3 ms·

This must be for show. If ByteDance or anyone else is sufficiently motivated to distill OpenAI’s models, it can’t be prevented. You can simply pay other people

This must be for show. If ByteDance or anyone else is sufficiently motivated to distill OpenAI’s models, it can’t be prevented. You can simply pay other people to collect the data for you.

blueboo 3y ago
“Simply”? It would be quite a project to do that at sufficient scale undetected, no?
- anon373839 3y ago
  How do you figure? It would look like people from all over the place running API workloads on a wide range of tasks and topics (if the goal is to distill the model generally, that’s what you’d need). So many people are already using GPT-4 to synthesize fine-tuning datasets, I think it would be invisible. However, you might wonder what the goal is. This “API distillation” is good for teaching a pretrained model how to do lots of things. But the end result is always constrained by the quality of the pretrained base model, and API outputs don’t help at all with that part.
yard2010 3y ago
With the backing state power they don't really need to pay anyone
hoofhearted 3y ago
You just block their IP’s; Gain intel from internal employees; Have the expensive law firm start sending C&D’s and prepping that law suit. If you tell someone that they can’t use your product in a certain manner, and then they go to extra lengths to circumvent your measures to gain a profit for themselves, then there is going to be big civil and criminal legal problems. State sponsored or not; Bytedance has a bank account and business license here in the U.S.
- renonce 3y ago
  They can outsource their API requests to third party “AI startup” who proxy their requests to OpenAI for a small fee
hoofhearted 3y ago
P.S. if I was in the position of OpenAI and Sam Altman, and Bytedance kept playing games after we had slapped their hand; I would just “shadow ban” them based off IP and user history, and serve them back garbage results to throw off their models :] Garbage in = Garbage Out

This must be for show. If ByteDance or anyone else is sufficiently motivated to distill OpenAI’s models, it can’t be prevented. You can simply pay other people to collect the data for you.

blueboo 3y ago

“Simply”? It would be quite a project to do that at sufficient scale undetected, no?

anon373839 3y ago
How do you figure? It would look like people from all over the place running API workloads on a wide range of tasks and topics (if the goal is to distill the model generally, that’s what you’d need). So many people are already using GPT-4 to synthesize fine-tuning datasets, I think it would be invisible. However, you might wonder what the goal is. This “API distillation” is good for teaching a pretrained model how to do lots of things. But the end result is always constrained by the quality of the pretrained base model, and API outputs don’t help at all with that part.

yard2010 3y ago

With the backing state power they don't really need to pay anyone

hoofhearted 3y ago

You just block their IP’s; Gain intel from internal employees; Have the expensive law firm start sending C&D’s and prepping that law suit. If you tell someone that they can’t use your product in a certain manner, and then they go to extra lengths to circumvent your measures to gain a profit for themselves, then there is going to be big civil and criminal legal problems. State sponsored or not; Bytedance has a bank account and business license here in the U.S.

renonce 3y ago
They can outsource their API requests to third party “AI startup” who proxy their requests to OpenAI for a small fee

hoofhearted 3y ago

P.S. if I was in the position of OpenAI and Sam Altman, and Bytedance kept playing games after we had slapped their hand; I would just “shadow ban” them based off IP and user history, and serve them back garbage results to throw off their models :] Garbage in = Garbage Out