Andrew Ambrosino on why Codex and ChatGPT are merging and why AI has flipped product work
Lenny's Podcast
Andrew Ambrosino leads product and engineering for the Codex desktop app at OpenAI. Their background spans design, engineering, product management, and founding companies. In this conversation with Lenny Rachitsky, the central question is how AI is changing product work now that anyone can build anything. Ambrosino's answer is that the product process has been turned upside down. Implementation used to be the expensive step. The scarce thing now is judgment about what to build and how to shape it, which Ambrosino, a little reluctantly, calls taste.
Ambrosino also warns against the more extreme conclusions people draw from this. They argue that PRDs are not dead, that the design process has not simply disappeared, and that eliminating roles is a mistake. They also explain how Codex grew from a developer tool into something OpenAI employees across every function use, and why that led the company to fold Codex and ChatGPT together.
For scale, Lenny cites these figures in the introduction: Codex usage has grown sixfold since January, it has over 5 million weekly active users, and nearly all OpenAI employees use it weekly, not only engineers. Ambrosino has publicly stated the goal of making Codex "the best desktop app that has ever existed." They explain that the quality bar has to be high enough that people never hesitate to open it for the next task, the way people reflexively open a browser tab.
The inverted product process
Ambrosino says the hardest part of leading product teams right now is that the process has inverted. With current models, from OpenAI or anyone else, they believe anyone starting from scratch can stand up whatever feature they want. Everyone at OpenAI has effectively unlimited tokens, and Ambrosino describes the staff as very agentic with lots of ideas, so everyone is building everything.
Traditional product process worked the other way. Even after teams moved past waterfall, the process assumed implementation was expensive. Teams de-risked everything up front through documents, research, and prototypes, because designs and prototypes were cheaper than real code. That assumption no longer holds. Ambrosino guesses that for one feature the team badly needs, there are probably "90 different uncoordinated" efforts implementing and trying it.
Ambrosino stresses that people are not doing fundamentally different jobs, and that skill sets and roles have not vanished. The order of the work has reversed. Once implementation is cheap, the costly part is curation. Out of 90 attempts, what is good about each one? What should be folded into something else? How should the idea be framed? Should it belong to another feature? How many segments should a toggle have?
Documents vs. prototypes: choosing the medium
Lenny summarizes the shift as "people used to write docs, now they build prototypes." Ambrosino pushes back. They have seen many product leaders declare that PRDs are dead and prototypes are in, and they don't believe it.
Their reasoning is that cheap implementation creates a temptation in both directions. Non-engineers, especially people who never had the time or interest to code, are tempted to skip straight to a prototype and say, "Let me just show you what I mean." Engineers, Ambrosino has noticed, are tempted to produce large numbers of documents "that are not worth reading." Their point is not that documents are bad. When implementation is abundant, you have to pick the right format for the point you are making. If the goal is product clarity in a vague area, a document may be best. If the goal is to put an interaction pattern in people's hands and stress-test it, a prototype is better.
Lenny connects this to a concept a previous guest shared, the "primal mark": the first mark an artist makes shapes everything that follows. A prototype made too early can anchor everyone to the wrong idea. Ambrosino agrees and adds a second problem. In the old world, the medium itself signaled where an idea was in the process. Anything that looked like the production app must be late-stage, meaning assumptions had been de-risked, design had reviewed it, and it served a real business goal. That signal existed because resources for building were only granted after de-risking. Now the link is broken. An exploration can look production-ready while having the wrong model of where research is going, what users want, or what the business needs. So teams have to say explicitly what stage a thing is at, whatever form it takes.
What "taste" actually means
Asked what good taste looks like in practice, Ambrosino cites a recent tweet from Linear's head of product (they weren't sure of the attribution). The tweet argued that people overemphasize the aesthetic side of taste, and used Paul Graham as an example of someone with clearly great taste who wears cargo shorts.
For Ambrosino, taste includes all of the following:
- Aesthetics. For example, noticing that an interaction animation doesn't match its semantic meaning because it is "too snappy for what it's actually trying to convey." They admit they probably focus on this too much.
- Systems thinking. How the piece fits into the larger system.
- Direction. Where the product is heading and which theme a piece belongs to.
- Presentation. How the work is framed and communicated.
They see the real taste question as the broader one: if we can build anything, what should this be, what is the goal, and how do we get there? Knowing what to work on, how to present it, and which medium to use is, in their view, becoming the most important skill in every field.
Why AI is still weak at design
Lenny notes that AI output is rarely great design, and people can usually tell "Claude design" or "Codex design" at a glance. Ambrosino, making clear they don't work in research and might "get yelled at," offers practical reasons and harder ones.
The practical reasons are these. First, design is harder to grade than code. Code can be checked by whether it compiles and does what it should. Building a loop that teaches a model what good design is requires human taste as part of the feedback mechanism, which is more tedious and onerous. Second, labs historically invest in capabilities that accelerate AI research. In the early era of coding models, it was obvious that writing correct code would speed up research. The same case is hard to make for design. Ambrosino expects these practical obstacles to go away and models to get "pretty good at design."
The harder problems are murkier. One is culture and novelty. Ambrosino recalls a period, probably last year, when every new website seemed to copy Linear's. A model producing Linear's site would have looked impressive once, but a model that outputs Linear's website every time is not solving the real challenge. In software engineering you almost want the model to lean on known patterns. In design, randomness and novelty matter.
The other is the abstraction layer between design and code. Having written and supervised much of the early Codex app's code, Ambrosino describes design decisions that live in the codebase: this element in one corner should share X, Y, and Z with that element elsewhere. Their example is a company rebrand. The shallow version is updating 263 components one by one. The deep version is understanding the shared semantics between things that look different but are both lists conveying the same interaction pattern. That layer, they say, still feels out of reach for current technology.
Lenny adds that the Codex app itself was a new paradigm, neither a terminal nor an IDE but a chat interface that codes and shows you the code. He suggests that inventing something like that is where human creativity still matters. Ambrosino agrees, "for now."
Is the design process dead?
Lenny brings up Jenny Wen, head of design for Claude, who argued on the podcast that the design process is dead. Ambrosino says they probably agree on a lot, and that they weren't a fan of the formal design process even before AI. From their time running a startup and hiring designers, they recall a snarky article about "the case study factory." It criticized designers being trained to value the process above outcomes, on the assumption that anything that went through research, divergence, and convergence was guaranteed to be good, even if nobody liked or used it.
Ambrosino notes that the process has always assumed implementation is expensive and you can only afford to build once, so you must exhaust the problem and solution space first. Tools like Figma and Origami let teams pull interactive prototypes earlier to get insights sooner. That was real, even if it became a meme about executives asking for a prototype and expecting it to work. Now all of implementation can be pulled into that early stage. The result is a mismatch: people see a fully polished prototype and ask, "Can we release this now?" In fact, it is still early-stage exploration, and nobody is saying so.
So Ambrosino says "the design process is dead" is both true and false. Designers have more tools than before. They can put ideas into the live product and A/B test them. Many companies, OpenAI included, now keep a "baby" version of the product, which Ambrosino calls "baby Codex." It is a dramatically simplified codebase that approximates the production app's interactions, so it is fast to vibe-code over: what if the sidebar worked like this, what if a pane opened with a group chat? If you are attached to the specific day-to-day tools of the old process, Ambrosino says, it is dead and you won't have a good time. But the overlay of the process, knowing and stating where you are in it, matters more than ever.
Role collapse on the Codex team, and its limits
Asked whether design reports to them, Ambrosino jokes that it "depends on the week." The teams sit together and work closely. They say the Codex org has seen more role collapse than other parts of the company or the economy. They attribute this to Codex being a technical product for engineers: the designers "speak engineer," product managers write code, and Alexander Embiricos, the Codex product lead, has a master's in computer science, which Ambrosino notes they themselves do not.
Ambrosino describes each person as defined less by fences between disciplines and more by the "average" of where they work. A designer does plenty of coding and product work, but on average their dots cluster in design. The process is also shaped by dogfooding. The team tries to do as much as possible inside the app even when it isn't the best tool, so that it becomes the best tool. They often deliberately don't improve their own process so they feel the product's gaps. Ambrosino calls this "a deeply uncomfortable place to be."
Lenny asks whether "member of technical staff" titles point to a future where everyone is simply a builder. Ambrosino notes that this title tradition comes from research-focused companies and likely predates OpenAI (they mention Xerox). They then state a concern: eliminating roles can eliminate the idea that disciplines are specialties with knowable best practices. They have heard of companies getting rid of the product role, which they call "a terrible idea." In those cases, a discipline with real best practices, real processes, and a history of real failures gets abandoned because people think, "Oh, I wrote some code."
They welcome losing the "this isn't your lane" boundary, but they argue that not everyone can work on everything in either breadth or depth, which is why managers won't go away. They also say engineers are often guilty of assuming engineering is a skill while other roles are "just people vibing." Their example: you can use Excel, but that doesn't mean you can work on the finance team.
What is changing, in their view, is that switching roles and learning a role's best practices have become easier, and effectiveness is less tied to mastering a specific tool. They say they spent a long time feeling they shouldn't be a software engineer because they didn't care about assembly or memorizing TypeScript syntax. That kind of tool-based gatekeeping is what's eroding. Beyond that, they think people hyperbolize.
