3 ms·
This is why you can't actually generalize your models to work for all contexts. If you want to anthropormphize them, they're natural poets and will find all th
by swatcoder 3y ago
This is why you can't actually generalize your models to work for all contexts.
If you want to anthropormphize them, they're natural poets and will find all the forgotten and degraded metaphors in your language and throw them back at you when you're not expecting it. You can't answer general question, write code, and generate stories all in the same system because the formal rules of those contexts are different and this poet's really into ignoring formal rules. That's kind of its thing.
Or in more mechanical terms, they work in huge, fuzzy associative networks not formal concepts and don't make the strict distinctions we do. You could even look at that as a unique and positive characteristic to leverage, if you weren't trying to prove that they're subsitute human thinkers.
- behnamoh 3y ago> You could even look at that as a unique and positive characteristic to leverage, if you weren't trying to prove that they're subsitute human thinkers. This is the way I think about LLMs: Lovely Little Minds. But your comment points to the opposite. Can you expand the part that you talk about "fuzzy associative networks"? What sort of unique and positive characteristics can we leverage?