Hi Kenny, I appreciate that you and others might want LLMs to be more capable than they are, esp. as they can feel like that at times. But in your comments above and below, you\u2019re pointing to a difference that doesn\u2019t make a difference:<\/p>\n
First, I said \u201cAt their core, LLMs cannot\u2026\u201d There\u2019s a lot of nuance that my brief, general description naturally doesn\u2019t capture. E.g., you\u2019re right that LLMs can be made more accurate by adding layers of processing as safeguards, e.g., RAG. <\/p>\n
Still, those layers and nuances don\u2019t magically transform LLMs into something else. At the end of the day, they still answer your queries by saying things that other people might say in response (and people can say all kinds of wacky things).<\/p>\n
In your example above, LLMs are just double-checking their initial output against what those human raters might say.<\/p>\n
In your example below (on math), they\u2019re also just double-checking their initial output against what human experts in math might say. Those LLMs still aren\u2019t calculators, because they can still get it wrong.<\/p>\n
I\u2019m not saying it\u2019s useless that LLMs answer your queries based on how other people might answer you. For some domains that are well known, e.g., \u201cWhat is the capital of France?\u201d, what other people might say in response would likely coincide with the fact of the matter. The reason why is that LLMs have been trained on tons of examples of \u201cParis is the capital of France.\u201d So, LLMs can indirectly model the world by modeling our language.<\/p>\n
But in less-known domains, they are much more likely to fabricate or falsely hallucinate, including giving a particular voice (or article or paper, etc.) disproportionate weight, just because there aren\u2019t many other competing voices. Here are a couple<\/a> examples <\/a>from this year, not older LLMs.<\/p>\n Anyway, you don\u2019t need to believe me, but just ask AI, if you think it\u2019s more capable than I do. Here\u2019s what Google\u2019s AI<\/a> thinks, and you can get similar results with other AI:<\/p>\n Query:<\/p>\n Is this essentially correct?:<\/p>\n \u201cAs large language models (LLMs), they model languages, given a large training dataset to draw\/remix from. This means they aren\u2019t designed to model the world, or domain knowledge, or whatever. They just model what other people might say in response to your prompt. Sometimes that can be useful, other times it s not; so, your mileage may vary.\u201d<\/p>\n AI:<\/p>\n Yes, your statement is\u00a0essentially correct\u00a0and aligns with the prevailing \u201cstochastic parrot\u201d or \u201cfancy autocomplete\u201d view of Large Language Models (LLMs).<\/p>\n It accurately highlights that LLMs are trained to predict the next word (token) in a sequence based on statistical patterns in massive datasets, not to \u201cunderstand\u201d the real world, possess domain knowledge, or reason logically in the human sense.<\/p>\n Here is a breakdown of your points based on current understanding\u2026<\/p>\n