First of all…bro…please use punctuation. This completely period-less wall of text was a nightmare to read lmao.
An OS controlled completely by AI sounds nightmarish to be honest. The amount of wrong outputs your typical LLMs generate would steer your PC into oblivion i tell you.
ASSISTED by AI is another thing though. There are already some tools out there that let you turn on some AI overlay that can help you figure out what you want to do. Even for the terminal.
But wouldn’t want that on my PC tbh. First of all, it would make the electricity bill for using your PC pure hell… imagine you have to fire up your GPU to 100% usage everytime you want to do something…
Yeah…you’re basically talking gibberish, sorry to tell you.
GPT4 doesn’t have to be reverse engineered, their software is nothing special and the open source programs available are more than competitive.
The part that can’t be pirated is the model itself. The training data that was formed into the model. That you can’t pirate or reverse engineer. You could steal it by hacking them and leaking it. But it would be incredible inefficient to run because OpenAI basically didn’t give a single fuck about efficiency and you need really fcking high end GPUs to run that.
So, to conclude.
ChatGPT “software” is absolutely boring and the alternatives out there are both plenty and in many cases superior. The interesting part is the model, which can’t be reverseengineered and even if, can’t be run on most hardware.
What you want is unrealistic and to be honest, completely unnecessary. We don’t want or need ChatGPT’s model. The open source models that already exist are in many cases better than GPT4; But they’re focused on specific tasks. Yes - there is no single model that can compete with GPT4 as is. But you can get yourself models for focused tasks and those “combined” definetly CAN compete with GPT4.
And the open source softwares like oobabooga are so much more capable than GPT4’s stupid minimalistic webui…
Well you wont find any “pirated” GPU host… Its nothing that can really be pirated because we’re talking about hardware here.
Afaik google colab has enough VRAM to train and they don’t really at the source material you put in there (again, as long as you don’t trigger any CSAM filters or so)
Runpod.io for example, you can just rent a GPU per the hour, train your own model and you’re good to go. I mean as long as you don’t upload criminal stuff (CSAM, etc.) you can train what you want on there.
Otherwise, training your own models will become easier by the day for home hardware. Look at fast it went for Image AI to be able to train your own model. You can do that with 12GB or even 8GB afaik now. I don’t think it’ll take very long until you can just let your 8GB card chuck away at some training data.