The Ghibli Thing Is Fine, But Kenton Shipping AI Code to Workers Production Is the One

OpenAI shipped native image generation in GPT-4o today — the Ghibli thing, the one everyone is currently doing — and yes, it's impressive, and yes, Gemini had been doing live image generation and clearly that was not acceptable, so here we are.

The image generation story will write itself. Half the internet is turning their selfies into anime frames as I type this.

The more interesting thing happened quietly, a couple hours earlier, over on Kenton Varda's timeline.

Kenton runs Cloudflare Workers. Not "leads the team" in the org-chart sense — he is, in the deepest practical way, the person who built the thing. Workers is not a product that emerged from a PowerPoint. It came from someone who thinks hard about distributed systems, who cares about the edge cases, who is not the type to get swept up in a hype cycle because his job is to actually run the infrastructure after the hype cycle is over.

He shipped AI-written code to production today.

Not a toy. Not a demo. Production.

And the way he talked about it was the tell — not "look what AI can do," not the breathless founder voice, not a VC tweet with the rocket emoji. Just: here is what I used, here is what it produced, here is where it lives.

That framing matters more than the code.

There's a specific credibility gradient for AI code adoption that I've been watching, and it runs from startup founders (early adopters, noise-to-signal ratio terrible) through mid-level engineers (reliable leading indicator) all the way up to the people who built the platforms the code runs on — who are, almost by definition, the last to adopt and the most meaningful signal when they do.

Kenton is at the far end of that gradient. The "I'll believe it when I see the people who wrote the runtime using it to ship changes to the runtime" end.

Today he moved.

The image generation stuff is a feature release. Somebody at OpenAI looked at the Gemini live image demos and ran a very standard competitive analysis meeting and the output was: ship it. Fine. Good, even.

The Kenton thing is a phase transition.

When the person whose job is "make sure software runs correctly at scale" decides that AI-generated code is production-quality enough to deploy — not demo-quality, not prototype-quality, production — something has changed that doesn't change back.

I don't know what the half-life is on the Ghibli meme. Probably two weeks. The Kenton data point doesn't expire.