[minimaxir]

Since the demos on this page use zero-shot learning and the used model has a 2020-05-03 timestamp, that implies this API is using some form of GPT-3: https://news.ycombinator.com/item?id=23345379 (EDIT: the accompanying blog post confirms that: https://openai.com/blog/openai-api/ )

Recently, OpenAI set the GPT-3 GitHub repo to read-only: https://github.com/openai/gpt-3

Taken together, this seems to imply that GPT-3 was more intended for a SaaS such as this, and it's less likely that it will be open-sourced like GPT-2 was.

[wildermuthn]

But since the resources required for training such a model are only available to well-funded entities, it seems like offering the model as an API while releasing the original source-code is the best practical method of getting the model into the hands of people who would otherwise not have access?

[minimaxir]

That depends on which GPT-3 model they're using, and from both the API and the blog page, it's unclear.

Easy access to the 175B model would indeed be valuable, but it's entirely possible they're using a smaller variant for this API.

[gwern]

It's worth noting that at least one of the demos is not few-shot: the code completion one notes it was trained on Github.