Team shape and "zone defense"
Asked how big the Codex team is, Ambrosino gives their usual answer: "somewhere between 10 and a few thousand." They mean it seriously. Everything at OpenAI, including model research, computer and browser use, model personality, and front-end infrastructure, feeds into this product. Still, they don't accept PRs from thousands of people on whatever they want. The desktop team has double-digit engineers, about half that many designers, and a few product people. Common traits are agency and taste. Many are former founders or people who did "founder-shaped" work at larger companies. OpenAI lets teams grow large, and they are mostly ICs.
Ambrosino and Embiricos use a "zone defense" analogy for product managers. If two product people are working too closely, that is usually a bad sign. The PMs spread out like a force-directed layout, looking for gaps. With chaotic idea generation everywhere and year-long top-down plans no longer working, taste-makers need to guide ideas from inception to what the product should be, which requires coverage across the whole area. They also want to hire product-minded engineers, so that a flood of code doesn't need full-team review for product coherence.
Lenny describes the most valuable person now as someone with high agency and high taste who can shepherd an idea from start to finish. Ambrosino agrees and links it to IC versus management. Management isn't going away, but everyone is now both. An IC no longer types code character by character. They manage agents and work streams, while managers do the same thing at a coarser granularity. Beyond command of their discipline, Ambrosino looks for the taste to separate signal from noise when tokens are unlimited and "we can't just be doing slop."
Planning when the models keep changing
Ambrosino says people are frustrated with them about planning all the time, and that the team does nothing revolutionary. The shorter the horizon, the more detail a plan gets. They do plan nine months out, but that plan stays hazy, because any precision added now is false precision. Speaking for the applied product side, not research, they note that anything planned in November might have held for December but "isn't what happened."
What planning does require is a view of which model capabilities are likely to arrive on what timeline. At their previous company, when the team began building features on top of models, the normal product process broke down. It became: list everything you might want to do in the next year or two, prototype all of it, decide what is ready now, let the rest sit, and retry each item whenever the models leap forward. Whether a feature worked depended on whether the model was smart enough, not on the feature's shape.
Ambrosino's clearest example is the Codex app itself. They are "very confident" that the same app, released in February, would have "absolutely failed" if it had been ready in November. In their view the only difference was the models in between. A product with the exact same shape would have had completely different results a few months apart.
Building ahead of the models, and being "too AGI-pilled"
This leads to a pattern: build things that don't work yet and keep them until models catch up. Ambrosino says the team does a lot of this, but again people must be clear about which stage something is in. Many still have the muscle memory that code written means code shipped. Instead, they say, working code is now an artifact you can test against future models.
Their example is browser-based agency. OpenAI had Operator in ChatGPT, then agent mode inside the Atlas browser, and now an in-app browser in Codex. Ambrosino sees a thread through Operator, Atlas, Codex, and ChatGPT that is "fundamentally the same feature." Re-releasing it with different levels of intelligence changes the outcome completely. So they push people not to conclude that a feature is bad just because it isn't working yet.
The opposite failure is excess ambition, which Ambrosino associates especially with research: "at the limit, the model can just do this." Their example is the original Codex release, Codex on the web. You gave the model a task, it went away, and came back with finished work. The form factor wasn't radical, but the model didn't do the task well enough. Then Claude Code launched as a fully local tool. It was not hooked up to the cloud, it asked the user questions, and it didn't pretend you could delegate your whole life to it. According to Ambrosino, that worked much better because it matched where the models were. "We were too AGI-pilled for the moment," they say. In the past, market positioning told you the shape and communication of a product. Now you might need to release the same thing six times, without changing its shape, before it works.
On pushing teams toward ambition, Ambrosino notes that once a product exists, people naturally find paper cuts and optimize, and users on Twitter help by pointing them out. That work matters, but it's why OpenAI also keeps a culture of bottom-up exploration. Just as the Codex app disrupted ChatGPT in some ways, they expect it to be disrupted by some future effort. One team can't always do both the disruptive work and the maintenance of quality, so the process has to allow for both.
Beyond loops: autonomous development
Asked what the frontier is, Ambrosino jokes that "loops are so last week." By last year's standard, 100% of the product's code is AI-written, so the useful question is whether that code is written supervised or unsupervised. They welcome the moving goalposts as a sign of progress.
The team has explored autonomously developed software and "harness engineering," such as an overnight run that does garbage collection on the codebase. One problem they say all models share is that they tend to increase complexity. "If research is listening at any company, please make the models better at deleting code." That becomes a serious obstacle when development runs on autopilot, on both the codebase side and the human side. How do you teach a model which feature requests to build, which to ignore, which to group and reframe, and how to build the right abstractions? All of this is improving, but Ambrosino says they are not yet at the point of setting up a loop that "improves the app" by listening to Twitter, Slack, and email, though they're trying to get there. Asked whether a model could one day just be told to win the market or make a billion dollars, they say they're not in the business of saying never or always.
How Ambrosino runs their own job in Codex
Ambrosino describes their own use of Codex as following the product's evolution. At first the goal was personal: make the app good enough to write the Codex app with it. That created a tight dogfooding loop: hit something you can't do, fix it, do more. After launch, people started using it for different kinds of work, and Ambrosino's job shifted toward hiring and product discovery. So they started using the app for that too: building spreadsheet models, running internal deep research on past efforts in an area, and tracking what everyone is working on so they can steer things that are off track.
A set of releases around May introduced the in-app browser, computer use, and artifact creation to Codex. Ambrosino calls it the "Codex for almost everyone" release, and says it was the first time they "vibe-coordinated" a release. They kept a Notion doc listing everything that had to happen and automated gathering updates from pull requests and Slack channels to update the status tracker. That's common now, but at the time they felt at the bleeding edge of release management.
Today they wake up to a daily brief summarizing which of their roughly 3,000 Slack channels need attention. They can reply, for example asking for five questions to answer. Setup is still exploratory. They created a scheduled task describing which channels and topics matter, categories to watch, and context. For the first few runs it needed steering, and instead of editing instructions directly they could tell it things like "de-emphasize this workstream" or "this came up and wasn't in the brief." Ambrosino acknowledges this is the chatbot's core problem: they have time to set it up because it counts as product discovery, but people outside OpenAI shouldn't have to figure it out. Lenny notes this makes Codex work a lot like OpenClaw. Ambrosino says that if the app lacks, say, a Slack connector, it will offer to add one, because the least the product can do is let people ask it how to do things. They add that they don't think that's enough.
Lenny shares an example. He built an email filter in Codex that labels unsolicited cold email. Setup required configuring Pub/Sub in the Google Cloud console, an interface he found painful, so he asked Codex to do it. It took over his computer and clicked through the setup. Ambrosino sums up the model's attitude: "I don't care if you don't have a connector, man. I'll just start clicking." They add that deciding when the agent should use a connector, the in-app browser, the Chrome extension, or computer use is designed largely by feel.
They also describe how personal workflows feed the product. Everyone at OpenAI builds different personal systems, and when themes emerge, the team considers making them first-class. Memory is one example. Many people inside and outside OpenAI build an Obsidian vault or Notion space as a "mind palace" for the agent. Ambrosino thinks people shouldn't have to, and that a memory feature should handle it. Other workflows reflect how a specific person does their job and should stay personal. Deciding which becomes a primitive is ongoing.
The browser, and running SaaS apps inside Codex
Lenny raises Dan Shipper's prediction that people will run their SaaS apps inside Codex instead of Chrome. Ambrosino says Shipper Slacks them daily with requests. They review OpenAI's attempts at browser-shaped activity: Operator, ChatGPT agent mode, Atlas, now the in-app browser, plus a Chrome extension that connects the app to Chrome.
Some of the obstacles are, in their words, very boring. The app originally launched on Electron, where in-app browsers are "kind of janky," so the browser was meant only for testing front ends during development. The team then moved to the OWL stack that powered Atlas, which brought multiple tabs and enterprise security so users can log into their sites. The open question is what the browser should be. Is it only for the agent, fast and without Playwright's latency, while you keep using Chrome yourself? Or is the app for everything, including your own browsing? That path isn't well traveled and raises tedious problems like keyboard shortcuts: should they map to VS Code, Chrome, Linear, or something new? Lenny notes the app has to serve both people who have never built anything and power users like OpenClaw's Peter Steinberger. Ambrosino says they're "not convinced" they'll win Peter over from the terminal, but will keep trying.
From developer tool to home base: why Codex and ChatGPT are merging
Ambrosino tells the story behind the merger. Codex began as a CLI. The desktop app was built with conviction but some uncertainty, as a developer tool. It would not be an IDE but a "right-sized surface," more than a chatbot, where you could see code but not edit it. In January and February, before public launch, internal dogfooding showed clear product-market fit for engineering and research workflows. The team focused on raising the quality bar.
Then people from marketing, comms, finance, legal, and essentially every discipline started using the Codex app, even though Ambrosino says it was "actively hostile to these people." It showed them code and asked for approval to run rg. OpenAI responded by bringing Codex into other surfaces meant for those users, including the ChatGPT desktop app and the Atlas browser, as general knowledge-work tools. Then came "the most annoying problem": nobody would leave the Codex app for the apps supposedly built for them.
