undefined | Better HN

0 pointsseydor3y ago0 comments

That's very misleading. The current large LLMs learn tons and tons of "reference maps" that map how words (tokens) are referring (attending) to other tokens within a certain context. In the process they learn a map that may be the equivalent of what we humans call "syntax", and they apply it to completely different contexts and inputs. It's not memorization, nor simple pattern matching, but something that allows them to be creative

0 comments

3 comments · 1 top-level

thomastjeffery3y ago· 2 in thread

The LLM does not know what a word or a context are. It doesn't have that kind of categorization in its model or anywhere else.

Tokens don't even match word boundaries: they are intentionally misaligned to better accommodate contractions.

> In the process they learn a map that may be the equivalent of what we humans call "syntax", and they apply it to completely different contexts and inputs.

Exactly. Alongside those patterns, they model all of the other patterns that are present in the text. There is no categorization between these patterns: the word "syntax" has no bearing on the model whatsoever, except where it literally exists in the text being modeled.

LLMs do not parse. They tokenize, but not to match any predefined grammar.

It's not the LLM itself "being creative". It's the content of the text: the patterns that are modeled. Those patterns exist independent of the LLM. They come from the complex human behavior that is, "writing".

seydorOP3y ago

I dont actually see the point of using human-derived notions of 'language', 'syntax' , 'know', 'category' and even 'thinking' or 'parsing' when making precise arguments. Those are all vague, subjective notions that we made for ourselves but nature does not need to fit to our mold.

The fact is, LLMs are the best quantitative model of the human communication code, so good that it can replicate it quite convincingly. Better than any model we had made before using our imprecise subjective notions, which were quite like alchemy. Whether it will prove to be 'thinking' or not, remains to be seen in the future.

> the word "syntax" has no bearing on the model whatsoever,

We don;t know that. One would have to look into the model weights for that and this analysis does not exist yet. If one can provide a decent definition of syntax, it may be found to reside somewhere in its weights. same for other subjective notions.

thomastjeffery3y ago

> I dont actually see the point of using human-derived notions of...'thinking'...

> Whether it will prove to be 'thinking' or not, remains to be seen in the future.

Maybe you should take your own advice.

You are really missing my point. Humans use language, syntax, thinking, etc. We know what those patterns are. LLMs do not. LLMs don't approach text with the same models that humans use: they model the text directly. Any pattern, whether it be a word or sentence or even a pattern that humans have no name for at all: the LLM models that pattern.

What we humans call "syntax", "word", "idea", etc.; LLMs don't. They don't have any names for any things. They just have the collection of things in front of them, and implicitly model that collection.

An LLM doesn't think objectively about its model, either. It only constructs continuations by modeling prompts. It is the content of the prompt itself that determines the continuation, not "thought".

j / k navigate · click thread line to collapse