Twenkid, Wow. Lots of words. I don't mind detail, but words are slippery.
If you actually want to do stuff, it's better to keep the words to a minimum and start with concrete examples. At least until some minimum of consensus is agreed. Trying to focus on your concrete questions... On Sat, Jun 22, 2024 at 10:16 AM twenkid <twen...@gmail.com> wrote: > > Regarding the contradictions - I guess you mean ambiguity Yeah, I guess so. I was thinking more abstractly in terms of grammatical classes. You can never learn grammar, because any rule you make always ends up being violated: AB->C, **except** in certain cases... etc. But probably ambiguity at the meaning level resolves to the same kinds of issues, sure. > * BTW, what is "not to contradict"? How it would look like in a particular > case, example? Oh, I suppose any formal language is an example of a system that doesn't contradict. Programming languages... Maths, once axioms are fixed, would be another example of non-contradiction (by definition? Of course the thing with maths is that different sets of possible axioms contradict, and that contradiction of possible axiomatizations is the whole deal.) > What do you mean by "the language problem"? "Grammar", text compression, text prediction... > The language models lead to such an advance: compared to what else, other > (non-language?) models. Advance, compared to everything before transformers. In terms of company market cap, if you want to quibble. > Rob: >Do you have any comments on that idea, that patterns of meaning which > can be learned contradict, and so have to be generated in real time? > > I am not sure about the proper interpretation of your use of "to contradict"; > words/texts have multiple meanings and language and text are lower resolution > than thought if they are supposed to represent the reality "exactly" lower > level, higher precision representations are needed as well In terms of my analogy to maths, this reads to me like saying: the fact there are multiple axiomatizations for maths, means maths axioms are somehow "lower resolution", and the solution for maths is to have "higher precision representations" for maths... :-b If you can appreciate how nonsensical that analysis would be within the context of maths, then you may get a read on what it sounds like to me from the way I'm looking at language. Instead, I think different grammaticalizations of language are like different axiomatizations of maths (inherently random and infinite?) You're not the only one doing that, of course. Just the other day LeCun was tweeting something comparable in response to a study which revealed transformer world models seem to... contradict! Contradict?! Who'd 've thought it?! The more resolution you get, the less coverage you get. Wow. Surprise. Gee, that must mean that we need to find a "higher precision representation" somewhere else! LeCun's post here: https://x.com/ylecun/status/1803677519314407752 ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T682a307a763c1ced-M268d0affcf74d745427a406e Delivery options: https://agi.topicbox.com/groups/agi/subscription