There are many purposes and uses for LLMs, but probably the most important one in my view is to retrieve and use information from a broader array of copyrighted sources than just web pages. Many web pages infringe on copyright, but in a somewhat occluded way that is challenging for rights holders to verify on its face. With LLMs, the training set can include a lot of copyrighted material (original works of authorship fixed in a tangible medium of expression within the statutory time period of protection), and then produce a nonoriginal work of non-authorship through a pastiche of its training data.
Using LLMs to clone authors doesn't really make sense because of how they work. It asks the system to do something that it's not really set up to do. What it is good at is pulling information from many sources (both restricted and non-restricted) and then producing non-infringing derivative works very quickly and at reasonable cost.
"I tell authors to imagine that years from now the only readers of your book will be your grandchildren. Write the book that you would want them to read. Don’t make concessions to agents, editors, or marketing departments that would detract from the book that your want to give to your grandchildren." THIS.
Great idea - love the book club concept. I am 63 and my dad is 86. I am working on the project you outlined with me, my dad and a reasonably large IT team. I plan to start with Eleven Labs for the voice clones. I think Tyler’s podcast with John Stuart Mills proves the viability of the base case and it will further improve with time (recall the decades long human genome decoding). Implement and iterate. Onward! What a great time to be alive (or cloned).
I don't trust LLMs. It would not surprise me if training one on a few book club sessions meant that its own sessions regurgitated past session content rather than the style. Maybe they will get better some day, maybe they are good enough now and I only see the silly ones, but I have seen too many silly LLM examples to trust them.
The risk is training the LLM on previous book discussions and expecting it to regurgitate only the style, not the content, for future discussions about different books.
I worked with neural nets about 20 years ago for about six months, trying to predict credit card fraud in simple cases. One of the accepted practices was that you trained it on some random half of a huge dataset, then tested it on the other half, and did that over and over until you were satisfied that it was ready to use in the real world on fresh data. I don't know how LLMs differ from those neural nets, but I don't think a dozen (say) past discussions is enough data for training and testing, especially to extend it to possibly hundreds of future book discussions whose participants change over time.
But I'm no expert on either neural nets or LLMs. Things change a lot in 20 years and my six month project on simple credit card fraud may not have much bearing on book discussionsn in human languages.
These are such odd notions to have at this particular juncture. Not to make light of “book clubs” (though I don’t understand what this phenomenon is) or that they should be led by AK - he has a natural gift for concision, detachment, and clarity (of his own thoughts anyway) - y’all know that nobody reads anymore, right?
Just because the sort of people who read a blog such as this, view the phone as “perpetually something to read in one’s pocket” - does not mean this is how 95% of people regard it.
And books - forget it. We could go back to having books published by monks copying them out by hand and the American public would not notice.
Books are a hard one for me unless we literally mean book as a physical format. Generationally the US political class is old, beyond old on average, and they still read books, i.e. If you goal is to possibly influence a US decision maker via some sort of intellectual argument, books are probably still where it's at for another decade or two.
Books as a concept though, i.e. a long manuscript with a formalized structure including a start and end isn't going anywhere at least in the fiction world, I could agree it's dead in non-fiction. The youth still eat up "lore" books from Warhammer 40,000 novels to manga to Web books (I forget the term for them). My kids and others like her to my surprise voraciously consume Korean, China, and Japanese full length web novels (which if printed would run hundreds of pages like a "real" book) as a sort of, as far as I can tell, long form text only pulp manga similar to the aforementioned 40K novels for American men; likewise romance novels seem as popular as ever even among young women.
Is there a future market outside pulp, I don't really know but people forget things like Dorian Grey (the novelet, not novel), Chaucer, etc were all considered pulp in their day.
Perhaps the estates of those who had interesting minds should consider selling the deceased's work for training purposes. It might be a useful research tool to be able to query "Please list all of Oriana Fallaci's written references to cicadas" and get something reliably complete. However, so long as an interesting mind is still alive and functioning it seems as if the time spent fooling with AI is a bad trade off for whatever original thinking it may have displaced.
There are many purposes and uses for LLMs, but probably the most important one in my view is to retrieve and use information from a broader array of copyrighted sources than just web pages. Many web pages infringe on copyright, but in a somewhat occluded way that is challenging for rights holders to verify on its face. With LLMs, the training set can include a lot of copyrighted material (original works of authorship fixed in a tangible medium of expression within the statutory time period of protection), and then produce a nonoriginal work of non-authorship through a pastiche of its training data.
