All podcasts / Lenny's Podcast / Summary

Why OpenAI is merging Codex and ChatGPT and the future of knowledge work | Andrew Ambrosino

2026-06-28 - 70 min - source - Read full transcript
Lenny Rachitsky (host)Andrew Ambrosino

Key insights

Implementation is no longer the scarce resource in product work; curation and taste are.
Andrew argues the product process has inverted: instead of exhaustively de-risking one idea before building it, teams now spin up dozens of parallel prototypes for the same problem because implementation is nearly free. The hard part shifted to deciding which of 90 competing explorations is actually good and how to fold them together.
taste-and-curation
PRDs and prototypes are not obsolete, they are situational, and picking the wrong medium anchors people prematurely.
Andrew pushes back on the popular claim that 'PRDs are dead, prototypes are in.' He says a document is still right for creating product clarity in a vague area, while a prototype is right for stress-testing an interaction pattern; the mistake is defaulting to a fully polished prototype so early that people mistake it for a launch-ready decision.
design-process
AI models still lag noticeably behind humans at design specifically, for structural reasons distinct from coding ability.
Design is harder to grade than code (there's no equivalent of 'does it compile'), and frontier labs have historically prioritized capabilities that accelerate their own AI research, which correct code does and good design does not. Andrew expects this practical gap to close over time.
design-process
The deeper design gap for AI is architectural, not visual: propagating a systemic change across a codebase's shared abstractions.
Beyond aesthetics, Andrew says models still struggle with the interplay between visual design and code structure - e.g. after a rebrand, updating 263 components one by one (shallow) versus recognizing the shared semantic pattern connecting visually different components (deep). That abstraction layer still feels out of reach for current models.
design-process
Roles haven't vanished, they're now defined by where you spend your time, not a fixed job title.
On the Codex team, designers write code and PMs write code; a person's role is 'the average' of what they actually work on, not a fenced lane. Andrew credits this to the product originally being built for engineers, so designers and PMs on the team already spoke a technical language.
role-collapse
Product-market timing can matter more than product shape: the same feature can fail or succeed purely based on model capability at launch.
Andrew says the Codex app released in February would have 'absolutely failed in the market' if it had shipped in November with the November-era model, even with an identical design. He draws a direct line through OpenAI's own history of similar-shaped agent features (Operator, Atlas agent mode, Codex) succeeding or failing based on whether the underlying model was ready, not whether the concept was right.
ai-product-development
Eliminating product/design roles wholesale is a mistake many companies are making right now.
Andrew calls it 'by the way, a terrible idea' when companies get rid of the product role and declare everyone a builder, because it discards accumulated discipline-specific best practices that took years to develop. He distinguishes this from removing rigid lane boundaries, which he welcomes.
role-collapse
OpenAI is merging Codex and ChatGPT because non-engineers kept choosing the developer tool anyway.
When OpenAI built separate surfaces for other functions (marketing, finance, legal) with tailored versions of ChatGPT and Atlas, employees kept using the Codex app instead even though it was actively hostile to non-engineers (showing raw code, asking to run shell commands). That forced the realization that developer tool versus general knowledge-work tool is a false dichotomy, so the two products are converging into one home base.
codex-chatgpt-merger
Teams should build things the current model can't quite do yet and revisit them each model generation, instead of killing a feature after one failed attempt.
Andrew describes releasing early, weaker versions of a feature (like the in-app browser or Operator-style agent) as creating artifacts to re-test against future models, rather than treating a weak first attempt as proof the feature concept is bad. The team deliberately keeps a backlog of prototyped-but-shelved ideas to revive as models improve.
ai-product-development
AI agents are starting to extend past their native UI by controlling other apps directly, including building their own integrations on demand.
An OpenAI videographer had Codex edit Premiere Pro projects by manipulating the underlying project files, and when that hit a limit, Codex built itself a Premiere Pro extension to control the app directly. Andrew separately describes asking Codex to complete a Google Cloud Console setup via literal computer-use clicking when no API connector existed.
codex-chatgpt-merger
Long-range roadmaps are mostly false precision right now; only near-term plans deserve real detail.
Andrew says the team plans in detail for the short term but deliberately keeps anything nine months out 'very hazy,' because adding precision to a plan that far ahead just wastes time given how fast model capability changes the feasible feature set.
ai-product-development
Hiring on the Codex team filters for high agency plus taste to separate signal from noise, not raw coding output.
With unlimited tokens available to everyone, Andrew says the differentiator isn't who can generate the most code but who has the taste to know which of many possible outputs is actually good, paired with the accountability to own outcomes rather than just activity.
taste-and-curation

