Re:How LLMs Work, Explained Without Math – Miguel Grinberg
by TeMPOraL on Monday May 06, 2024 @07:17AM
(#40273328)
Point is that the simplest way to excel in next token prediction in the way human consider correct - which is rated by how people feel the predictor mimics a human understanding - is to actually have a world model and other components of human understanding.
Understanding and compression are the same thing. LLMs are fed a huge chunk of totality of human knowledge, and optimized to compress it well. They for sure aren't doing it by Huffman-encoding a multidimensional lookup table.
Re:How LLMs Work, Explained Without Math – Miguel Grinberg
by civilized on Monday May 06, 2024 @09:02AM
(#40274147)
> Point is that the simplest way to excel in next token prediction in the way human consider correct - which is rated by how people feel the predictor mimics a human understanding - is to actually have a world model and other components of human understanding.
This is a speculative theory for why a next token predictor might sound like it knows what it's talking about. Not something we actually know.
Re:How LLMs Work, Explained Without Math – Miguel Grinberg
by jameshart on Monday May 06, 2024 @07:20AM
(#40273345)
I mean, I think it was implied that humans judge the ‘next token prediction’ ability of LLMs as being good based on the quality of the overall output.
Re:How LLMs Work, Explained Without Math – Miguel Grinberg
by civilized on Monday May 06, 2024 @09:05AM
(#40274174)
In which case you have a trivial point rather than a backwards argument: "the output seems like it knows what's it talking about, and the easiest way to explain that is if it really knows what it's talking about."
This argument is backwards. Humans don't measure the next token prediction ability of the agents they speak to, human or AI. We rate speakers on whether they seem to understand what we say in context and respond by contributing useful information and analysis.
The attributes you're saying can be inferred from known superior next token prediction ability are the things we can actually detect and measure, at least qualitatively. Next token prediction quality is not measurable by humans in any human-meaningful way. Improving test cross entropy by 50% doesn't mean anything to us. It is irrelevant except as a mechanism to train LLMs.