3 ms·

Reddit has a new AI training deal to sell user content

35 points by namanyayg 3y ago

ncr100 3y ago
This means Reddit is selling the intellectual contribution of its users, on its face. If the comments and posts are artwork, or music, would that also be legitimate usage by reddit according to their Terms of Service? "The user is the product." Appears to be true again, now there.
- pavel_lishin 3y ago
  > If the comments and posts are artwork, or music, would that also be legitimate usage by reddit according to their Terms of Service? I don't see why they wouldn't be. In any case, I'll be spending this week scrambling my old comments & probably deleting my submissions. Reddit's sufficiently user hostile at this point that I can find other places to entertain myself online.
  - kjkjadksj 3y ago
    All your comments are crawled aready, there’s no point in deleting a few miliseconds after posting.
- nullindividual 3y ago
  What about the works that their users have stolen and (re)posted without permission?
- lelandfe 3y ago
  When you post something on Reddit, you give Reddit a license to do more or less whatever it wants with the thing posted. https://www.redditinc.com/policies/user-agreement https://www.redditinc.com/policies/user-agreement > When Your Content is created with or submitted to the Services, you grant us a worldwide, royalty-free, perpetual, irrevocable, non-exclusive, transferable, and sublicensable license to use, copy, modify, adapt, prepare derivative works of, distribute, store, perform, and display Your Content
  - thebruce87m 3y ago
    > When Your Content is … submitted to the Services, you grant us a worldwide From the way this is worded it sounds like if someone else posts (submits) your content you are somehow granting this despite being unaware.
    - bick_nyers 3y ago
      IANAL but I think the "you grant us" part is what is key. If a license is not yours to grant, then that license is not theirs to receive.
      - Red_Leaves_Flyy 3y ago
        The models trained there on are rotten fruits and will be destroyed under court order. Frankly, a cartel with the power to change their countries laws would be wise to spend a few billion building out a state of the art ai lab and blatantly training their models on everything they can download. Everyone would use it. Would the us actually invade ($country) because hyper realistic Taylor swift and Mickey Mouse porn is a tap and swipe away?
    - lelandfe 3y ago
      "Your Content," not "your content." It is legalese jargon, defined as: "Content created with or submitted to the Services by you or through your Account" And: "By submitting Your Content to the Services, you represent and warrant that you have all rights, power, and authority necessary to grant the rights to Your Content" Of course, how many people posting memes on Reddit have the rights to do so? How many know they're in violation of the user agreement in doing so, hah.
K0balt 3y ago
Wow, didn’t see that coming…. (lol) Hopefully, the dataset will be pirated and released into the de-facto public domain, since it was always considered by the users to be a commons, not an intellectual slave farm.
- pavel_lishin 3y ago
  What's to pirate? Anyone can crawl reddit now, right?
  - K0balt 3y ago
    Nope, not in bulk. They have antiscraper limits in their api. But we could web scrape it with an AI agent looking for quality training data… It would be a TOS violation, so technically piracy I think still.
    - kjkjadksj 3y ago
      People probably just do it the old fashioned way with a botnet doing scraping under the rate limit.
      - K0balt 3y ago
        Show HN: I built a botnet to scrape Reddit and why I used OrbitDB and rust
neom 3y ago
Yesterday: https://news.ycombinator.com/item?id=39409499 https://news.ycombinator.com/item?id=39409499
artninja1988 3y ago
This just shows why scraping content to use for models has to be free for everyone or incumbents are gonna pull stuff like this. Open source has to be able to compete with big tech, who can train all their stuff on their users data, as per their tos and big rights holders.
zeknife 3y ago
I seriously doubt that adding reddit posts is going to increase the output quality of any LLM
- KTibow 3y ago
  It's been a big part of the training data since the start, as glitch tokens demonstrate
ChrisArchitect 3y ago
[dupe] Some more recent discussions: https://news.ycombinator.com/item?id=39409499 https://news.ycombinator.com/item?id=39409499 https://news.ycombinator.com/item?id=39404051 https://news.ycombinator.com/item?id=39404051

