> I definitely get this. The thing that gives me hope is that you only need to poison a very small % of content to damage AI models pretty significantly. It helps combat the mass scraping, because a significant chunk of the data they get will be useless, and its very difficult to filter it by hand
It'd be great if the code returned by this project is code that doesn't work. Imagine if all these models are being trained with code that looks OK but in the end it just bullshit. I'd be amazing.
Miasma is just a wrapper around the "Poison Fountain". You can check out the explanation and sample some of their content here: https://rnsaffn.com/poison3/
It's pretty much exactly what you're describing: content that looks correct but is deeply insane.
> I definitely get this. The thing that gives me hope is that you only need to poison a very small % of content to damage AI models pretty significantly. It helps combat the mass scraping, because a significant chunk of the data they get will be useless, and its very difficult to filter it by hand
It'd be great if the code returned by this project is code that doesn't work. Imagine if all these models are being trained with code that looks OK but in the end it just bullshit. I'd be amazing.