Super nice piece. As someone a little more familiar with information theory, I tend to think of the LLM content generators as "decoders". Where the input prompt is a series of oriented codewords that represent encoded(compressed) bits of prose, which after initial reconstruction is run through a higher level error correction system to match a likely output.
As far as the capability of a LLM, there is a weakly parallel problem in computer science, where there are more subroutines than there are names to call them. Plus we want natural language code words, which limits the code book, limiting the level of compression, and thus the overall capacity of a LLM.
I also liked your larger point that most likely these LLM will just add to the noise floor.
Thank you for taking the time to write this thoughtful take on LLM content generators.
Ah, the "decoder" metaphor is interesting and one that hadn't occurred to me — fits in with the common joke about how humans will use ChatGPT to transform a few bullet points into a multi-paragraph email, and then use it again to transform the email back into bullet points for the recipient.
Another thing I've thought about the eventual capabilities of LLMs is that the training involves narrowing the probability space of their possible outputs (I remember seeing early versions that generated gibberish in the form of a recipe, or a letter, then "narrowing down" until there are actual words, then further until the words look like cooking ingredients, etc). There's a sweet spot where the space is narrow enough to get outputs that "sound right"/are correct but not so narrow as to appear deterministic and automated — which is around where we are now, I think. More data/training seems like it would result in more narrowing, and less of the dazzling "original" effects that GPT sometimes produces. My sense is that diminishing returns are soon to set in if they haven't already. I'm not sure if my mental model for this is entirely correct, but it makes sense to me that it's good at producing code, since the probability space for a "paragraph" of correct code is so much smaller than a paragraph of convincing natural language of the same length.
I'm glad you liked the piece — thanks for the comment.
Two AI specialists were on Sam Harris podcast the other week and as far as I can gather (lots went over my head), they both thought GI was extremely dumb and incapable; one of them thought it can harm us because it's trash, other that we have nothing to fear because it's trash. https://www.samharris.org/podcasts/making-sense-episodes/312-the-trouble-with-ai This completely contradicts everything that we're reading in the lay commentariat sphere or even specialist media commentariat, which sees great changes afoot due to AI like Chat and ChaiGPT.
I've noticed the first job ad the other day that has the requirement "know how to use ChatGPT and relies on it a lot". Some assistantship position for Misha Gloubermann.
A fun exercise is asking it to supply quotations from novels illustrating a concept or theme (especially ones that haven't been written about much and don't have easily googleable quotations).
I'd be a bit nervous about anyone who said they "relied" on ChatGPT. I've used it to code SQL a handful of times and it helped, but it also came with a bunch of little bugs.
Super nice piece. As someone a little more familiar with information theory, I tend to think of the LLM content generators as "decoders". Where the input prompt is a series of oriented codewords that represent encoded(compressed) bits of prose, which after initial reconstruction is run through a higher level error correction system to match a likely output.
As far as the capability of a LLM, there is a weakly parallel problem in computer science, where there are more subroutines than there are names to call them. Plus we want natural language code words, which limits the code book, limiting the level of compression, and thus the overall capacity of a LLM.
I also liked your larger point that most likely these LLM will just add to the noise floor.
Thank you for taking the time to write this thoughtful take on LLM content generators.
Ah, the "decoder" metaphor is interesting and one that hadn't occurred to me — fits in with the common joke about how humans will use ChatGPT to transform a few bullet points into a multi-paragraph email, and then use it again to transform the email back into bullet points for the recipient.
Another thing I've thought about the eventual capabilities of LLMs is that the training involves narrowing the probability space of their possible outputs (I remember seeing early versions that generated gibberish in the form of a recipe, or a letter, then "narrowing down" until there are actual words, then further until the words look like cooking ingredients, etc). There's a sweet spot where the space is narrow enough to get outputs that "sound right"/are correct but not so narrow as to appear deterministic and automated — which is around where we are now, I think. More data/training seems like it would result in more narrowing, and less of the dazzling "original" effects that GPT sometimes produces. My sense is that diminishing returns are soon to set in if they haven't already. I'm not sure if my mental model for this is entirely correct, but it makes sense to me that it's good at producing code, since the probability space for a "paragraph" of correct code is so much smaller than a paragraph of convincing natural language of the same length.
I'm glad you liked the piece — thanks for the comment.
Two AI specialists were on Sam Harris podcast the other week and as far as I can gather (lots went over my head), they both thought GI was extremely dumb and incapable; one of them thought it can harm us because it's trash, other that we have nothing to fear because it's trash. https://www.samharris.org/podcasts/making-sense-episodes/312-the-trouble-with-ai This completely contradicts everything that we're reading in the lay commentariat sphere or even specialist media commentariat, which sees great changes afoot due to AI like Chat and ChaiGPT.
I've noticed the first job ad the other day that has the requirement "know how to use ChatGPT and relies on it a lot". Some assistantship position for Misha Gloubermann.
A fun exercise is asking it to supply quotations from novels illustrating a concept or theme (especially ones that haven't been written about much and don't have easily googleable quotations).
I'd be a bit nervous about anyone who said they "relied" on ChatGPT. I've used it to code SQL a handful of times and it helped, but it also came with a bunch of little bugs.