Ambrosino's takeaway is that "developer tool versus general knowledge work tool" is not a binary. Just as a person's role is the average of their work, a product can infer a lot from what a user is doing. Someone doing Excel work doesn't want git repository information. The product can start simple and grow more complex as needed, possibly with modes that make entry points legible. The team believes it has the right shape for deep, vertical work, and works closely with OpenAI's finance, science, and legal teams. If the right extensibility primitives and general model are in place, they argue, you can do anything with it. The challenge is generalizing.
Ambrosino says someone called the result a "super app," and they wish they hadn't, because now they hear about it all day. They describe it instead as a home base: a place to track everything you need to do across surfaces, where some work happens in the app and some happens in other apps it opens. For example, Codex has a built-in spreadsheet editor, but it's probably not good enough for the financial modeling behind OpenAI's multi-billion-dollar fundraising. So the app talks directly to an add-in in desktop Excel, and when it's done you can close Excel. The idea, they say, is not to draw a rectangle where everything must happen, but to have a home where you start, end, and automate work using whatever tools are needed.
The videographer who built a Premiere Pro extension
Ambrosino's favorite example of this model involves Brent, OpenAI's in-house videographer, who had to edit the launch videos for the Codex app, shot in the same room as the interview. Brent edited all of them with Codex, which became one of the early moments where the team asked what people were doing with this thing.
Brent started out of curiosity about whether Codex could edit video. Codex has no video-editing UI. It understood that Brent used Premiere Pro and could make some edits by changing the files behind the project, but not all of them. So Codex built itself a Premiere Pro extension, installed it, and then sent it instructions, such as asking it to change a marker inside Premiere. Ambrosino calls this "pretty nuts."
To them, it shows one of two opposite models the team is pursuing at once. In the first, Codex and ChatGPT work with specialized tools people already use, through connectors, computer use, or even extensions like this one, instead of OpenAI building a better video editor. In the second, from Dan Shipper's idea, users open their web apps inside Codex and let Codex do more with them. Lenny takes it as another reason to be more ambitious: just try it and see whether the model figures it out.
After recording, the show's producer mentioned they also edit with Codex, for simple tasks like cutting footage into three segments at pauses in conversation. Ambrosino responds that nearly every new use starts with a story like this. The product isn't designed for the job, but it is essentially a blank chatbot that can write code. It can do anything, so the question is what it is useful for.
Failure, and not getting married to your process
In the "failure corner," Ambrosino says this may be the first time in their career they haven't felt like they were failing. They spent years as a founder in heavily regulated spaces, which felt like a constant slog, and ended up selling the company "for parts." At the next startup, building AI tools in another locked-down industry, they tried things repeatedly without success. By their count they failed for "somewhere between 10 and 15 years" before this, and they're still surprised things are going well. Even now, they say, the project to merge Codex with ChatGPT has had countless micro-failures. The team proposes a shape, posts it in Slack, and gets a 2,000-message thread about how wrong they are. Ambrosino considers that candor the reason the external product has been good.
In the lightning round, Ambrosino, a parent of young kids, recommends The Gruffalo and The Big Orange Splot. In the second, a man on a street of identical houses gets orange paint spilled on his house and decides to redo it completely. The neighbors object at first, then one by one remake their own houses. Lenny reads it as a book about agency. Ambrosino names Linear as their favorite software product before Codex. Asked which is hardest among PM, designer, and engineer, they decline to rank them. They find the constant debate over whether designers, PMs, or engineers are "cooked" overblown. The new fluidity is refreshing, especially for people with agency, but some things shouldn't go away. People should find work that matters and figure out how to do it.
The most direct advice comes after the formal recording ends. Ambrosino says they don't know what advice to give, but if there is one thing: don't get married to your exact process. Get attached to the outcomes you are uniquely able to deliver, and change your process to try things. Being the best at understanding Figma's auto layout, they note, is not a durable advantage, since AI will get better at that too. They also add a caveat. They're wary of claiming this is how things will be for everyone. People at OpenAI self-select for figuring out the next thing, while most people won't be early adopters, and relearning everything constantly is genuinely tiring. Ambrosino personally hates repetition and loves doing the job differently every day, which, they joke, is why they can't negotiate: they don't want any other job.
90% of people at OpenAI use Codex. Not 90% of engineers. That's 90% of the entire company.
Yeah. This tweet the other day where you said that you intend to make Codex the best desktop app that has ever existed.
Yeah. The quality bar for Codex had to be so high that there was never like a hesitation that you have opening this app to do the next thing, that this was your natural choice. Just like people have kind of come to open a browser tab, right?
That's true. I know there's numbers constantly coming out about the records you guys are setting for usage.
I don't know. Like we'll see. A lot of people seem to like the app.
Why do you think AI and the top frontier models are just not good at design?
I think design's a little bit harder to grade because the human aspect of taste is like part of the feedback mechanism you need. That is still feeling a little bit out of reach with the current technology.
What does the shape of product team look like now versus a couple years ago?
Everybody at OpenAI is very agentic, has great ideas, and so everybody's building everything. And it's not that people are doing fundamentally different roles or focusing on different things. It's that it's backwards. The implementation is actually not the expensive part anymore. It's, dare I say, taste.
You feel like there's this collapse coming where everyone's everything and that's just the future, or do you think we're going to continue to be mostly divided up?
There are some things that I'm afraid of. I've heard a lot of companies be like, we're getting rid of the product role and everybody's just going to be a builder, and then what happens is
Today my guest is Andrew Ambrosino, product and engineering lead for the Codex app at OpenAI. Codex is quickly becoming people's go-to app for building products and also for non-product work like organizing files in your computer, drafting documents, doing data analysis, reading your emails, and a lot more. If you stick around for the end of this episode, we actually have a little clip from after we stopped recording where the producer in the room started talking about how he uses Codex in his editing work.
Since this January, Codex usage has grown 6x. They currently have over 5 million weekly active users. I suspect this number is quickly going to be out of date. Internally at OpenAI, nearly 100% of their employees use Codex weekly. And that is not just the engineers. Andrew is a designer turned engineer turned product manager who is building the app that more and more of the world is using to build their own products.
Before we get into it, don't forget to check out lennysproductpass.com for a year free of the hottest and most well-crafted AI products in the world, available exclusively to Lenny's Newsletter subscribers. With that, I bring you Andrew Ambrosino.
Andrew, thank you so much for being here. Welcome to the podcast.
Thank you for having... This is a rare in-person podcast. I rarely do this kind of thing. We'll see how it goes.
We'll see.
We'll see, people like these more. When we were preparing for this chat, I asked you what's the biggest thing you want people to get out of this conversation, and you said that it was how AI is changing the shape of product work. You're working at maybe the most bleeding-edge, AI-pilled software team there is. So you have a really interesting lens into where things are heading, where other teams are going to be in a year or two or more. What does the shape of product team look like now versus a couple years ago?
One of the hardest things to do right now as a leader building these products is just sort of the inversion of the process in my mind, which I think a lot of people have talked about, which is that anybody can build anything, right? Like I generally believe now that starting from scratch, if you talk to these models, ours, anybody else's really, you can stand up whatever feature you want, right? And that's not necessarily a hard part of software, but that's like, that's really cool. And I think that has created an environment where people are making all of this, right? You give people unlimited tokens. Everybody at OpenAI is very agentic, has great ideas, and so everybody's building everything.
Whereas I think, you know, you look back at product process that we've all run for a long time, and it's been a little bit opposite, right? It's been kind of research, ideation, maybe there was some prototyping, but it was, you know, even when we got past waterfall, it was still kind of flavored of like, the implementation is expensive, and so what you want to do is you want to derisk all implementation up front through documents, through research, through prototypes, because prototypes and designs are cheaper, was kind of the assumption there. And that's changed. That's like totally changed. And right now I'm sure there are 90 different explorations for... there's this feature that we desperately need to do that I'm sure there are 90 different uncoordinated teams like implementing and trying, right?
So I guess the short answer is like, it's backwards. And it's not that people are doing fundamentally different roles or focusing on different things, or that even skill sets have vanished, or that roles have just disappeared. It's that it's backwards, right? The implementation is actually not the expensive part anymore. It's, dare I say, taste. But it's the curation process. It's like, of those 90 attempts, like what's good about these? What should we fold into other aspects of this, right? How should we frame this? Should it be part of this other feature, right? How many segments should be in the toggle? You know, all of those things.
This episode is brought to you by our season's presenting sponsor, WorkOS. What do OpenAI, Anthropic, Cursor, Vercel, Replit, Sierra, Clay, and hundreds of other winning companies all have in common? They are all powered by WorkOS. If you're building a product for the enterprise, you've felt the pain of integrating single sign-on, SCIM, RBAC, audit logs, and other features required by large companies. WorkOS turns those deal blockers into drop-in APIs with a modern developer platform built specifically for B2B SaaS. Literally every startup that I'm an investor in that starts to expand upmarket ends up working with WorkOS. And that's because they are the best. Whether you are a seed-stage startup trying to land your first enterprise customer or a unicorn expanding globally, WorkOS is the fastest path to becoming enterprise ready and unblocking growth. It's essentially Stripe for enterprise features. Visit workos.com to get started, or just hit up their Slack where they have actual engineers waiting to answer your questions. WorkOS allows you to build faster with delightful APIs, comprehensive docs, and a smooth developer experience. Go to workos.com to make your app enterprise ready today.
