OpenAI charges per-word. You get an initial credit, but the "setup" prompt is so long that each request will be far more expensive than "normal" requests, since you'll be charged for the entire setup prompt with every single request.
That said, these new developments give us a new reason to have powerful computers. The last 10 years computers were way more powerful for most of the task for their conventional use cases but now computers are yet again way less powerful for the new frontier use cases.
Apparently GPT-3 runs on 350GB VRAM, so it requires specialised hardware, so the best we can do today is to sneak peak into the future of on-device GPT-3 like LLMs.
Not sure how often do people ask Siri questions but that seems like it would not be a significant cost for one user. If you were to build a service on top of it, that might be expensive though.
But I wonder how soon we’ll have fine-tuned or otherwise customized specialized LLMs that are cheaper to run? In the mean time this kind of thing is a really neat proof-of-concept of how ChatGPT’s abilities can be harnessed.
More information on the models: https://subscription.packtpub.com/book/data/9781800563193/2/...