Kind of sad that we have to rely on third parties with enough compute power to get a good experience with GPT. I obtained 'a' GPU and spent a week trying to learn how to fine-tune GPT with it, and the results were terrible. Maybe this isn't the kind of thing suitable for hacking on as a hobby, unless you're a researcher or employee with access to a lot of capital and dozens of GPUs.

Still wish the old talktotransformer model was released, instead of monetized behind a new company. I haven't been able to find a comparable model yet.