35 points by namanyayg 3y ago

This means Reddit is selling the intellectual contribution of its users, on its face. If the comments and posts are artwork, or music, would that also be legitimate usage by reddit according to their Terms of Service? "The user is the product." Appears to be true again, now there.

pavel_lishin 3y ago
> If the comments and posts are artwork, or music, would that also be legitimate usage by reddit according to their Terms of Service? I don't see why they wouldn't be. In any case, I'll be spending this week scrambling my old comments & probably deleting my submissions. Reddit's sufficiently user hostile at this point that I can find other places to entertain myself online.
- kjkjadksj 3y ago
  All your comments are crawled aready, there’s no point in deleting a few miliseconds after posting.
nullindividual 3y ago
What about the works that their users have stolen and (re)posted without permission?
lelandfe 3y ago
When you post something on Reddit, you give Reddit a license to do more or less whatever it wants with the thing posted. https://www.redditinc.com/policies/user-agreement https://www.redditinc.com/policies/user-agreement > When Your Content is created with or submitted to the Services, you grant us a worldwide, royalty-free, perpetual, irrevocable, non-exclusive, transferable, and sublicensable license to use, copy, modify, adapt, prepare derivative works of, distribute, store, perform, and display Your Content
- thebruce87m 3y ago
  > When Your Content is … submitted to the Services, you grant us a worldwide From the way this is worded it sounds like if someone else posts (submits) your content you are somehow granting this despite being unaware.
  - bick_nyers 3y ago
    IANAL but I think the "you grant us" part is what is key. If a license is not yours to grant, then that license is not theirs to receive.
    - Red_Leaves_Flyy 3y ago
      The models trained there on are rotten fruits and will be destroyed under court order. Frankly, a cartel with the power to change their countries laws would be wise to spend a few billion building out a state of the art ai lab and blatantly training their models on everything they can download. Everyone would use it. Would the us actually invade ($country) because hyper realistic Taylor swift and Mickey Mouse porn is a tap and swipe away?
  - lelandfe 3y ago
    "Your Content," not "your content." It is legalese jargon, defined as: "Content created with or submitted to the Services by you or through your Account" And: "By submitting Your Content to the Services, you represent and warrant that you have all rights, power, and authority necessary to grant the rights to Your Content" Of course, how many people posting memes on Reddit have the rights to do so? How many know they're in violation of the user agreement in doing so, hah.

K0balt 3y ago

Wow, didn’t see that coming…. (lol) Hopefully, the dataset will be pirated and released into the de-facto public domain, since it was always considered by the users to be a commons, not an intellectual slave farm.

pavel_lishin 3y ago
What's to pirate? Anyone can crawl reddit now, right?
- K0balt 3y ago
  Nope, not in bulk. They have antiscraper limits in their api. But we could web scrape it with an AI agent looking for quality training data… It would be a TOS violation, so technically piracy I think still.
  - kjkjadksj 3y ago
    People probably just do it the old fashioned way with a botnet doing scraping under the rate limit.
    - K0balt 3y ago
      Show HN: I built a botnet to scrape Reddit and why I used OrbitDB and rust

neom 3y ago

Yesterday: https://news.ycombinator.com/item?id=39409499 https://news.ycombinator.com/item?id=39409499

artninja1988 3y ago

This just shows why scraping content to use for models has to be free for everyone or incumbents are gonna pull stuff like this. Open source has to be able to compete with big tech, who can train all their stuff on their users data, as per their tos and big rights holders.

zeknife 3y ago

I seriously doubt that adding reddit posts is going to increase the output quality of any LLM

KTibow 3y ago
It's been a big part of the training data since the start, as glitch tokens demonstrate

ChrisArchitect 3y ago

[dupe] Some more recent discussions: https://news.ycombinator.com/item?id=39409499 https://news.ycombinator.com/item?id=39409499 https://news.ycombinator.com/item?id=39404051 https://news.ycombinator.com/item?id=39404051