Taste, such a buzzword. I want to come back to that. This idea of 90 prototypes, so interesting. So just to make sure I understand that. So there's an idea out there floating around OpenAI. What people used to do is write docs. Here's what we're going to build. Here's the feature. Here's the strategy. Today, what you're describing, which makes all the sense, is people just create a prototype. And what you're saying is people across the company have kind of similar ideas, and now instead of a doc they create their little prototype, and that leads to kind of 90 different things people can look at and maybe pick, here's a direction we want to go down. Is that the idea?
There's a lot of this. And you know, it's not just happening here. Like you've seen many product leaders say PRDs are dead, prototypes are in, and I actually don't believe this at all. I think that one of the interesting things that is happening right now is that because implementation has gotten so cheap across every medium, it's very tempting to jump straight to a prototype, especially if you're not an engineer, right? Especially if you've never been able to write code or never been interested or never had the time. It's really tempting to say like, PRDs are dead, let me just show you what I mean, right? What I've also noticed, though, is that for engineers, it's really tempting to write a lot of documents, a lot of documents that are not worth reading. This is no shade on people writing documents. It's that if implementation is abundant, then it's really important to pick the right format for the point you're trying to make. If that point is product clarity around a vague area, then it might actually be a document. If what you're trying to do is get something in people's hands to try out and to stress test an interaction pattern, it's a prototype. But I think like this is kind of the funny thing now, which is that like it's really important to pick the medium.
There's this term that a podcast guest shared that I think about when you say this, which is, it's called the primal mark. When a designer or a painter or an artist just creates the first mark on a painting or a piece of art, that mark is what you start to respond to. And so everything kind of trickles down from that first mark you make. And what I'm hearing you saying is sometimes the prototype is the wrong first thing to do, because then you're just responding to this prototype versus a different idea, versus a bigger idea. So I love hearing this. So like everyone's just like, "Okay, forget it. No more writing, no more docs, no more PRDs." You're saying they're actually still useful for specific use cases.
Yeah. I think too there's this part of the previous world, which was that the medium implied, it had baked in a lot of signal around where in the process something was, right? So if you're seeing something that feels like the app in production, that means that it's late in the process, that assumptions have been derisked, that, you know, design has looked at this, that this is a good business goal, right? And now those things are sort of divorced, right? And the reason it was that way is because it was hard to get resources to build the thing until it was properly derisked, and now that's like just out the window, right? And so I think it's really important to start saying, look, we can have prototypes, we can have documents. Are we clear around what this is doing? Right? Because to your point, you do not want to over-anchor on this thing that was meant to be an exploration, but now it looks so production ready that like, oh, visually it's ready for prod, but it's not actually the right model of where the research is going or what users are asking for or what's right for the business, right? Not to overdo the taste thing, but it's once again the taste to know what to work on, how to present that information, how to achieve the goals, what medium to use, is emerging as like the most important thing to do. And that's it. That's in every field.
What is taste when you talk about good taste? Is it what you describe, deciding here's the thing we're going to invest in? Is it also, once you have a thing, is this right, is this the thing to ship? Talk about, when you think about what is good taste, good judgment, what does that look like concretely? Because people hear this word, they're like, "Oh, I have good taste. I know it." What does it look like in practice?
Yeah. It's funny. There was a tweet, I'm too online, there was a tweet, I think it was yesterday, from the head of product at Linear, I might be getting that wrong, sorry to anybody, who said people overemphasize the aesthetic part of what taste means. And they used Paul Graham as an example, saying Paul Graham clearly has great taste and wears cargo shorts, right? Like, you know, we've got to tease out what taste means a little bit. And there's a lot of nuance here. I think it's all of the above to what you mentioned. There is an aesthetic part to it. But there's also a systems thinking part of it. Like how does this fit in the system? There's a where are we going, and what theme is this part of? There's how to present this. A lot of it is wider context. And you know, obviously there are parts of taste that are like, okay, this interaction animation doesn't fit the semantic meaning it's supposed to, right? Like it's too snappy for what it's actually trying to convey. And that's incredibly important, and I focus probably too much on that. But there's the like, what should this be, like if we can build anything, what's the goal here and how do we get there, that I think is actually the real taste question here.
When I hear things like this, I always wonder where will human brains continue to be valuable as AI becomes stronger and better and doing more of the work, and it feels like taste is a part of it. Something I think about along these lines is just AI is still very bad at actual design. Like the output of AI is not great. Rarely is it like, this is it, they nailed it. And it's always like, oh, this is Claude design, this is Codex design. Why do you think AI and the top frontier models are just not good at design today? And do you think they'll get there? Do you think we'll get to a place of like, holy moly, we're done?
Yeah. I tend to think that there are some practical reasons why it's lagged and also some harder problems to crack. I'm not in our research, or like, I'm sure I'll get yelled at for saying this. I think design's a little bit harder to grade than software, and that, you know, creating a loop where you can train the model on what's good design and what's bad design is just a little bit more tedious and onerous than, you know, does the code compile, does it do what it's supposed to, right? Because the human aspect of taste is like part of the feedback mechanism you need. I also think that the labs historically invest in making their models good at things that accelerate AI research, and that in the early era of coding models, it was very clear that the model being able to write correct code would accelerate research, right, in a way that you can't really make the same case for design. Not that getting good at design isn't important, it's that it's not directly in that flywheel, right?
Those are practical reasons, and, you know, those will go away. Like these models will get pretty good at design. There are some kind of murkier things that are going to be really tough. I have kind of a short list of them. One is there is an aspect of culture to what is considered good design, in that, you remember, it was probably what, last year, where like every new website that came out was just a copy of Linear's website, right? Like Linear's website: great design, great taste. Like if a model did that, I'd be like, wow, this is incredible leaps here, right? If I have a model that outputs Linear's website every time, that's not the challenge here, right? There's an amount of novelty that is more important in design than it actually is in software engineering. Like software engineering, you almost want it to over-index on known patterns, right? Whereas design, it's like, no, there's an element of randomness here and novelty, right?
There's also the, you know, to me, like I spent a lot of time writing code or, you know, supervising code on the early Codex app, and even as the models get good at design, there's sort of an abstraction layer that is an interplay between the software design and the code that's being written. Like this thing over here in this corner should share X, Y, and Z in the codebase with this thing down here, right? And that's a little bit different than saying the model needs to be a better designer, especially on the, you know, that's not visual, that is visual design, but it is significantly deeper. It's about the abstractions, in that, like, oh, if tomorrow our company did a rebrand, the shallow version of this is that we have to, you know, update 263 components one by one. The deep version is like the
semantics between these two things that look different, like they're both in lists that have this style that conveys this interaction pattern to the user. And I think that is still feeling a little bit out of reach with the current technology, right, that abstraction layer. So I think, as we've gone through this process, right, we started the Codex app in November and we weren't using it full-time; now we use it for everything. That's been a journey, but now the things that we actually do while using it are different things. So what was the question?
I know, that was an amazing answer. Speaking of design and being creative, the Codex app when it came out, it's such a new thing that nobody had seen before. It's not a terminal thing. It's not an IDE thing. It's this chat thing that codes and you could see code.
Yeah.
To your point, it feels like it'd be hard for AI to be like, here's a whole new paradigm for how to code. And that feels like where human brains continue to be valuable for now, is creativity almost, and coming up with something new versus patterns of things that have been done before.
Yeah. I mean, I totally agree. Let's give it up for the human brain. [laughter]
For now. As we were getting ready for this, you said that you were listening to the episode with Jenny, who is the head of design for Claude Code and Cowork and such, and she had this whole thesis that the design process is dead. There's no time for design. Things are moving too fast.
Just build now, and design is kind of steering things as things move along. You're implying you have kind of a different perspective on the design process.
We probably agree on a lot of this, Jenny and I. I wasn't a fan of the design process proper. I agree with her take that it is dead, and I genuinely was not a fan of this process before AI. Like I—
Can you describe the process real quick, just for people to think about the design?
So I mean, when I ran a startup a number of years ago, we would do design hiring, and there was this sort of snarky article that came out about the case study factory, and it was midsurp era stuff, right? And it was that designers are being taught about this process and valuing that above all else, above all outcomes even, right? And if something went through this process, two things were true. One, it would be good, and the process would guarantee quality and guarantee impact. And also that the thing was good if it went through that process, even if you don't like it and nobody uses it. It's like the process, of user research and the divergence and the convergence, is the right framework. It was always a little academic.
But I think this is really exposing some areas where it falls down, especially because of the speed of implementation. And once again, that process is sort of predicated on the assumption that implementation is expensive and that you can really only afford to build once. And so you need to fully, exhaustively go through the problem space and the solution space before implementing, right? And then we saw with Figma and Origami and all of these tools that you can fast-forward some of the insights by pulling interactive prototypes earlier into the process, right? That you can simulate production. And there ended up being sort of a meme about executives just being like, well, can we just do a prototype and then expect it to work. But this thing was real, right? This became part of the design process proper, right? We pulled prototyping into that. The problem now is that you can pull all of the implementation into that.
And there's a mismatch between, I think, a lot of assumptions. Again, you see this fully polished prototype that looks like it's ready to go out the door, and enough people at a company see that and they're like, can we release this now? But we're actually in that early design process stage, and nobody's just saying that, right? This is where we are with a bunch of multiplayer exploration, right? Ninety people will have this idea, it'll look really polished, but it's like, no, that's the design process now, right?
