Why OpenAI is merging Codex and ChatGPT and the future of knowledge work | Andrew Ambrosino
OpenAI's Codex lead reveals why 90% of the company uses a developer tool despite it being "actively hostile" to non-engineers—and what this says about the collapse of traditional product roles.
Read OriginalMy Notes (4)
Agency. Exactly. You can just do things. Marry the outcome, not the process
And it also speaks to how I sort of see IC versus management, which is that it's not that management is going away. It's not that everyone's an IC, but everyone's kind of both now. If you're an IC, you're not typing code out character by character. You are managing something. You're managing agents, you're managing work that is happening that comes together to do a certain thing. If you're a manager of teams, you're doing the same thing, just at a different granularity.
Everybody's sort of defined less by the fence and the boundaries of where design stops and engineering starts, but more the average of where they're working. So if you average up all of the things that somebody on our design team does, there's plenty of code-writing things, there's plenty of things that are product work, but on average their dots are over here.
the most valuable person right now is someone who can take an idea from idea to done with the taste to know this is great. Just shepherding throughout, this obsession with making it awesome, this kind of high agency, high taste person
The same product ships six times before it works, and the shape never changes
- Product used to fail because of shape and communication. Now it often fails because the model underneath wasn't smart enough yet.
- You may need to release the same thing six times before it lands, with the shape unchanged each time.
- Whether a feature is good stopped being a question about its design and became a question about intelligence.
- Codex proves it. Andrew is confident the February Codex app would have failed in the market if it had shipped in November. Nothing about the product differed. The only change was the models between November and February.
- Operator, Atlas, Codex and ChatGPT all carry the same underlying feature. Operator didn't work out. Very cool idea, wrong moment.
- Re-releasing it with better intelligence changed the outcome completely.
- So don't be stubborn about calling something a bad feature. It might not be ready yet.
Summary used for search
TLDR
• AI has inverted product development: implementation is now cheap, taste/curation is expensive—leading to 90 uncoordinated prototypes instead of PRDs
• The Codex app would have failed if launched in November vs February (same product, smarter models = different outcome)—so teams now build features that don't work yet and wait for AI to catch up
• Product managers now operate in "zone defense"—spreading out for coverage rather than owning swim lanes, since everyone can build anything
• OpenAI tried building separate tools for different personas but everyone stayed in Codex, revealing the future is a unified AI home base that coordinates all work
• Andrew automates his entire job inside Codex (Slack briefs, release management, status tracking)—the app even builds its own Premiere Pro extensions when needed
In Detail
Andrew Ambrosino describes a fundamental inversion in how products get built at OpenAI. The traditional process assumed implementation was expensive, so you derisk everything upfront through documents and research. Now implementation is free—anyone can build anything—which creates a new problem: 90 people building 90 different prototypes of the same idea, all looking production-ready but actually in early exploration. The bottleneck has shifted from "can we build this" to "should we build this and how should it work"—pure taste and curation.
This shift is forcing role collapse, but not in the simplistic "everyone's a builder now" way. OpenAI uses a "zone defense" model where product managers spread out to provide coverage rather than owning features, since engineers are product-minded and designers write code. Roles are defined by the average of what you do, not rigid boundaries. The dangerous version is companies eliminating product/design entirely and losing decades of discipline-specific knowledge. The healthy version is more overlap with specialists who can work across the stack.
The most striking insight is about timing and model intelligence. The exact same Codex app would have failed in November but succeeded in February—the only difference was model capability. This means teams now build features that don't work yet, let them sit, and retry with each model improvement. OpenAI initially tried building separate tools for different personas (Codex for engineers, ChatGPT for others), but discovered everyone stayed in Codex despite it being "actively hostile" to non-technical users. This revealed the right shape: a unified home base that coordinates work across tools, whether through connectors, computer use, or building its own extensions (like when a videographer used Codex to build a Premiere Pro plugin so Codex could edit videos for him).