Bot traffic has surpassed humans on the Internet. The winners in the next infrastructure cycle are the companies building trust rails for machines: agent identity, intent verification, API-native content delivery.
Isn’t AI training on itself a well known thing to avoid ? If I remember correctly the “” performance “” goes to shit very quickly when you train a model on it’s own output.
I doubt serious AI actors will make that mistake.
But on its own, the way they just opened their torrent client and started downloading made me furious.
In France you can still get caught downloading illegally and it can have serious consequences. But for AI businesses, copyright holders seem to look the other way. Businesses have extra rights to citizens and it’s completely unfair.
Yeah, I think we’re talking about the same thing. I thought the article I linked was the one I had read about model inbreeding, but now that I look at it a bit closer, it’s probably the product of model inbreeding itself. ;) I thought there was an article published this year about the problem, but now I can’t find it to save myself. It’s possible that I’m hallucinating. My memory is worse than ChatGPT’s.
Isn’t AI training on itself a well known thing to avoid ? If I remember correctly the “” performance “” goes to shit very quickly when you train a model on it’s own output.
I doubt serious AI actors will make that mistake.
But on its own, the way they just opened their torrent client and started downloading made me furious.
In France you can still get caught downloading illegally and it can have serious consequences. But for AI businesses, copyright holders seem to look the other way. Businesses have extra rights to citizens and it’s completely unfair.
AI developers might be smart enough to filter their own output, but AI training is still taking in output from their competitors’ models, usually without realizing it.
I read carefully this article but I think it is not about the issue I was mentioning.
I was talking about “model collapse” and this seems rather about multiple models training on similar datasets (shared learning ressources).
Yeah, I think we’re talking about the same thing. I thought the article I linked was the one I had read about model inbreeding, but now that I look at it a bit closer, it’s probably the product of model inbreeding itself. ;) I thought there was an article published this year about the problem, but now I can’t find it to save myself. It’s possible that I’m hallucinating. My memory is worse than ChatGPT’s.