Tying the design process to mediums, media, that's the scary part. It's that designers have more tools now to do this process with, right? You can put stuff into the current product and you can A/B test it or just use that as a prototype. Many companies right now have this idea of a baby version of the product, like baby Cursor. You've seen this on Twitter. We have baby Codex, right? A dramatically simplified codebase that approximates all of the interactions of the production app and therefore is a lot quicker to vibe code over, right? Because you can be like, well, what if the sidebar worked like this, or what if a pane came in and had a group chat here? What if XYZ, right? That's a huge tool that's part of the design process.
So to say the design process is dead, I feel like it's both true and false, right? If you are tied to the tools, the exact day-to-day specifics of the process, then yeah, it's dead. You're not going to have a good time. But to throw the process out completely, or throw out the overlay of the process, the "hey, we're at this point in the process," that is still more important than ever.
It's really interesting because you have a background in every function. If people look at your LinkedIn, it's engineer, designer, product manager, founder. Now you oversee the desktop app, and I think design is not under your purview. Is that right? Is there a separate design team, or are they under your—
Depends on the week.
Okay, good.
We work very closely together. We believe in all sitting together, being embedded. Like, reporting lines, I don't—
They shift weekly. What does the design process look like on the Codex team?
Yeah, there's been a lot written about role collapse, existential role collapse. There are no roles anymore. We haven't seen that. We have seen more role collapse in the Codex org than, I think, other parts of the company and other parts of the economy. I think part of this is that this was a technical product for engineers, and so our designers speak engineer, right? Our product managers speak technical language and write code. Alexander has a master's degree in computer science. I do not have a master's degree in computer science. So we've seen a lot of role collapse.
And I think that one of the ways that we describe how the groups work together is that there's significantly more overlap in the roles than there used to be. And everybody's sort of defined less by the fence and the boundaries of where design stops and engineering starts, but more by the average of where they're working, right? 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, right, if you draw it out on a diagram.
And this sort of speaks to the process too, especially because the entirety of the Codex app has been informed by the dogfooding loop. There is a desire among all of us to try to do as much as possible in the app, even when it's not the best tool, so that it can become the best tool. And so a lot of design we all work on by using the app and saying, okay, what's broken about this? This is a whole thing we do, which is that we often don't improve our process so that we can make the product better to do it, which is a deeply uncomfortable place to be in. But week to week it's changing.
I love this point so much, that your role is the average of what you spend your time on. If most of your work is PM work, then okay, you're a PM for now.
Yeah.
If it's engineering, you're an engineer for now. I feel like, was OpenAI the first company to call people member of technical staff?
No, I believe that this might have started with Xerox. The first company I interned at, a company called Up There, did the same thing.
This has been around, but it's much more common now. It is kind of a tradition in research-focused companies.
Right, okay, got it. So it emerged from research, but I feel like it's such a sign of where things might be headed, this idea of, we're just going to call everyone member of technical staff. Your function isn't set. You're not in this bucket of the PM or the org of the design. Do you feel like that's where we all head long term? Do you feel like functions will continue to exist? Like there's still the PM skill set and the eng skill set and the design skill set.
Yeah.
And people are like, I'm a designer. Or do you think, like people call it builder, do you feel like there's this collapse coming where everyone's everything and that's just the future, or do you think we're going to continue to be mostly divided up?
There are some things that I'm afraid of. And I think that some companies like to be very extreme about getting onto the bandwagon of whatever people say is going to happen. And I think part of the danger in eliminating the concept of roles is that it can dangerously eliminate the idea that things are specialties with knowable best practices, right? 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 a builder. And then what happens is this whole discipline of product that's been built up and has real best practices, real things that have been tried and failed, and real processes, that just gets abandoned because people are like, "Oh, I wrote some code," right? That's not a great place to be in.
I think the boundary of, so-and-so, this isn't your lane, I welcome that part going away. But there's a balance here where not everyone can work on everything, for one, both in terms of breadth and depth, right? This is why managers are not going to go away. Not everybody can work on everything. And also every discipline has a skill component to it, which I think a lot of engineers are guilty of not recognizing, like, well, engineering has a skill to it, it's writing code, and other roles are just people vibing. It's like, no, that's not how it works, right? Yes, you can use Excel, but you cannot work on the finance team, right? That kind of stuff.
Yeah. I think there's also just, do you want to be doing this work? Do I want to be—
I think more of it actually now is that it's easier to switch roles. It's easier to learn the best practices. It's easier to not tie your effectiveness in a role to the ability to use the exact tool, right? It's more of, can you get yourself into this mindset, learn which things work and which don't, and then focus on it, right? I spent so long feeling like I should not be a software engineer because I didn't care about assembly language or memorizing TypeScript syntax, you know? And there have always been parts of these roles that are sort of gatekeeping, that are like, well, no, being good at this role is being good at this tool. And I think that's what's kind of starting to erode. I just think people hyperbolize all of this.
What does your team look like on the Codex team? How many engineers, designers, PMs? What's the makeup of the team right now?
Every time people ask me how many people are on the Codex team, do you remember my answer to this? Yeah, I'm like, it's somewhere between 10 and a few thousand. I mean, it's a fake answer, but it's real in that we do see this as the culmination of what everybody works on here. Everything that goes into model research, everything that goes into how models are good at CUA and browser use, everything about model personality, all of the product work around front-end infrastructure, all of it is this product. At the same time, we are not accepting PRs daily from thousands and thousands of people on whatever they want.
So we've got a team of double digits of engineers, probably half that on the design side, a few product people, although product here is kind of more of a zone defense play. And I think one thing that is very common among everybody on the Codex side, or on the desktop side, is agency and taste, right? A lot of former founders, or people who were at larger companies doing founder-shaped things. A lot of people with immense taste. At OpenAI we let teams get very large, so we haven't done that "hey, there's no management," but the teams are quite large, right? It's mostly ICs, and I think that's good.
You use this term zone defense for product work, and I think that's really interesting. It kind of maps to the design shift also, just like you're there to manage and coordinate. Talk a little bit more about what that looks like. What does zone defense look like for a product person?
Yeah, and I have had a lot of conversations with Alexander about this analogy, which is that if two product people are working too closely, that's often not a good signal. And as a product org you sort of want to do this force-directed activity where you're like, where are the gaps? Especially in this new world where curation and steering and alignment are a lot of it, where there's a ton of chaos happening, people throwing ideas all over the place, right? The whole top-down, year-long planning thing is not going to work. And so now it's like, we need the tastemakers to guide things from inception to what the product should be. And that means you basically want company coverage. And so you spread out and you say, "All right, who's best at what? Let's create some space between us so that we've got full coverage," right? And that's kind of how it goes.
And then you fill in the gaps and you're like, look, we want to hire engineers who are product-minded. We don't want it to be that we've got a bunch of people writing a bunch of code that needs the full team reviewing it for product coherence, right? We want everyone to have these skills, but I think what people go deep on has to change, right?
This is definitely a thread I've been noticing over and over talking to folks like you: one of the most valuable people right now is someone who could take an idea from idea to done with the taste to know this is great. Just shepherding it throughout, this obsession with making it awesome, this kind of high-agency, high-taste person, exactly as you described. Is that kind of the way you think about here's who we're hiring, here's who's going to do really well in this new world?
Yeah, I think that's the core piece right now. 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, right? If you're an IC, you're not typing code out character by character, right? You are managing something. You're managing agents, you're managing work that is happening, right, 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, right? I generally look for, obviously, command over the discipline, but then the taste to say, hey, you're going to have unlimited
tokens, and we can't just be doing slop. You need to be able to determine what's signal, what's noise, in a world of just infinite content.
You mentioned planning. At the pace things are moving, it's become very hard to plan roadmaps. I imagine especially in your world. People are very frustrated with me all the time on this, because things are just constantly shipping, things are changing, right? How do you plan on your team? How far ahead are you thinking, and what does a plan look like? Is it a spreadsheet? Is it an MD file? What's kind of the output of a plan?
Yeah, I don't think we do anything revolutionary on that. We're not clever about planning. I think the basic gist is the shorter term something is, the more detail it needs. And it's not that we don't plan for 9 months out, it's that that just has to stay very hazy, because any amount of precision that you add to a 9-month plan right now is false precision, and you're just going to waste time. You can say stuff, right? But nothing that we planned... I think research is different, so I'm not speaking for research here, but on the applied side, when we do product, anything that you could have planned in November may have been true for December, but isn't what happened, right?
So it's hard. It is really hard to do planning. We generally need to know what do we think models are able to do on what timeline. At my last company, I kind of saw this shift where we were starting to use the models to drive features, and the product process fell down. It basically had to be like, let's list out all of the things that we think we are interested in doing for the next year or two. Let's prototype all of them, decide which things are ready now, and then just let the others sit and bake, and then every time there's a new leap in models, let's try that thing again with it swapped out. Because the whole premise of whether features were good or not was based on whether they were smart enough, not the shape of them.
So this is a great story about the Codex app. 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 that the only difference was the models between November and February, right? And I think there's a lot to that, that this product with the exact same shape, its outcomes were totally different depending on just a few months of timing.
