If you are familiar with the concept of an NP-complete problem, the weights are just one possible solution.
The Traveling Salesman Problem is probably the easiest analogy to make. It’s as though we’re all trying to find the shortest path through a bunch of points (ex. towns), and when someone says “here is a path that I think is pretty good”, that is analogous to sharing network weighs for an AI. We can then all openly test that solution against other solutions and determine which is “best”.
What they aren’t telling you is whether people traveling that path somehow benefits them (maybe they own all the gas stations on that path. Or maybe they’ve hired highway men to rob people on that path). And figuring out if that’s the case in a hyper-dimensional space is non-trivial.
thejevans@lemmy.ml 1 week ago
uh sure. My point is that sharing weights is analogous to sharing a compiled binary, not source code.
teawrecks@sopuli.xyz 1 week ago
Yes, and I don’t like the common comparison to binary blobs, and I’m attempting to explain why.
It is inherently safer to blindly run weights than it is to blindly execute a binary. The issues only arrise if you are then blindly trusting the outputs from the AI. But you should already have something in place to sanitize outputs and limit permissions, even for the most trustworthy weights.
It’s basically like hiring someone and wondering if they’re Hydra; no matter how deep your background check is, they could always decide to spontaneously defect and try to sabotage you. But that won’t matter if their decisions are always checked against enough other non-Hydra employees.