3 ms·

I don't understand why some peoole thinks any infringing content can be singled out and removed. Aren't LLMs giant coefficient matrices like, a punched out cro

I don't understand why some peoole thinks any infringing content can be singled out and removed. Aren't LLMs giant coefficient matrices like, a punched out croissant dough, made of all training data plyed over? How can you say you can remove one specific ply out of dough and declare that ever potential effect that the offending ply had created is now completely removed?

keiferski 3y ago
I’m no LLM expert but I think there is a distinction to be made between the LLM dataset and the output it gives to the user. What you’re suggesting is that it’s difficult to remove something from the dataset, which may be the case. But that doesn’t mean the user will necessarily be able to access it. My guess is that this is much easier to attack from the user end.
- numpad0 3y ago
  Removing something from the dataset requires full training from scratch(~$100 million for base unaligned GPT-4). You can't like, edit the database file and keep the AI. the database file _is_ the AI.
selimthegrim 3y ago
I think given the use of ply for AlphaGo/chess engines this is a pretty cool metaphor.
anonymousab 3y ago
The "reasonable" singular removal is more about coming up with ways to block prompts that can produce infringing content, and having filters on the other end to catch infringements before they are published to the user. It's an endless whack-a-mole that never actually addresses the problem but might look good enough to the legal system or to keen supporters. Barring some major breakthrough, the actual answer is to train a new model without the infringing data. I think some of the people saying "remove it from your model" are aware of this and are simply being glib and needling; "you've created this infringement monstrosity, so surely you made sure to include a way to deal with this problem without throwing away all of your work, right?"

keiferski 3y ago

I’m no LLM expert but I think there is a distinction to be made between the LLM dataset and the output it gives to the user. What you’re suggesting is that it’s difficult to remove something from the dataset, which may be the case. But that doesn’t mean the user will necessarily be able to access it. My guess is that this is much easier to attack from the user end.

numpad0 3y ago
Removing something from the dataset requires full training from scratch(~$100 million for base unaligned GPT-4). You can't like, edit the database file and keep the AI. the database file _is_ the AI.

selimthegrim 3y ago

I think given the use of ply for AlphaGo/chess engines this is a pretty cool metaphor.

anonymousab 3y ago

The "reasonable" singular removal is more about coming up with ways to block prompts that can produce infringing content, and having filters on the other end to catch infringements before they are published to the user. It's an endless whack-a-mole that never actually addresses the problem but might look good enough to the legal system or to keen supporters. Barring some major breakthrough, the actual answer is to train a new model without the infringing data. I think some of the people saying "remove it from your model" are aware of this and are simply being glib and needling; "you've created this infringement monstrosity, so surely you made sure to include a way to deal with this problem without throwing away all of your work, right?"