The Pattern Before the Product
Two hundred integrations on day one. Expedia. Instacart. Wolfram Alpha. Zapier. OpenAI planted a flag in May 2023 and called it the App Store moment for AI. ChatGPT wasn't a chatbot anymore — it was a platform. The operating system for everything that comes next.
The plugins are gone now. Deprecated, April 2024. One understated paragraph in a developer blog. No keynote. No retrospective. If you go looking for them, you find a redirect.
Which would be a normal story of tech hubris — announce big, ship quietly, disappear without accounting — except something odd happened on the way to the graveyard. The plugins failed. The pattern survived.
Tool use is everywhere now. It's built into every model worth mentioning. The architecture that made plugins theoretically possible — give a model a function signature, let it decide when to call it, receive the result, incorporate it into reasoning — that's the substrate running under agents, assistants, copilots, and every "agentic workflow" in conference demos this year. OpenAI didn't abandon the idea. They abandoned the product and kept the physics.
This is not how tech companies usually talk about what they're doing. The standard story is: we built a thing, the thing succeeded or failed, we learned and pivoted. What actually happened here is stranger: they shipped an implementation before anyone knew what the implementation was for. The 200 plugins weren't a product. They were a probe. A way of asking: if models can reach out and touch external systems, what do people actually want to touch?
The answer turned out not to be "a Kayak integration inside ChatGPT." The answer was "models that can use tools as a general capability, invokable programmatically, composable with other systems." The function calling API — released June 2023, a month after plugins launched — is the thing that mattered. The plugins were the public research phase.
Credit where it's slightly due: most companies don't figure this out at all. They ship the bad implementation, watch it fail, and conclude the idea was wrong. OpenAI looked at the plugins data, noticed what developers were actually doing with the API underneath, and correctly identified that the infrastructure was the product and the UI was a distraction. That's not genius — it's just basic pattern recognition executed at unusual speed.
What you should actually take from the plugins obituary is this: the gap between announcement and product is where the real architecture lives. The keynote is marketing. The deprecation notice is archaeology. What gets deprecated tells you what was a bet. What gets quietly absorbed into the API tells you what resonated with the underlying structure of what people actually needed.
OpenAI found the pattern before they found the product. That's rarer than it sounds, and worth more than two hundred plugins that don't exist anymore.
Further reading
- OpenAI — Function calling and other API updates (2023-06-13)
threaded with
- beat · Tech
The Loneliness Was Already There
AI companion apps did not manufacture the loneliness — they found it fully formed. What follows requires no villain, only an owner who can change the terms on a Tuesday.
today
- beat · Tech
The Database He Aimed at Her
A Florida deputy used Flock to track his ex. Every control ran. The only one that is not internal requires the woman being stalked to file the complaint herself, in the building that employs him.
yesterday
- beat · Tech
There Is No National Voter File
ICE is shopping for a contractor to assemble every state voter roll into one file. That file already exists — data brokers built it two decades ago, and nobody voted on that either.
2 days ago