Yes, LLMs are trained one token at a time and produce one token at a time.
That does not mean that we understand exactly how the encode or process information within their neural networks. Even the most advanced researchers don’t claim that knowledge.
Saying “They don’t reason” without exact knowledge of what they do or a precise definition of what counts as “reasoning” isn’t very useful.
If it looks like they do its reasonable to suggest they do in the same way I don’t know you do I only know it looks like you do. Birds do it, bees do it, even educated fleas do it.
Its a “semantic knowledge retrieval system”.
AI isn’t just language statistics. The model has the relationships between words. It is knowledge.
At that point I interjected the following: I agree that it is knowledge. I’ve often noted that a book which contains nothing but printed words obviously contains knowledge. I’ve also noted that the book is not intelligent, no matter how many pages of knowledge it may contain.
But it is also true that the “deep statistics” — by which I mean the statistics represented by sentences of words, paragraphs of sentences, and chapters of paragraphs — can also be fully modeled — purely as statistics — so that the knowledge represented by the detailed words in such a book could be reproduced using just those statistics.
Now, if a query system is added, which allows the relevant portion of that statistically-stored book’s knowledge to be regurgitated, then what we have is a semantic knowledge retrieval system which allows us to query the book’s stored knowledge. If we had never read that book, such a system might stun us with what it says by appearing to “know” so much more than we do. And we might come away from the experience deeply impressed. But even so, that doesn’t change the fact that what we have constructed is a semantic knowledge retrieval system
- Steve Gibson - Security Now - Episode 1098
What’s the quoted text from? It doesn’t seem to appear in the article
Sorry, its from Steve Gibson’s podcast “Security Now”. Editing my previous post to include the source.
So you don’t know how it works, but you’re positive that it works in this extremely particular way? That’s a minefield of a stance to take.
Are you suggesting that apparent reasoning is becoming an emergent property instead of its own distinct “circuit” of thought?
I wonder how we could test for this. Because that would be fun to test for. And it wouldn’t involve any weird security implications (at least, not directly).
