In your faq you say you have no plans to make nostalgebraist-autoresponder use gpt-3, why not? Will it be too expensive to finetune?
For one thing, I currently have no way to use GPT-3 at all: I signed up for OpenAI’s waitlist long ago and haven’t heard back.
(I’m not surprised or mad about this, I expect they’ve gotten a huge number of requests and I’m not especially high profile)
Even if I did have API access, it would be too expensive to host/serve, even without finetuning.
As for finetuning -- which, for my project, is a hard requirement -- I doubt I’d be allowed to do it at all. It’s only available at their “Scale” pricing tier, which looks aimed at corporate clients:
To sum up, using GPT-3 in production is not really comparable to using GPT-2 in production in terms of cost and accessibility.
GPT-2 is accessible to hobbyists. You can run GPT-2 on standard data center GPUs like T4s, which you can rent from Google for prices ranging from “low” to “zero” depending on your tolerance for timeouts. You can finetune it pretty quickly on a standard TPU (not even a pod).
GPT-3 was a deliberate exercise in pushing the limits of current hardware. The level of hardware and support needed to run it in production is not something average hobbyist can expect to access.