What Developers Are Building With GPT-3 (and Why the API Feels Complicated)
A look at the chatbots, code helpers, and writing tools developers are prototyping on GPT-3's private beta API, and the tension around Microsoft's exclusive license.
If you’ve been anywhere near dev Twitter since the summer, you’ve seen the demos. Someone types a plain-English description of a website layout and GPT-3 spits out working HTML. Someone else feeds it a natural-language prompt and gets back a passable SQL query. There are chatbots that hold a conversation with a consistent-ish persona, writing assistants that draft marketing copy from a one-line brief, and code helpers that autocomplete entire functions from a comment. None of it is perfect — the outputs wander, hallucinate facts, and occasionally produce nonsense with total confidence — but the range of things people are throwing at this one model is the real story.
What’s notable is that almost none of these builders have access to GPT-3’s actual weights or architecture details. OpenAI opened a private-beta API back in the summer, and that’s the only door in. You send text, you get text back, and everything in between is a black box. For prototyping, that’s turned out to be fine — plenty of interesting products don’t need you to peek under the hood. But it’s a very different relationship to a model than developers are used to with, say, an open-source library you can fork and inspect.
The API-versus-model gap
That gap became a lot more concrete on September 22, when Microsoft announced an exclusive license to GPT-3’s underlying model. The API access that outside developers have been building on is still there, run through OpenAI, but the model itself now has one licensee with a deal that nobody else gets. It’s an unusual structure: broad, metered access to the outputs, tightly held access to the thing producing them.
That structure is exactly what’s drawn criticism through the first half of October. OpenAI started as a nonprofit with an explicitly open mission — the name itself was a promise. An exclusive license to a flagship model, even paired with a public API, reads to a lot of people as a pretty clean departure from that founding idea. You can still build a chatbot on top of GPT-3 today, but you’re building on rented infrastructure with one very large, very specific landlord now holding special rights to the foundation.
None of this has slowed the prototyping down, at least not yet. If anything the API’s ease of use — no infrastructure to manage, no training runs, just prompts and responses — is exactly why so many small teams and solo developers have been able to ship something in an afternoon. Code helpers seem to be the category with the most obvious near-term product potential: a tool that turns a comment into a function, or a description into a regex, is a genuinely useful thing regardless of what’s happening upstream in licensing deals.
Where this goes next is an open question. Will OpenAI expand API access and pricing to something sustainable for indie developers, or will the interesting capabilities increasingly funnel through Microsoft’s products first? I don’t think anyone outside those two companies knows yet. For now, the beta waitlist is still the entry point, and the demos keep getting more ambitious than the terms of access.