Using LLMs to clone authors doesn't really make sense because of how they work. It asks the system to do something that it's not really set up to do. What it is good at is pulling information from many sources (both restricted and non-restricted) and then producing non-infringing derivative works very quickly and at reasonable cost.
"I tell authors to imagine that years from now the only readers of your book will be your grandchildren. Write the book that you would want them to read. Don’t make concessions to agents, editors, or marketing departments that would detract from the book that your want to give to your grandchildren." THIS.
Great idea - love the book club concept. I am 63 and my dad is 86. I am working on the project you outlined with me, my dad and a reasonably large IT team. I plan to start with Eleven Labs for the voice clones. I think Tyler’s podcast with John Stuart Mills proves the viability of the base case and it will further improve with time (recall the decades long human genome decoding). Implement and iterate. Onward! What a great time to be alive (or cloned).
I don't trust LLMs. It would not surprise me if training one on a few book club sessions meant that its own sessions regurgitated past session content rather than the style. Maybe they will get better some day, maybe they are good enough now and I only see the silly ones, but I have seen too many silly LLM examples to trust them.
Makes sense not to trust them unchecked but I don't see significant consequences as a central figure for a book club.
The risk is training the LLM on previous book discussions and expecting it to regurgitate only the style, not the content, for future discussions about different books.
I worked with neural nets about 20 years ago for about six months, trying to predict credit card fraud in simple cases. One of the accepted practices was that you trained it on some random half of a huge dataset, then tested it on the other half, and did that over and over until you were satisfied that it was ready to use in the real world on fresh data. I don't know how LLMs differ from those neural nets, but I don't think a dozen (say) past discussions is enough data for training and testing, especially to extend it to possibly hundreds of future book discussions whose participants change over time.
But I'm no expert on either neural nets or LLMs. Things change a lot in 20 years and my six month project on simple credit card fraud may not have much bearing on book discussionsn in human languages.
We're both 70 & grew up in Saint Louis
I love books-I read a book a week.
Real books, by real authors.
“Write the book that you would want them (your grandchildren) to read.” - Great advice that can be expanded to how to live your life.
These are such odd notions to have at this particular juncture. Not to make light of “book clubs” (though I don’t understand what this phenomenon is) or that they should be led by AK - he has a natural gift for concision, detachment, and clarity (of his own thoughts anyway) - y’all know that nobody reads anymore, right?
Just because the sort of people who read a blog such as this, view the phone as “perpetually something to read in one’s pocket” - does not mean this is how 95% of people regard it.
And books - forget it. We could go back to having books published by monks copying them out by hand and the American public would not notice.
You should have stopped when you said you don't understand the book club phenomenon.
And the end of books is high exaggerated. At minimum it has in no way happened yet.
The end of book stores is something completely different
Books are a hard one for me unless we literally mean book as a physical format. Generationally the US political class is old, beyond old on average, and they still read books, i.e. If you goal is to possibly influence a US decision maker via some sort of intellectual argument, books are probably still where it's at for another decade or two.
Books as a concept though, i.e. a long manuscript with a formalized structure including a start and end isn't going anywhere at least in the fiction world, I could agree it's dead in non-fiction. The youth still eat up "lore" books from Warhammer 40,000 novels to manga to Web books (I forget the term for them). My kids and others like her to my surprise voraciously consume Korean, China, and Japanese full length web novels (which if printed would run hundreds of pages like a "real" book) as a sort of, as far as I can tell, long form text only pulp manga similar to the aforementioned 40K novels for American men; likewise romance novels seem as popular as ever even among young women.
Is there a future market outside pulp, I don't really know but people forget things like Dorian Grey (the novelet, not novel), Chaucer, etc were all considered pulp in their day.
Perhaps the estates of those who had interesting minds should consider selling the deceased's work for training purposes. It might be a useful research tool to be able to query "Please list all of Oriana Fallaci's written references to cicadas" and get something reliably complete. However, so long as an interesting mind is still alive and functioning it seems as if the time spent fooling with AI is a bad trade off for whatever original thinking it may have displaced.