|
|
|
|
|
|
by vitamark
29 days ago
|
|
|
"For an LLM, it is exactly the opposite". Is it really? LLMs don't have words inside, for the most part they operate by applying transformations to a vector that does not contain any words at all. Words that come out of an LLM are just what sampler gives us by looking at the vector that is the result of those transformations. Does this vector contain a world model? Some form of thoughts or reasoning to arrive at the result? That's an open question really. Sure, rerunning that whole process for each token might not be the best solution, although that's an open question too. But saying that LLMs operate on words first is too big of an oversimplification |
|
|
It is, but it has been explored in various forms. Anthropic have some interesting papers where they can map out concepts to some extent as they exist inside the vectors of the LLM, and they can see e.g. that vectors relating to some concept can appear there some iterations before the LLM emits the tokens that spell it out. (though also, quite frankly, this shouldn't be all that surprising. If they were just picking the next token with no representation of what might follow it the result would be more like the 'one-word-at-a-time' party game than coherent text. They used poetry with a rhyming scheme as an example because it's something where you need to have some idea of what might rhyme with the line before and aim at it at the start of the next line in order to have any degree of success).