This episode is brought to you by Mercury. Radically different banking loved by over 300,000 entrepreneurs, and now with Command. I've been a customer of Mercury's for over 6 years. I have never once thought about leaving. Mercury is basically what happens when banking is built by product people, not by bankers. They make it so easy, dare I say fun, to send invoices, move money around, set up virtual cards for folks on my team. Does your bank have an API, a terminal-native CLI, or an AI-ready MCP server? I don't think so. And just recently, they launched Command, a conversational interface built directly into Mercury, which acts as your financial operator. I've been using Command to transfer money around, to figure out what categories I've been spending the most money in, analyze my cash flows, and just today I used it to find out how much I've made from a specific sponsor over the past year. I just ask, "How much have I made from X over the past year?" 10 seconds later, I have an answer. It is so freaking cool. Visit mercury.com to learn more and apply online in minutes. Mercury is a fintech company, not an FDIC-insured bank. Banking services provided through Choice Financial Group and Column N.A., members FDIC.
This is definitely a thread on this podcast: build things that are not yet working that will work when the model gets better. And there's this other thread of ambition: be more ambitious with the things you take on. So is this just a way you approach things, like, let's just build a bunch of things that may not work yet, we'll just have them around and wait for a model to catch up? Is that kind of the approach?
Yeah, I think we have a lot of that. I think sometimes the challenge is you have to be very clear, again, about what stage of the design process that's in. People still have this muscle memory of, oh, I wrote the code for this thing, therefore we should put it out there. It's like, no, no, no, that means you have an artifact now that we can test against for future models, right? This happened with the in-app browser in the app that we have, right? We had a kind of working version. I mean, go back to Atlas, we had agent working inside of Atlas, and that was pretty cool. We had Operator before that in ChatGPT, right? That didn't work out. Very cool idea. There's some thread that you can draw between Operator, Atlas, Codex, ChatGPT that's fundamentally the same feature, but the re-releasing of it with different intelligence totally changes the outcome here. And so I push people not to be stubborn about, no, this isn't working, so it's a bad feature. No, it might not be ready yet.
There's also this aspect of, especially in research, there's always a desire to be the most ambitious and to say, okay, but at the limit the model can just do this, and it just doesn't work on the product side. If you go back to the original Codex release, basically what it was is Codex web, and it wasn't good for interacting with. It was like, you give the model a task and it's going to go off, do the task, come back to you finished. Doesn't sound that radical. The problem is it didn't do the task that well. It wrote code, it was good, but that form factor was too early. And then Claude Code comes out, totally local, not hooked up to the cloud. It doesn't pretend to be as... it's not as AGI-pilled, right? It's going to ask you questions, it's going to sit there, you can't just delegate your life to it. That worked way better, right, because that's the point that the models were at there, right? So we were too AGI-pilled for the moment. And I think about that lesson a lot on this stuff. It used to be that, you know, the market told you all these things about the shape of the product, about the communication of the product, and now it's like, no, you might need to release this thing six different times before it works, and the shape might not change at all.
It's so interesting to hear about all the variables you have to think about building product now. There's the timeline for the models and the research and how smart it gets. There's people's ability to even understand, this is how you could build software in the cloud and this is the future, like get people prepared for this new future. And then just what you can build as a team. And I love that Codex example, because it comes back to this idea of ambition, and I want to hear if there's anything there for you, this thread of just be more ambitious, because these models can do so much more than you even imagine, and sometimes it's too ambitious for the market and they're not ready for it. Do you think about that at all, just pushing your team to be more ambitious, because it's so much easier to just do things that may have felt crazy hard in the past?
Yeah, this is a core challenge. Once there's a product that exists or a feature that exists, it's really easy for people to find paper cuts and optimize. And they should, and people on Twitter like to remind us of this, and I thank them for that. People should be focused on the features that exist and making them more reliable and better. But this is why we also have a culture of bottoms-up exploration here, because sometimes, in the same way the Codex app came and disrupted ChatGPT in some way, right, this thing will get disrupted by a future effort, and that's part of the design. You can't always as one team be good at both the disruptive piece and the maintaining a product and its quality piece. So the point is to design a process that allows for both.
Kind of zooming out a little bit, if you think about the progression that we've been on of AI impacting how we build product, it's insane how far we've come from, as you said, we used to write all our code by hand, artisanally created human code, to AI writing 100% of our code, to, you actually put it this way, that now coding is steering the AI. And when you think about what percentage of my code is written by AI, it's almost like how many times did I have to steer it in the right direction is the AI version of coding. And now that there's agents and loops and all these things, what's kind of the latest frontier from what you've seen of how people are building? Is it loops? Is there something else, of just like the most AI-forward teams, here's how they operate now, that people may not be aware of?
Yeah, I mean, loops are so last week, man. We talked about this. One of the big questions is always, well, how much of the product is AI-written? And it's always hard to answer that question, because if you're using the goalposts from last year, it's like, well, 100% of our product right now is AI-written code. So the question is more like, well, okay, fine, is the code written supervised versus unsupervised, right? And that's a totally different thing. I welcome the moving of goalposts, because that means we're making product progress. There have been a lot of explorations here around autonomously developed software, a lot of harness engineering stuff, a lot of different explorations. Like, okay, well, what if you came in overnight and did garbage collection of the codebase to clean it up, right? One thing that I think all models suffer with right now is they usually increase complexity. If research is listening at any company, please make the models better at deleting code. But that becomes a problem right now when you try to put development completely on autopilot. And it's both on the human side and the codebase side. So feature requests, right? How do you teach a model which features to build, which ones to ignore, which ones to kind of group together and reframe a little bit? How do you teach a model how to build the right abstractions, right? All of this is getting better. I don't think we're at a place yet where we're like, we've just got to set up a loop that's like, improve the app, and listens to Twitter and listens to Slack and listens to email. We're not there yet, but we're trying to make it happen.
Do you think we'll get there? Do you think we'll get to a place where it's just like, grow, win, /goal make money, make me a billion dollars? Win the market.
I don't know, man. I am not in the business of saying never or always or whatever.
Yeah. How are you using AI in your work as a product leader, engineering leader? What are some ways that you use it that maybe people may not be aware they can use the app for?
Yeah, I think I have the best job in the world right now. But one of the things that makes it very fun is that when we were developing the original Codex app, the goal for me personally was to make it the thing that I wrote the code with, right? I was like, I need to make this so good at development that I can build the Codex app with this. And the Codex app at that time was a development tool, right? And we did that super quick dogfooding loop, because you've got your personal dogfooding loop where you're like, oh, I can't do this thing, I should fix that so that I can do the thing. Now I can do the thing. Now I can do more things, right? We released that, and then the next challenge was, hey, people are starting to do some different-shaped things with this, right? And now I need to grow this, and so I need to hire a few people and help. So then my role changed at the same time that the role of the app needed to change. So I'm like, okay, I need to do more product discovery here. I need to figure out the right loops for seeing what everybody's working on and steering things that are off track. And so all of a sudden that's what I started using the Codex app for, right? I did still write code. I've tried to align my own usage of it with the problem that we're trying to solve, right? And now I'm like, I need to build a spreadsheet that models this out. I need to do an internal deep research on all of the efforts that have gone into this area of research for the next version of this.
There's a release, or a series of releases, in May-ish that introduced the in-app browser, computer use, and artifact creation to the Codex app. That was, I think, our Codex-for-almost-everyone release. And everybody knows the term vibe coding; I think that was our first vibe-coordinated release, where I had a Notion doc somewhere with everything that needed to happen, and I was automating going out to gather updates from pull requests, from Slack channels, and updating the status tracker. Now this is pretty commonplace, but at the time I felt like I was at the bleeding edge of how to manage a product release. In short, the way that I use the Codex app is basically: what has my job grown into, and how do I make this thing able to do everything I need to do? I will get up in the morning, I will see the daily brief that I have from everything from the 3,000 Slack channels that I'm in, like which things need my attention. I can kind of message back and be like, "All right, give me five questions and I'll answer them," and I can do that.
How do you set that up? What's the workflow for somebody to set that up, because that sounds amazing.
Again, I think we're still in the discovery phase on a lot of this. And so right now it's like I'm making an automation, or scheduled task, that says, go through my Slack channels, these are the things that I care about and think are most important. So I'm still kind of defining that: these are things to watch out for, different categories, here's some context. And that'll get set up as an automated task, and then the first few times it runs it might need some steering. And luckily with this app, I don't have to find out how to edit the instructions. I can just be like, hey, next time this runs, can you please worry about this instead, or can you deemphasize this workstream, or, hey, this thing happened and it didn't come up in a brief, can you make sure that stuff in that shape comes up. So I can kind of coach it along the way. It'll update the way that it notifies me, stuff like that.
Amazing.
I think in the future, though, this has been a core problem with the chatbot shape, right? I know how to set this up. I have time to set this up, because for me it's product discovery to set it up. But if you're not working at OpenAI, not developing this, you don't want to have to figure out all this stuff. We need to figure out that shape of things.
Yeah. What I'm hearing is, I don't think people realize that your app can act a lot like OpenClaw. People were so excited about, you just talk to it, set up this thing, check on this thing for me every day and then tell me what's going on. It's starting to become a part of all the products, which is amazing. So the way somebody would set this up is they just talk within the app and say, I want to set up an automation to do this, look at my Slack, and these are the things I want to do. Great.
