Skip to main content
17 events
when toggle format what by license comment
Feb 14 at 16:23 comment added gaborous 1.5 years later, the exact same idea and theoretical underpinnings of the proof with the addition of some experimental demonstrations are now published by another team in a Nature paper.
Jun 30, 2024 at 21:16 comment added gaborous @WernfriedDomscheit Thank you very much, I did not know there was a name for this concept (despite working in AI!), thank you very much for pointing this out!
Jun 19, 2024 at 6:59 comment added Wernfried Domscheit This topic is commonly known as model collapse
Feb 7, 2023 at 22:48 comment added Rahul Its interesting, how researchers label Machine-generated data while implicitly fetching data for training. ?
Jan 28, 2023 at 13:14 comment added gaborous @Cerbrus AI not another party but a technology, just like the (AI-based) recommendation system of SO/SE. Unless we consider SO to be a closed-loop system, which I find implausible since it is a publicly accessible website, I think it is extremely short-sighted to dismiss any non-SO-centric perspective. The perspective I offer is just more systematic, accounting for other factors and the wider technological landscape. I make no claim it should be the primary perspective, I explicitly wrote it is an alternative one, a "food for thought" if you prefer.
Jan 27, 2023 at 8:28 comment added Cerbrus @gaborous The point of view of the other party is, frankly, irrelevant when discussing SO policy.
Jan 26, 2023 at 21:18 comment added Peter Mortensen Re "it's very easy to workaround these filters with slightly modified models or post-processing": Yes, but fortunately the worst abusers also have the minimum effort attitude (in other words, they couldn't be bothered).
Jan 26, 2023 at 20:47 comment added user400654 i simply don't see how this addresses the policy at all. You aren't contesting the policy, you state you agree with it... but that's what the upvote button is for. Nothing else in the answer is relevant to the policy decision.
Jan 26, 2023 at 20:42 comment added gaborous @Cerbrus There is no filter that is 100% effective, and it's very easy to workaround these filters with slightly modified models or post-processing, there are already loads of posts like that on social networks by AI users (not even researchers). And I never wrote that it's SO/SE responsibility to provide usable training data, I only provided an alternative viewpoint that also converges towards the same solution. It seems there is a radicalization of "us vs them" mindset, sad to see. Even though my opinion agrees with the ban, it's ditched just for taking the point of view of another party.
Jan 26, 2023 at 20:36 comment added gaborous @JAD Your opinion sounds very manichean. When not used for spamming, AI has wonderful applications that already changed our current world. In any case, hating a tool is fruitless, just like hating a knife. It can be used for good and bad. I only argue that even from the point of view of AI researchers and users, there is no medium to long-term benefit to accepting AI generated answers, only (catastrophically bad) downsides. So whether you look from the POV of SE users, or from AI users/researchers, banning is one of the few reasonable solutions.
Jan 26, 2023 at 20:34 history edited gaborous CC BY-SA 4.0
added 1 character in body
Jan 26, 2023 at 20:18 history edited Peter Mortensen CC BY-SA 4.0
Added some context.
Jan 26, 2023 at 12:24 comment added Cerbrus The training data could be filtered to remove AI-generated content (As that content is detectable). Doesn't mean SO should allow AI-generated content, but it's not really SO's responsibility to provide usable training data.
Jan 26, 2023 at 9:29 comment added JAD Honestly, this sounds like entirely the problem of the researchers. I'm not sure why sites being spammed by these AIs would be sympathetic to the further creation of more of these models.
Jan 26, 2023 at 7:25 history edited Wai Ha Lee CC BY-SA 4.0
added 4 characters in body
Jan 26, 2023 at 6:43 history edited gaborous CC BY-SA 4.0
added 43 characters in body
Jan 26, 2023 at 6:37 history answered gaborous CC BY-SA 4.0