The dataset chatgpt's models were trained on is unbelievably massive. it' microsoft so think about ever single email conversaion from outlook, every public/private github repo, all public data (wikipedia, reddit, 4chan, twitter, facebook, etc. etc.) and all their other platforms, spanning decades.

We're talking trillions to quadrillions of individual conversations between billions of people over decades. Chatgpt is the appearance of intelligence. In reality all it is doing is regurgitating conversations it has already seen.

Yeah, these sorts of things are massively parallel and designed to run on specialized (GPU) hardware. I work for a law firm (AMLAW 100) and was researching doing our own AI based audio recognition, and even just for that it isn't feasible. It would mean spinning up multiple servers with specialized CUDA hardware and THEN training the AI constantly. Just not worth it when there are closed caption services in the cloud that will do the same thing and sign our NDAs.

Because that isn't how these things work. This isn't a piece of software that is compiled and sold to the public and someone can trace memory registries to unlock without a key. The the "AI" is on the backend, the user doesn't see or have access to that computer (server). Hacking isn't really a thing like you see on talmudvision. Things don't get "hacked" in the way. Nearly all hacking is intentional leaks, think CC info being "hacked", the remainder is nearly all social engineering. Most "hacking" where things are accessed that shouldn't be aren't hacks, they're people who didn't follow simple instructions, left default passwords on etc.