35 ms·

Isn’t it much more likely that there are a lot of book reviews and summaries in its training set from which it can synthesize its own?

VWWHFSfQ 3y ago
I think it's actually much more likely that they just dumped a bunch of book PDFs in the training folder and let it go to work. I seriously doubt any of these AI companies are being even the least bit careful about the data they're lapping up for training
- throwuwu 3y ago
  Training data quality is hugely important so I very much doubt they don’t curate the text they use.
sureglymop 3y ago
Where are the summaries from? I would say it's much more likely that a shadow library was scraped but if course that is also seemingly impossible to prove. One may be able to somewhat test that by asking for a summary of a book/text only available on a shadow library.
- pbhjpbhj 3y ago
  You could ingest all reviews that are extent in the online corpus, remove from the book all quotes found. Then ask the AI if a distinctive triples, say, of words appeared in their book, somehow, you'd probably need prompt engineering to get past "While I don't have access to the full text of the book [...]". A little maths and you might prove beyond reasonable doubt that the LLM was trained on the book. As a step towards a PoC I looked at https://www.amazon.co.uk/Bedwetter-Stories-Courage-Redemption-Pee-ebook/dp/B0045I7FYW https://www.amazon.co.uk/Bedwetter-Stories-Courage-Redemptio... and found a reference to "Boys' Market Manchester" which seemed like a Googlewhack-ish (unlikely) triple of words. Then I asked ChatGPT about it: Me: Has Sarah Silverman ever written about Boys' Market Manchester ChatGPT ChatGPT: As of my knowledge cutoff in September 2021, I do not have any information indicating that Sarah Silverman has written specifically about Boys' Market Manchester. Sarah Silverman is an American comedian, actress, and writer known for her stand-up comedy and her work in film and television. While she has written books and has often shared personal anecdotes in her comedy, I couldn't find any specific references to Boys' Market Manchester in relation to her work. However, please note that my information might not be up to date, as Sarah Silverman's career and activities may have evolved since then.
mint2 3y ago
Except they have a documented paper trail showing illegal book repos were used in training
- throwuwu 3y ago
  Link? Also have to prove they used all books and did 0 curation to remove copyrighted material
itronitron 3y ago
If book reviews and summaries were part of the training set, wouldn't that imply that OpenAI's LLM is more like a search engine in that it produces the input text based on a prompt?

VWWHFSfQ 3y ago

I think it's actually much more likely that they just dumped a bunch of book PDFs in the training folder and let it go to work. I seriously doubt any of these AI companies are being even the least bit careful about the data they're lapping up for training

throwuwu 3y ago
Training data quality is hugely important so I very much doubt they don’t curate the text they use.

sureglymop 3y ago

Where are the summaries from? I would say it's much more likely that a shadow library was scraped but if course that is also seemingly impossible to prove. One may be able to somewhat test that by asking for a summary of a book/text only available on a shadow library.

pbhjpbhj 3y ago
You could ingest all reviews that are extent in the online corpus, remove from the book all quotes found. Then ask the AI if a distinctive triples, say, of words appeared in their book, somehow, you'd probably need prompt engineering to get past "While I don't have access to the full text of the book [...]". A little maths and you might prove beyond reasonable doubt that the LLM was trained on the book. As a step towards a PoC I looked at https://www.amazon.co.uk/Bedwetter-Stories-Courage-Redemption-Pee-ebook/dp/B0045I7FYW https://www.amazon.co.uk/Bedwetter-Stories-Courage-Redemptio... and found a reference to "Boys' Market Manchester" which seemed like a Googlewhack-ish (unlikely) triple of words. Then I asked ChatGPT about it: Me: Has Sarah Silverman ever written about Boys' Market Manchester ChatGPT ChatGPT: As of my knowledge cutoff in September 2021, I do not have any information indicating that Sarah Silverman has written specifically about Boys' Market Manchester. Sarah Silverman is an American comedian, actress, and writer known for her stand-up comedy and her work in film and television. While she has written books and has often shared personal anecdotes in her comedy, I couldn't find any specific references to Boys' Market Manchester in relation to her work. However, please note that my information might not be up to date, as Sarah Silverman's career and activities may have evolved since then.

mint2 3y ago

Except they have a documented paper trail showing illegal book repos were used in training

throwuwu 3y ago
Link? Also have to prove they used all books and did 0 curation to remove copyrighted material

itronitron 3y ago

If book reviews and summaries were part of the training set, wouldn't that imply that OpenAI's LLM is more like a search engine in that it produces the input text based on a prompt?