This bit of the devs' statement strikes me as really odd:

"Anyone who claims that you can 'train' your own model from scratch solely on your own data is either misinformed, is conflating the concepts of 'training' and 'finetuning', or has enormous amounts of cash and time."

I've literally done this myself in Python. It's a Markov chain text generator trained solely on my own text. It won't take prompts like ChatGPT, but it's absolutely not finetuning an existing model either.