The GPT-3 App Store Nobody Officially Built

OpenAI opened the GPT-3 API to public beta back in November, and in the months since, access has kept loosening. What started as a tightly gated research demo is now the substrate for a small wave of startups. Most of them are doing one of three things: generating marketing copy, powering chatbots, or — the one I keep coming back to — finishing your code for you.

The copywriting tools are the easiest sell. Feed GPT-3 a product description and a tone, get back five headline variants. It’s a solid fit because the cost of a bad output is low: you skim, you pick, you edit. Chatbots are messier — GPT-3 wasn’t trained to stay on-topic or refuse gracefully, so most of what I’ve seen bolted on top is scaffolding to keep the model from wandering into nonsense mid-conversation.

The code-completion angle is the one worth watching closely, though. A handful of small teams have been building editor plugins and web tools that take a comment or a function signature and generate plausible code underneath it. The quality is uneven — GPT-3 wasn’t trained specifically on code, it just absorbed a lot of it incidentally from scraping the public web, including GitHub. But “uneven and occasionally surprising” describes most GPT-3 output, and for boilerplate — a regex, a simple API call, a test scaffold — uneven is often good enough to save real time.

What’s interesting is that none of this is centrally coordinated. OpenAI isn’t (as far as anyone can tell) building a flagship coding product itself right now — it’s letting third parties experiment on the API and presumably watching what sticks. That’s a very different rollout strategy than shipping a single polished tool, and it means the “AI pair programmer” concept is being stress-tested in public, by people with no inside information, using a general-purpose language model that was never specifically optimized for the job.

I’d bet the current crop of code-completion tools built on the raw GPT-3 API represent a floor, not a ceiling. If a foundation model can produce decent completions from general web-scraped code with zero code-specific fine-tuning, the obvious next move for anyone with the resources is to train a variant specifically on source code — more data, tighter loss on the task, maybe some way to check outputs actually compile or pass tests. Whether that’s OpenAI, Microsoft-owned GitHub (which has the world’s largest pile of permissively licensed code sitting right there), or someone else entirely, I don’t know. But the ingredients are all visible right now: a capable base model, an API that’s easy to build against, and a swarm of small teams proving out the use case for free.

The bigger question is what happens to the developer workflow if this actually gets good. Autocomplete already changed how people write code once, moving from “recall the exact syntax” to “recognize it from a list.” A model that can draft whole functions from a comment would be a bigger shift than that — less about recall, more about editing and verifying someone (something) else’s first draft. That’s a different skill, and not everyone will like the trade.