Yeah. It'll set it up for you. If it doesn't have a Slack connector, it'll say, "Can I add the Slack connector?"
Yes or no? You can hit yes to that. Like, the least that we can do is make it so that if you don't know how to do something in the app, they can just ask it, right? Yeah.
I don't think that's enough, but I think that's the least we can do.
Yeah. A good example is I built this little app that filters spammy email for my inbox. So every email that comes in, and I built this in Codex, every email that comes in, it looks at it and decides, is this unsolicited kind of cold email stuff that I don't want to look at, and labels it and puts it somewhere else. And to set that up, one of the steps was you have to go into like the Google Cloud console and set up all these Pub/Sub API things and triggers. I don't know if you've ever used that interface. It's like so annoying and slow, right? So I was like, wait, what if I ask you to do it? And I was like, okay, cool. Do this for me. And it just, you described computer use. I've never actually seen this happen on my computer before. It just takes over my computer and starts going there.
It's like, I don't care if you don't have a connector, man. I'll just start clicking.
Yeah. And it figures it out. It's crazy just to like watch it doing its thing.
Designing the decision boundary between connectors, when to use the in-app browser versus your Chrome extension that's connected versus computer use, is interesting and all done through just like feeling it out. I saw a great Twitter thread the other day where they describe all these three and what you use it for.
Yeah. So, this person described it really well.
These personal workflows are really interesting because some of them really click. Like, you know, people are trying all sorts of stuff. Everybody's making these personal systems. You ask everybody here what they do and everything's going to be different, and then certain themes arise and we're like, you know what, that should be a first-class experience in the app. Like, we should take this thing that everybody seems to be setting up and just make that work. And I think memory is sort of in the shape where we've had a lot of people, and a lot of people at other companies too, are like, well, I set up an Obsidian base or a Notion area and I tell it how to basically build my mind palace and how to put, it's like, I don't know if everyone, like, you shouldn't have to do that. Like, there should be a memory feature that does that for you, right? That's pretty generic. So that's one, but then there are other things like, they're your process of your job, and there's something that, like, yeah, you should set that up. But I think we're constantly wading through what's working for individual people, like what should enter the product versus stay, like, no, that's just how you do your job, right? What should become a primitive, what shouldn't.
Yeah, this is the taste and judgment you spoke of earlier, you know, citing these things. I want to talk about this browser use piece a little bit because I think people don't realize how powerful this is and what it could be used for. Reminds me, I don't know if you watched when Dan Shipper was on the podcast, from Every, he had this prediction that we're going to start using Codex to run our SaaS apps inside.
Yeah.
So, instead of going in Chrome,
I know, he Slacks me about this every day asking for stuff.
Do you feel like this is where things go, where we're just working within the Codex app using Notion and Linear and Salesforce inside, with your agent kind of helping you along, or do you think that's kind of a different direction?
It's been, yeah, it's been really interesting because obviously we've had a few attempts at browser-shaped activity, right? Operator and ChatGPT agent mode and Atlas, and now we have the in-app browser inside of the desktop app. We also have the ability to install a Chrome extension where the app connects to Chrome. Like, we've had a lot of shapes at this, and I think we've learned a lot of different things. There's a lot at play. There's a lot of really boring things at play. Like, you know, we originally launched the app. It's an Electron app. The things that you can do with in-app browsers there, it's kind of janky. So we had the in-app browser for development. It was for testing your front end in development. And we were like, it's not really for anything else, guys. Like, it's a developer tool, right? And then we switched over to our OWL stack, which had powered the Atlas browser, and so now, you know, multi-tab, and we've got enterprise security so that you can actually log into all your websites if you're, you know. So we've been iterating on this. I think the tough thing has always been what should the shape of this browser be? Like, is this something that is only for the agent, right? That, like, you've got Chrome, you open Chrome, you do your thing in Chrome; if you ask the desktop app, it opens up this browser that it can control really quickly, doesn't have the latency of Playwright, whatever. Or are we trying to say this app is for everything and we want you to use this as a browser? And those have a lot of trade-offs. It's not a super, you know, well-traveled path, right? Like, most browsers are browsers at the top level that have got browser tabs. This creates a lot of really boring but tedious problems, like keyboard shortcuts, right? Are we trying to do key mapping to VS Code or to Chrome or to our own thing or to Linear? Like, you know, we want to have some sort of muscle memory that carries over, but we've got all these things that have shape, like subshapes of different products out on the market. What do we do? And this just highlights how extra challenging this app is, where you have to allow it to work for somebody that's never built anything
from the more basic user to, like, power, like Peter,
OpenClaw, trying to code with it.
I'm not convinced I'm going to get Peter to use the app. I think he might be the last terminal holdout. But I'm going to keep trying.
Okay. Let me zoom out for a moment and talk about kind of the big picture of where you're taking all this. What's kind of the vision for Codex? Where does this go? What's it going to look like? You know, I don't know, a year or two, 10 years.
We had Codex as a CLI, right? And then we decided to build this app. And, you know, we were a little uncertain about the app, but had a lot of conviction in what it could be, as, to start, a developer tool, right? Yeah. And it wasn't going to be an IDE. It was going to be this right-sized surface where it was sort of a chatbot, but it was more than that, and you could see the code, but we weren't going to let you edit the code. There's a really interesting thing that happened at OpenAI in January and February, and it was before we actually released the Codex app, which is that we'd started to dogfood the Codex app. And what we were finding, like, we were converging on some pretty clear internal PMF on engineering, right, and research workflows. They were thrilled. They were loving it. We were like, all right, we just got to get the quality bar up before we release it to the world. We're convinced that this will be a thing. But then at the company we spun up a few other workstreams to say, hey, this Codex effort is onto something with these coding agents, and we have people from marketing, from comms, from finance, from legal, from basically every discipline who are using this Codex app even though it is actively hostile to these people, right? It is trying to show them code, it's trying to ask for approval to run rg on the, you know, it's doing all of these things that are actively not the right product surface for them. So why don't we take our other surfaces and add Codex to them? Like, let's add it to the ChatGPT desktop app. Let's add it to the Atlas browser, right? And let's essentially take the lessons of Codex and make it more general for a general knowledge work tool, right? And those efforts went for a little bit, and the most annoying problem happened, which is nobody would leave the Codex app for the apps that were allegedly for these other personas. And I think the lesson in all of this was just that the whole developer tool versus general knowledge work tool, like, there's a lot of nuance here that isn't just one or the other. And I think we believe really strongly in this, and that there are certainly, in the same way that we talk about the average of your role is what your role is now, this is true on the product side too. People who are doing Excel work don't want to see git repository information. We know that. But we also know we can tell a lot from what they're doing about what kind of work they do, and we can start simple and grow the product complex as we feel is needed, right? Doesn't mean we don't have modes, right? You might want some modes for organizing your stuff and to sort of be legible about the ways that you enter the experience, right? But we really believe strongly that what we've built here is the right shape to take on really deep, vertically focused things, right? We work deeply with our finance team, with our team working on science, team working on legal, right? And we say, like, if we can build the right extensibility primitives in the right general model, then you can do anything with this, right? And then our challenge is, well, how do you generalize it? But this is kind of going back to the best desktop app that we can build, like, what does that look like? And so, you know, was Codex the developer tool, ChatGPT, like, where does this go? This is how we think about it.
It is so interesting, the point you made that the Codex app was so, like, you did such a good job getting people to be aware it existed, and so good to use and fun to use, that everyone's starting to use that versus the ChatGPT app. So clearly the direction is combine them so that you're not creating this confusion, which I know is, you know, things people have been talking about, this idea of bringing them together.
Somebody called it a super app, and I wish they hadn't said that, because now I have to hear about the super app all day every day. We'll get past it. Great.
Okay. But is that kind of the, let's not call it a super app, but the idea is like one place people go to do all the things? Is that the general idea, or TBD?
Yeah. I think what we see here is that it's a great home base. It's a great place to keep track of all of the things that you have to do across different surfaces, and some of those things you do all of it in the app. Some of those things the app opens other apps to do, right? The app can connect to Excel so that, you know, yes, it has a spreadsheet editor inside the app. Is that good enough for people doing financial modeling at OpenAI for raising billions of dollars? Probably not. And so the app talks directly to the add-in in Microsoft Excel on your desktop. When it's done, you can close Excel, right? And so it's not just about, hey, we're drawing a rectangle on the screen and everything needs to happen in that rectangle. It's, this thing should be a home for you where you start work, you end work, you automate work, and it uses whatever you need to do, right?
There's a great story about, we had some videos that we shot in this room for the original launch of the Codex app, and our in-house DX videographer Brent then was tasked with editing all these videos, right? And he edited all the videos with Codex, which was one of the early, like, whoa, what are people doing with this thing, right? And the process for why he decided to start using Codex was really interesting. He started just because he was curious if Codex could edit videos. And so Codex is not a video editor per se, right? It doesn't have any of that UI in it, but it was able to understand that he used Premiere Pro. It could do some edits by editing the files that were backing what was on screen in Premiere Pro, but it couldn't do everything. So naturally, what Codex then did was built itself an extension that could be installed into Premiere Pro that it could then talk to and say, "Hey, Premiere Pro extension, can you please change this marker inside of the Premiere Pro app." That was pretty nuts when we saw that happening.
