GPT-3, One Year On: What Businesses Are Actually Building

It’s been roughly a year since OpenAI opened up broader access to the GPT-3 API, and enough time has passed that we can stop talking about GPT-3 in the abstract (“look what it can do in a demo”) and start talking about what people have actually shipped with it.

The clearest early winner is copywriting. A handful of startups have built products that take a short brief — a product description, a value proposition, a few bullet points — and spit out marketing copy, ad variations, or blog outlines. This is a genuinely good fit for the technology: the task is short-form, tolerant of some weirdness, and the human is going to edit the output anyway. You’re not shipping GPT-3’s raw text to a customer; you’re using it to blast through the blank-page problem.

Code assistance is the second big cluster. Tools that autocomplete functions, explain what a snippet does, or convert a comment into working code have shown up from multiple teams building on the API. It’s not “write my app for me” — anyone who’s played with these knows the output needs review — but as a way to cut down on boilerplate and Stack Overflow round-trips, it’s already useful enough that some developers say they don’t want to give it up.

Then there’s chatbots, which is the category I’m most skeptical of. GPT-3 is great at sounding fluent and reasonable in a single exchange. It’s much shakier at staying consistent over a long conversation, remembering what it said three turns ago, or refusing to wander off-topic. The chatbot products I’ve seen lean hard on prompt engineering and guardrails to keep things on rails, and it shows — the good ones feel curated, the rough ones feel like talking to someone who just wandered in from another conversation entirely.

The DALL-E question mark

Back in January, OpenAI teased something adjacent but different: DALL-E, a model that generates images from text descriptions. The demo images — chairs shaped like avocados, illustrations of imaginary animals — made the rounds because they were genuinely startling in how literally the model seemed to understand compound, invented prompts.

But it’s worth being clear about where that stands right now: DALL-E is a research preview. There’s no public API, no code release, nothing a developer can currently build on. It’s a glimpse of a direction, not a product category yet. If and when OpenAI does open it up the way they did with GPT-3, I’d expect a similar pattern to repeat — a flurry of demo tools first, then a slower shakeout of what’s actually durable to build a business on.

That pattern is the real story of GPT-3’s first year, honestly. The interesting products aren’t the ones that just pipe a prompt through the API and show you the result — they’re the ones that figured out which narrow slice of a workflow the model is reliably good at, and built the surrounding product to catch it when it isn’t. A year in, that’s still the hard part, and it’ll probably still be the hard part whenever DALL-E or whatever comes next gets its own API key.