamental concept is to learn the statistics of text, and in the process it models the syntax via long-range connections successfully. There is no indication that it actively generates "concepts" or that it knows what concepts are. In fact the model is not self-reflective at all, it cannot observe its own activation or tell me anything about it.

[circuit10]

There is an indication, you can find it by clicking on this post.

The self-reflection part is probably true, but that’s not strictly necessary to understand concepts

[seydor]

it is important in order for us to accept it as an agent that understands things, because self-reflection is so important and obvious to us.

[timschmidt]

I'm still waiting for someone to prove beyond a shadow of doubt that humans have a single one of these features we're debating about the presence or absence of in LLMs.

[seydor]

there is no way to prove because those are subjective to humans. LLMs would have to at least show they have a subjective view (currently the 'internal world' they report is inconsistent)

[timschmidt]

The internal worlds of humans are inconsistent: https://en.wikipedia.org/wiki/Shadow_(psychology)

[mr_toad]

> learn the statistics

> what concepts are

How do you know concepts aren’t just statistics?

[seydor]

"concept" is ill-defined , it s a subjective thing that humans invented. It is probably not possible to define it without a sense (a definition) of self.