4 ms·

This reminds me of that awesome analysis of all the speech of a diplomat's visit in Asimov's original Foundation book: > Hardin threw himself back in the chair

This reminds me of that awesome analysis of all the speech of a diplomat's visit in Asimov's original Foundation book: > Hardin threw himself back in the chair. “You know, that's the most interesting part of the whole business. I admit that I thought his Lordship a most consummate donkey when I first met him – but it turned out that he is an accomplished diplomat and a most clever man. I took the liberty of recording all his statements.” >... When Houk, after two days of steady work, succeeded in eliminating meaningless statements, vague gibberish, useless qualifications—in short all the goo and dribble—he found he had nothing left. Everything canceled out. Lord Dorwin, gentlemen, in five days of discussion didn't say one damned thing, and said it so that you never noticed. I'm pretty sure that we're now at the level of AI where it's possibly to fully automate such an analysis, such that even if the original content is entirely corrupted by product placement, the AI could cut it out to leave only the valuable information, if any remains. The only question is whether the AI will be on the user's side or the advertiser's side.

bambax 2y ago
What will happen is that when you ask the AI to summarize a book to remove the fluff, it will inject random mentions of how the main character decided to drink a Coke. Let's go with truly open models! you say. That way we can be sure there are no shoddy behind-the-scenes deal going on between the model provider and some company or government. But the ads are in the training data, they are part of the fabric of the world. You can't get rid of them except if you do the training yourself, which is a huge amount of work, and maybe impossible (because model providers escape copyright laws, and you can't).
- falcor84 2y ago
  I agree about the difficulty, but am optimistic that if enough of us wanted to achieve this, we could train it as a distributed effort, with the coordination of a non-profit like the EFF. The question is whether we care enough.
- tensor 2y ago
  Actually, training it yourself is probably better in terms of copyright. If you train an LLM that only you use, from books you own, I don't see how that is actually illegal. As long as you're not selling the LLM or content it generates is there any legal issue?
  - Terr_ 2y ago
    > If you train an LLM that only you use, from books you own, I don't see how that is actually illegal. But it won't be functional, because Large Language Models requires far more samples than your private library, at least if you want it to do things that you couldn't solve with a much simpler deterministic search. For example, usefully extending the document "Which book had the sarcastic one-armed antagonist?" with another sentence that happens to also be the answer.

bambax 2y ago

What will happen is that when you ask the AI to summarize a book to remove the fluff, it will inject random mentions of how the main character decided to drink a Coke. Let's go with truly open models! you say. That way we can be sure there are no shoddy behind-the-scenes deal going on between the model provider and some company or government. But the ads are in the training data, they are part of the fabric of the world. You can't get rid of them except if you do the training yourself, which is a huge amount of work, and maybe impossible (because model providers escape copyright laws, and you can't).

falcor84 2y ago
I agree about the difficulty, but am optimistic that if enough of us wanted to achieve this, we could train it as a distributed effort, with the coordination of a non-profit like the EFF. The question is whether we care enough.
tensor 2y ago
Actually, training it yourself is probably better in terms of copyright. If you train an LLM that only you use, from books you own, I don't see how that is actually illegal. As long as you're not selling the LLM or content it generates is there any legal issue?
- Terr_ 2y ago
  > If you train an LLM that only you use, from books you own, I don't see how that is actually illegal. But it won't be functional, because Large Language Models requires far more samples than your private library, at least if you want it to do things that you couldn't solve with a much simpler deterministic search. For example, usefully extending the document "Which book had the sarcastic one-armed antagonist?" with another sentence that happens to also be the answer.