3 ms·

Well an LLM certainly doesn't store the training data in a compressed format, because otherwise it'd be sci-fi levels of compression. So what are you suggesting

Well an LLM certainly doesn't store the training data in a compressed format, because otherwise it'd be sci-fi levels of compression. So what are you suggesting it is doing?

chaxor 3y ago
I wouldn't be so sure that there isn't "any" form of compression going on - that just seems silly. It seems that at some level it is *fantastic* lossy semantic compression. This is one of the more fascinating aspects of LLMs from a computer science perspective - almost all of human knowledge and experience in text from (several petabytes of data in a database) can be 'cleverly semantically compressed' to fit into a database of just a few hundred GB (NN params). That's some astounding lossy compression algorithm, unfortunately with a less than efficient decoder (temp=0). However, it isn't always 'plagiarism' either. 'Memorization' vs 'Plagiarism' vs 'understanding' is actually very tricky in LLMs, because having a good world model requires memorizing some facts, as well as some (potentially) "ontological" constructions. In addition, quoting a famous person is very useful for imparting some wisdom occasionally, which is also memorization, and can be seen as plagiarism if the author is left off. But sometimes this occurs (common sayings) and no one bats an eye. One good example of why explicit memorization is really needed is to provide exact references (list of authors, year, and article title) - so clearly making a NN not rewarded for memorizing is not ideal. Hence, *very occasionally*, you get some 'bad memorization'. But it is worth noting that the alignment achieved with these models for such an enormously complex task is quite good.
gumballindie 3y ago
An audio file doesnt store musical notes and a digital book does not store ink and paper. Using algorithms either of the two can however reproduce audio and visual content sufficiently similar, albeit not identical. An llm or generative model doesnt store audio files and digital books, compressed or not. But using algorithms it can reproduce audio and visual content sufficiently similar, but unlike my previous example it’s also made of bites. Just like an audio file and digital book are not copies of the physical product, an ai “data source” is not an identical or compressed copy of the original bites. That’s why it can be used to generate output identical to the original data even tho the data it uses is not a verbatim copy. It’s a new way of storing data, optimised for stochastic generation. It is a powerful tool but still relies on original content.
- gmerc 3y ago
  Ya but we have not outlawed tools that create audio files nor have we outlawed audio formats. So?
  - gumballindie 3y ago
    No one is asking for outlawing ai, that would be ridiculous. But as we regulate what we can do with audio files and audio creation tools we can also regulate what we can do with ai.

Well an LLM certainly doesn't store the training data in a compressed format, because otherwise it'd be sci-fi levels of compression. So what are you suggesting it is doing?

chaxor 3y ago

I wouldn't be so sure that there isn't "any" form of compression going on - that just seems silly. It seems that at some level it is *fantastic* lossy semantic compression. This is one of the more fascinating aspects of LLMs from a computer science perspective - almost all of human knowledge and experience in text from (several petabytes of data in a database) can be 'cleverly semantically compressed' to fit into a database of just a few hundred GB (NN params). That's some astounding lossy compression algorithm, unfortunately with a less than efficient decoder (temp=0). However, it isn't always 'plagiarism' either. 'Memorization' vs 'Plagiarism' vs 'understanding' is actually very tricky in LLMs, because having a good world model requires memorizing some facts, as well as some (potentially) "ontological" constructions. In addition, quoting a famous person is very useful for imparting some wisdom occasionally, which is also memorization, and can be seen as plagiarism if the author is left off. But sometimes this occurs (common sayings) and no one bats an eye. One good example of why explicit memorization is really needed is to provide exact references (list of authors, year, and article title) - so clearly making a NN not rewarded for memorizing is not ideal. Hence, *very occasionally*, you get some 'bad memorization'. But it is worth noting that the alignment achieved with these models for such an enormously complex task is quite good.

gumballindie 3y ago

An audio file doesnt store musical notes and a digital book does not store ink and paper. Using algorithms either of the two can however reproduce audio and visual content sufficiently similar, albeit not identical. An llm or generative model doesnt store audio files and digital books, compressed or not. But using algorithms it can reproduce audio and visual content sufficiently similar, but unlike my previous example it’s also made of bites. Just like an audio file and digital book are not copies of the physical product, an ai “data source” is not an identical or compressed copy of the original bites. That’s why it can be used to generate output identical to the original data even tho the data it uses is not a verbatim copy. It’s a new way of storing data, optimised for stochastic generation. It is a powerful tool but still relies on original content.

gmerc 3y ago
Ya but we have not outlawed tools that create audio files nor have we outlawed audio formats. So?
- gumballindie 3y ago
  No one is asking for outlawing ai, that would be ridiculous. But as we regulate what we can do with audio files and audio creation tools we can also regulate what we can do with ai.