Rings very true to my experience. The issue with LLMs is once you exit the domain the common path, you quickly realize you are back to solving problems. I like to think of LLMs more as a big, conversational search engine. Synthetic data and boundary pushing are still difficult if not impossible with current tech. That said this will eventually get solved but we are much further from it than we think.

Rings very true to my experience. The issue with LLMs is once you exit the domain the common path, you quickly realize you are back to solving problems. I like to think of LLMs more as a big, conversational search engine. Synthetic data and boundary pushing are still difficult if not impossible with current tech. That said this will eventually get solved but we are much further from it than we think.

https://x.com/VictorTaelin/status/1844969648904663126

Neat paper out of Apple measuring the performance of a bunch of LLMs on a mathematical reasoning benchmark as the questions are varied: changing names, changing numbers, adding relevant clauses, and adding irrelevant statements. o1 seems to be in a class of its own, though it still falters when no-op statements are added. https://arxiv.org/abs/2410.05229

Can you quote this bitch in here

https://warpcast.com/elffjs/0x20aa84bd