4 ms·

This, or a lawsuit like it is going to be the SCO vs IBM of the 2020's, to wit: a copyright troll trying to extract rent, with various special interests cheerin

by andy99 3y ago

This, or a lawsuit like it is going to be the SCO vs IBM of the 2020's, to wit: a copyright troll trying to extract rent, with various special interests cheering it on to try and promote their own agenda (ironically it was Microsoft that played that role with SCO). It's funny how times have changed and at least now a louder group seem to be on the troll's side. I hope to see some better analysis on the frivolity of this come out. There may be some commercial subtlety in specific cases that doesn't depend on scraping and training, but fundamentally using public internet data for training is not copying, is fair use, and is better for society as a whole than whatever ridiculous alternative might be proposed. edit: I'm speaking about training broadly capable foundation models like GPTn. It would of course be possible to build a model that only parrots copyrighted content and it would be hard to argue that is fair use.

dannyr 3y ago
NYTimes has a paywall. Is that public internet and therefore fair use?
- yreg 3y ago
  They don't have the paywall up if you identify as a search engine scraper, so it is kinda public internet. (I'm not claiming it's fair use.)
jamiek88 3y ago
Using the words troll and frivolous undermines your otherwise decent point and in fact goes against your point. The law isn’t settled, it’s a genuine legal question mark. It ain’t frivolous or trolling or ridiculous.
logicchains 3y ago
>It's funny how times have changed and at least now a louder group seem to be on the troll's side Because for many people, their views on current events are whatever the "thought leaders" working for the NYT and similar publications tell them to think.
phkahler 3y ago
>> There may be some commercial subtlety in specific cases that doesn't depend on scraping and training The key is to stop calling it "training" and use "learning" or just "reading". The argument from NYT will probably be that LLMs are just a fancy way to compress or abstract information and spit it back out. In which case "training" seems to support their case?
- mycall 3y ago
  I don't recall the source, but when people read, they typically only remember 20% of what they read (or heard?). Machine training encodes much more than 20%, so it is much closer to copying than training. Now the emergent abilities that come from this could be considered learning and dare I say imagination (which is the opposite of copying).
xbar 3y ago
That is an irrelevant comparison. This is theft and monstrous profit from theft. For actual justice this should be a class action suit of the world vs. OpenAI/Microsoft and the financial consequences should be company-ending for OpenAI. Otherwise, you have incented everyone in the AI industry to steal as much as they can for as long as they can.

by andy99 3y ago

dannyr 3y ago

NYTimes has a paywall. Is that public internet and therefore fair use?

yreg 3y ago
They don't have the paywall up if you identify as a search engine scraper, so it is kinda public internet. (I'm not claiming it's fair use.)

jamiek88 3y ago

Using the words troll and frivolous undermines your otherwise decent point and in fact goes against your point. The law isn’t settled, it’s a genuine legal question mark. It ain’t frivolous or trolling or ridiculous.

logicchains 3y ago

>It's funny how times have changed and at least now a louder group seem to be on the troll's side Because for many people, their views on current events are whatever the "thought leaders" working for the NYT and similar publications tell them to think.

phkahler 3y ago

>> There may be some commercial subtlety in specific cases that doesn't depend on scraping and training The key is to stop calling it "training" and use "learning" or just "reading". The argument from NYT will probably be that LLMs are just a fancy way to compress or abstract information and spit it back out. In which case "training" seems to support their case?

mycall 3y ago
I don't recall the source, but when people read, they typically only remember 20% of what they read (or heard?). Machine training encodes much more than 20%, so it is much closer to copying than training. Now the emergent abilities that come from this could be considered learning and dare I say imagination (which is the opposite of copying).

xbar 3y ago

That is an irrelevant comparison. This is theft and monstrous profit from theft. For actual justice this should be a class action suit of the world vs. OpenAI/Microsoft and the financial consequences should be company-ending for OpenAI. Otherwise, you have incented everyone in the AI industry to steal as much as they can for as long as they can.