Andrew Ambrosino reveals he rewrote chatgpt.com from scratch in five weeks to unify the Codex and ChatGPT desktop experience

Andrew Ambrosino leads OpenAI's Codex desktop and Space teams. His DevDay appearance doubles as a reveal: the six-week social-media disappearance his followers noticed was a full rewrite of chatgpt.com from scratch.

The rewrite

Ambrosino says his team rebuilt chatgpt.com over five weeks to run on the same codebase as the Codex desktop app. The result is that the Codex desktop experience — its agentic workflows, persistent threads, file-system access — is now the default chatgpt.com experience for everyone.

The second change shipped alongside it: ChatGPT Work on desktop now runs in the cloud by default while retaining access to the local machine. That closes the awkward split between ephemeral ChatGPT threads and a separate Codex environment. Dots and Codex Cloud, both announced at DevDay, follow the same architecture: cloud-native, with local machine access layered on top when needed.

“One of them is that we completely rewrote chatgpt.com in the last five weeks to be the Codex app... So now we've got the same Codex desktop experience you're used to, the ChatGPT desktop experience is now chatgpt.com. And then the second merge that we've done is now by default, ChatGPT work on the desktop works on the cloud.”

Space and the multiplayer bet

Space is OpenAI's move into collaborative, agent-native workspaces. Ambrosino describes it as a way to curate shared work with teammates and their Dots, pulling in documents, Slack threads, and live data without leaving the ChatGPT surface. The pitch is that it brings Codex and ChatGPT into a "multiplayer era" — though it's early, and Ambrosino acknowledges the product has gone through multiple internal iterations before reaching this form.

Product timing and the iteration loop

Almost everything released at DevDay, Ambrosino says, is a "take five" on ideas OpenAI tried internally before the models were ready. Dots went through local and cloud variants. Space evolved from an internal knowledge-base product and something called Canvas. Codex Cloud is explicitly framed as version two: the original Codex was cloud-based, the team moved into a local-agent era, and now the full local-agent capability is back on cloud infrastructure. The pattern is consistent — ship early, pull back, wait for model capability to catch up, re-launch.

Dynamic UI and Ultrafast

The live weather chart with haptic feedback that surfaced organically in ChatGPT is the kind of on-the-fly UI generation Ambrosino's team has been working toward. He's candid that the challenge is real: you can't have an instant-chat model writing a full web app on every query. The research, he says, is now good enough to make it feel excellent. Ultrafast matters here specifically — the speed closes the gap between generating a dynamic chart and it feeling like a native interface rather than a loading screen.

Ambrosino's most concrete Ultrafast use case: tagging a Dot in a Space session, asking it to pull a DAO growth chart from live business data, and having it appear immediately. His framing — "time is money, save both" — lands as a product principle, not just a line.

DevDay itself was built with coding agents

Ambrosino notes that the entire DevDay event was planned using coding agents. They no longer look like coding agents on the surface, but the underlying infrastructure is the same. It's an anecdote, but it's also the clearest signal of where OpenAI thinks the adoption curve is heading: past developers, into operations, logistics, and everything else.

Ambrosino says he's going back into another six-week lock-in the following day.