It's a great model, right? There are these specialty tools that specialize in things, and so we're trying to do two things at once with Codex and now with ChatGPT. One is how can we seamlessly interact with these tools that you're already using and say, like, we don't need to build a better video editor for you, right? But Codex and ChatGPT can use that video editor, right? It can interact, can hand stuff off to it, right? So how can we do that? And that's often through connectors or computer use or even extensions in this case, right? And then there is, you know, Dan Shipper's thing, which is, hey, I have these web apps that you can click around and use, but I want to be able to open these in Codex and have Codex do extra stuff with it, right? And so those are kind of like two models that are almost inverse of each other that we're doing a lot with both at the same time.
This Premiere story is interesting to me because it's another example of just be more ambitious with these AI jobs. Like, you may not know, maybe they could do this thing. It's almost just like, go try, go try it out, see if it figures it out. I'm going to take us to a recurring corner on the podcast that I call failure corner. And so the question for you is, people see people like you just killing it, just growing. Everything's winning. Codex is doing so great. This crazy career, everything's up and to the right. People may not see the times that things didn't work out and things that you launched that were failures. And so these stories are really important for people to hear, that it's not all just a win all the time. What's a story of a time you failed in your career that taught you something really important?
It's funny to hear that description played back at me, and like, this is perhaps the first time I've not felt like I was failing. I mean, I was a startup founder for a long time. I ended up selling the company for parts, essentially, right? And it was just, it was years. It was a slog. It was heavily regulated spaces. The whole thing felt like a constant failure. I went to this other startup and we were trying to do some AI tools in this also pretty locked-down, regulated industry, and that felt like just time after time of trying things and it not working. So to me it's been like, I've failed actually quite a lot, and you know, sometimes it's just a point in time where things line up, like skill set, passion, point in the market line up. We know with this project to bring what we've learned with the Codex app and marry it with ChatGPT, there have been I don't know how many micro failures on this, where we're like, this is the shape it should look like, and then throw that in Slack, and there's like a 2,000-message thread about how stupid we are. And it's like, this is the thing I love about OpenAI, is people will just tell us that, right? There's no holding back on when we fail with product things internally. It's why the external product has been pretty great, is because it goes through these cycles of like
2,000 "this sucks."
I failed for like, I don't know, somewhere between 10 and 15 years before getting to this point, so I'm still surprised every day that things are going well. Like, and I know it, I like, but
I think this is really important for people to hear, that you can have a lot of things not work out and then things start to work out super well, and it's just keep going and keep learning, I imagine, is a lesson. Well, with that, we've reached our very exciting lightning round. I've got five questions for you.
Great. Here we go. What are two or three books that you find yourself recommending most to other people?
See, man, I'm a parent now. I'm a parent of young kids, so I'm like, I don't know. There's one called The Gruff
to my kids. Oh my god. Our kid just got obsessed with The Gruffalo.
Yes.
We have like a bedtime chart and now it's like pajamas, brush teeth, books, Gruffalo, on blankie.
Yeah.
Yeah. It's so good.
So other books I'm reading right now are like that style. I'm like, actually The Gruffalo is not a terrible one. I feel like there's some less.
So sweet. Yeah. Although every kids book is about death. Like someone's eating someone, someone's killing. There's always like bad, like murder.
Yeah. Yeah. And destruction.
Like violence.
Even when it doesn't feel like violence to them, the words, there's like...
It creates like an arc and excitement.
Yeah.
Yeah. Okay. Great choice, Gruffalo.
I currently have a book backlog. I need to read all of them.
Any other children's books that your kid likes?
Okay. So, yes, actually, I am well versed on children's space. My favorite children's book ever is a very old one and it's called The Big Orange Splot or something along those lines. Look it up. It's great. Go get it. If you hate HOAs, go get it. It's about this guy, Mr. Plumbean, who lives on a street where all the houses are the same. It's a very neat street, and then one day a bird drops a large can of orange paint on his house and he says, "F it." Like, I'm going all in. So he goes to the store, he gets paint, hammocks, alligators, and he totally redoes his house, and the neighbors, they're up in arms, right? Like, you know, HOA, property value, whatever, whatever. It's not about HOAs, but I'm a very anti-HOA person. It's just really cool. And then one by one, the neighbors go talk to him, have what they refer to as lemonade, but it's a very convincing lemonade because one by one, they all start redoing their houses. Like one guy does like a boat, right? I think that's a good book. I think people need to read this, especially, you know, newbies.
What I hear from this is agency.
Agency. Exactly. You can just do things.
Just do things. Amazing. Okay, we'll keep going. Favorite recent movie or TV show you really enjoyed, if you've had any time?
So, The Magic School Bus is back on Netflix. It's a new animated one. It's got Kate McKinnon, because Miss Frizzle is now Professor Frizzle and she's, you know, around, but she's not the main Frizzle now. Kate McKinnon plays the main Miss Frizzle. Yeah, I always liked The Magic School Bus.
I've never used it.
Personally, I don't have time for movies. So, what I do is watch hour-long Netflix things back to back. You know that thing that people do?
Oh, I couldn't sit down and watch a whole movie, and then watch hour-long episodes and keep... Yeah.
I did. There's something addicting about that.
Favorite product you've recently discovered that you really love?
What would be a terrible answer? I feel like I'm discovering our product every day.
Beautiful.
That's so...
I think Linear. Linear does a great job. Like, until this, Linear was my favorite, at least software product.
Is that what you guys use to plan in?
Well, in theory.
Do you have a favorite life motto that you find yourself coming back to in work or in life?
I want to ask everyone who works with me this, because I feel like I'm not a motto person, and then people tell me stuff I say all the time.
Yeah. Like when we were chatting ahead of this, there's so many little nuggets that stood out to me that I've integrated into this chat. So I totally hear that. Okay, last question. You've been a PM, you've been a designer, you've been an engineer. Which is the toughest role of the three? Which is like the hard... to start a war?
[laughter]
Yes. I know. I think they're all very different. The things that make it tough for one person make it easy for others. There's so many takes on this and on this triad. What's going to happen with this triad? Like, designers are done, or should designers code? Are PMs cooked? Do we not need engineers anymore because PMs are going to write all the code? Or are designers going to PM now? And everybody's cooked and everybody's so back, and I don't know. There's some convergence. There's some fluidity that is being introduced that I think is refreshing and great, especially for people with agency that want to be able to just do the thing that needs to get done. And at the same time, like we talked about, there are some things that shouldn't go away. But I think people should find the stuff that's worth working on and go figure out what to do on those things.
That is a beautiful way to end it. Andrew, thank you so much for being here.
Thank you. Bye everyone.
Thank you so much for listening. If you found this valuable, you can subscribe to the show on Apple Podcasts, Spotify, or your favorite podcast app. Also, please consider giving us a rating or leaving a review, as that really helps other listeners find the podcast. You can find all past episodes or learn more about the show at lennyspodcast.com. See you in the next episode.
I love that thing about Brent in the Premiere.
Yeah.
That's cool. I've actually used it for edits as well. It's basically...
Simple stuff like, oh, could you just cut this into the three breaks, you know, like if there's a pause in conversation like the second, and Codex understands it.
Basically, every job, we feel like, starts with a story like this, which is the product's not designed for it, but it's sort of a blank chatbot that can write code. So it can do everything, but like what...
What are the useful things for it to do?
Right. I mean, people just have to have curiosity about the project. They have to...
You know, have an intentional outcome and use Codex as the platform for that, just to see what happens.
You know, because there's no risk at all.
Right. Just a few tokens.
Few tokens.
But of course, if you're working for OpenAI, there's less risk in that regard.
Yeah. You asked...
One of them, or a lot of them actually, like what skills are important, and then you've also had conversations about the cracked new grad versus the...
Like...
I don't know if you're married to the exact process you have right now, like that. I don't know what advice to ever give, but if there's one piece, it's like, do not get married to your exact process. Get married to the outcomes that you are uniquely able to deliver.
And then do things like change your process to try things. Like...
You just keep...
"I'm the best at understanding Figma auto layout." Like...
Right.
What are you doing?
Right? 'Cause AI is going to be better at that also.
You just keep spouting interesting things.
Keep it in. [snorts] It's crazy the level of self-awareness that's required to be successful with AI.
It is.
Yeah.
It's also why I'm nervous to ever say, like, this is how something's going to be, 'cause I think about my parents. They're open-minded people. They are into their careers, but the stuff that works here is just not going to work with everybody. There's no nice way of saying that.
You know, but the people who are here are self-selecting for, like, oh, I'm somebody who just figures out the next thing to do, and that's just not... most of the population will not be an early adopter of things.
And there's also just, it's kind of a bummer to have to relearn things all the time. Yeah.
You know, it's like, god damn it...
Just to learn a new thing again.
Yeah. Like, I hate repetition. This is like a me thing. This is why I'm not ever the best media person either, 'cause I just hate repeating myself.
Perfect.
Sucks to be a founder when you hate repeating yourself, 'cause you have to...
Repeater-in-chief.
And so to me, I'm like, if I can come in and do my job in a different way every day, love it.
But like, that's not...
You found product-market fit for your job.
Yes. Yeah. That's why I can't negotiate, because I'm like, well, I don't want other jobs. [laughter]
Amazing, man.
Article published · Updated