Books referenced

Media referenced

Companies

Techniques and frameworks

Summary

Andrew Ambrosino, product and engineering lead for OpenAI's Codex app, joins Lenny to explain why the previously separate Codex (developer tool) and ChatGPT (general assistant) are converging into a single product, and what that says about how product work itself is changing. His central claim is a full inversion of the traditional process: implementation used to be the expensive, precious resource that justified months of amateurish planning and de-risking before a single prototype got built. Now that any team can spin up a working feature from scratch in an afternoon, implementation is cheap and abundant, and the scarce resource has become taste and curation - the judgment to look at 90 parallel, uncoordinated attempts at the same idea and decide what's actually good.

That inversion reshapes almost everything else discussed in the episode. PRDs aren't dead, Andrew argues, contrary to the popular narrative, but the industry has gotten sloppy about matching the right medium (a document versus a prototype) to the right purpose, and prototypes that look production-ready are getting mistaken for validated decisions when they're really still early exploration. Design specifically still lags coding in what AI models can do well, for reasons Andrew traces to how hard design is to grade automatically and to labs' historical incentive to prioritize capabilities that accelerate their own AI research. The deeper problem isn't aesthetics, it's abstraction: models still struggle to propagate a systemic change, like a rebrand, across the shared semantic patterns connecting components that look nothing alike on the surface.

On roles, Andrew describes heavy collapse inside the Codex org specifically, where a person's function is defined as "the average of where they're working" rather than a fixed job title - designers write code, PMs write code, and the team explicitly organizes with a "zone defense" mentality, spreading tastemakers across the problem space instead of clustering. He's careful to distinguish this from the more extreme move some companies are making of eliminating the product function altogether, which he calls a terrible idea because it throws away real accumulated best practices along with the rigid lane boundaries that deserved to go.

The merger story itself comes from a very concrete internal failure: when OpenAI tried building separate, purpose-fit surfaces for non-engineering teams (marketing, finance, legal), those teams kept using the Codex app anyway, even though it was actively unfriendly to them - showing raw code, asking permission to run shell commands. That signal pushed OpenAI toward treating Codex and ChatGPT as one general knowledge-work home base that starts simple and grows in complexity as needed, rather than maintaining a permanent split between "developer tool" and "everyone else tool." Concrete examples reinforce how far this generalization already reaches: an OpenAI videographer got Codex to edit Premiere Pro footage by manipulating project files directly, and when that wasn't enough, Codex built itself an extension to control Premiere Pro; Andrew separately describes having Codex complete a tedious Google Cloud Console setup purely through computer-use clicking because no API connector existed.

Andrew closes by tying model timing to product outcomes directly - he's confident the Codex app that shipped in February would have failed outright if released in November with the earlier model, identical design and all - and by describing how his own team plans: full detail for the near term, deliberately vague past nine months, because false precision on a longer horizon just wastes effort given how fast the feasible feature set moves. The lightning round is unusually personal, centered on his kids' bedtime books (The Gruffalo, The Big Orange Splot) and a genuine admission that he doesn't have many hobbies outside work and parenting right now.

Notable Quotes

"The implementation is actually not the expensive part anymore. It's, dare I say, taste." - Andrew Ambrosino

"I've heard a lot of companies be like, we're getting rid of the product role, which I think is, by the way, a terrible idea. And everybody's just going to be like a builder." - Andrew Ambrosino

"I am very confident that the Codex app that we released in February, if that had been ready in November, it would have absolutely failed in the market. And the only difference was the models between November and February." - Andrew Ambrosino

"Nobody would leave the Codex app for the apps that were allegedly for these other personas." - Andrew Ambrosino, on why OpenAI is merging Codex into ChatGPT

"You need to be able to determine what's signal, what's noise, in a world of just infinite content." - Andrew Ambrosino, on what he looks for when hiring