Grab llama.cpp. Self-host on your machine in a VM that has GPU permission but not network. Presto, no big-tech data collection. No fingerprinting. No profiling. No ads.

I've used it like that for language translation. Sometimes I wanna read a German, Japanese or w/e page. Or say a few words to someone in their own language. Which my human brain can't do. Local models do a more than acceptable job for me.

You have to be really, really super careful about confabulation, sycophancy & other probs. Do not get into the habit of asking a q to the model and believing it. And don't use it agentically. But for page translations? Other low-importance tasks? Useful.

changes the fingerprint each session going to help

It's gonna be super hard to totaly defeat fingerprinting. And you'll never know if you really have. Local model + deny network, you can be almost totally confident. "Almost", b/c nothing is 100% perfect, ever. But you go from almost 100% chance of surveilence, to almost 0% chance of surveilence.