4 ms·

> the marginal value that the next buyer would buy Well the argument is that they should pay for the right to produce derivative works not for the physical cop

by qwytw 3mo ago

> the marginal value that the next buyer would buy Well the argument is that they should pay for the right to produce derivative works not for the physical copy.

TeMPOraL 3mo ago
Given the specific of the "derived works" being produced, do you feel obliged to pay royalties every time you cite or use some knowledge acquired from a book during your daily life?
- qwytw 3mo ago
  Well no, but I'm a human and not a commercial computer program so it's not particularly relevant. I'm would not be making loads of money making those citations either but that's only secondary. > derived works It's not evident that any works are being produced in the legal sense. Since LLM outputs are not considered copyrightable it's probably closer to using a search engine.
Kim_Bruning 3mo ago
Sure, but they're not producing derivative works. I'm told some of the recent Qwen models were trained on something like 140TB of data or so. I can put some of the Qwen models on a slightly large USB stick (~32G) . No way that contains the whole source text. No form of compression survives that. It's an old argument, it's still the simplest argument-by-numbers. The intuition this counters is "it's drawing from a big database" or "it memorized everything". It didn't. It still can recall passages at times. But then again, so can I for my favorite books and movies. And no I'm not saying it's human, to stop the next big jump. LLMs are LLMs. They have their own properties. Some of which happen to coincide with human properties. Just like a robot arm has some properties in common with humans. Similar requirements, similar constraints; you're going to have to expect similar behaviors and failure modes at least some of the time. But I haven't heard anyone claim a robot arm can edit a python program, nor have I heard of an LLM picking 100 widgets per minute.

by qwytw 3mo ago

> the marginal value that the next buyer would buy Well the argument is that they should pay for the right to produce derivative works not for the physical copy.

TeMPOraL 3mo ago

Given the specific of the "derived works" being produced, do you feel obliged to pay royalties every time you cite or use some knowledge acquired from a book during your daily life?

qwytw 3mo ago
Well no, but I'm a human and not a commercial computer program so it's not particularly relevant. I'm would not be making loads of money making those citations either but that's only secondary. > derived works It's not evident that any works are being produced in the legal sense. Since LLM outputs are not considered copyrightable it's probably closer to using a search engine.

Kim_Bruning 3mo ago

Sure, but they're not producing derivative works. I'm told some of the recent Qwen models were trained on something like 140TB of data or so. I can put some of the Qwen models on a slightly large USB stick (~32G) . No way that contains the whole source text. No form of compression survives that. It's an old argument, it's still the simplest argument-by-numbers. The intuition this counters is "it's drawing from a big database" or "it memorized everything". It didn't. It still can recall passages at times. But then again, so can I for my favorite books and movies. And no I'm not saying it's human, to stop the next big jump. LLMs are LLMs. They have their own properties. Some of which happen to coincide with human properties. Just like a robot arm has some properties in common with humans. Similar requirements, similar constraints; you're going to have to expect similar behaviors and failure modes at least some of the time. But I haven't heard anyone claim a robot arm can edit a python program, nor have I heard of an LLM picking 100 widgets per minute.