zlacker

Impossible. The only way to know what is happening is to have the code run on your own infra.

>>popalc+(OP)
That still doesn't mean much unless you're doing your own training or getting the weights from a trusted source, and neither of those mean much without knowing something about the data being trained on.

If someone is trying to influence your results, running the inference on your own infrastructure prevents some attack vectors but not some of the more plausible and worrying ones.

>>sfink+Hq
I don't think people are concerned about the models' math being biased/tainted (people know of it but that largely doesn't factor into the "security concerns" that people cite.) Typically, it's about how do we know that our data is not going to be seen by a 3rd party. That's what I'm speaking to. Running on your own infra, you can guarantee there are no phone-homes.