YouSaid · the spoken record

Alexander Embiricos

lines on the record: 106
first: 2025-12-14
most recent: 2025-12-14
sittings or episodes: 1
sources: podcast

Every line below is reproduced as it was said and linked to the record it came from. Nothing here is summarised or generated. Directory · Search · Corrections

“In a bunch of functions now, but I guess since you asked about product specifically, now answering questions much, much easier. You can just ask codecs for thoughts on that. A lot of PM type work, understanding what's changing, again, just ask codecs for help with that. Prototyping is often faster than writing specs. This is something that a lot of people have talked about. I think something that I don't think is super surprising, but something that's slightly surprising is like we see we're mostly building codecs for to write code that's going to be deployed to production, but actually we see a lot of throwaway code written with codecs now. That's kind of going back to this idea of like, you know, ubiquitous code. So you'll see someone wants to do an analysis. If I want to understand something, it's like, okay, just give codecs a bunch of data, but then ask it to build an interactive data viewer for this data, right? That's just too annoying to do in the past, but now it's just totally worth the time of just getting an agent to go do something.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah, I mean, I think mostly I just feel like much more empowered. I've always been sort of more technical leaning PM, and especially when I'm working on products for engineers, I feel like it's necessary to dog food the product. But even beyond that, I just feel like I can do much, much more as a PM. And Scott Beltsky talks about this idea of compressing the talent stack. I'm not sure if I've phrased that right. But it's basically this idea that maybe the boundaries between these roles are a little bit less needed than before because people can just do much more. And every time you someone can do more, you can skip one communication boundary and make the team that much more efficient. So I think we see it.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“In my mind, if we wanted to get to something like that, a friend you were talking about world, I think we really need to figure out how to get people to configure their coding agents to be much more autonomous on those later stages of the work.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah. So, like, we see this now. Like, we have a Slack integration for codecs. People love, you know, if there's something that you need to do quickly, people just like mention codecs. Like, why do you think this bug is happening, right? It doesn't have to be an engineer, even like maybe data scientists often here are using codecs a ton to just like answer questions like why do you think this metric moved what happened so questions you know you get the answer right back in slack it's amazing super useful but when it's as for when it's writing code then you have to go back and look at the code right and so The real like I think bottleneck right now is like validating that the code worked and like writing code review.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“That's super interesting. And I vent you the so if we go, if we went and asked them what the bottleneck to that productivity is, did they share what it is?”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“And so, in my mind, this is like a really interesting thing for us as open AI as we're building. So, for instance, when I think about launching a browser, which we did with Atlas, right? Like in my mind, one of the really interesting things we can then do is we can then contextually surface ways that we can help you as you're going about your day, right? And so we break out of this like, you know, we're just looking at code or we're just in your terminal into this idea that like, hey, like a real teammate is dealing with a lot more than just code, right? They're dealing with a lot of things that are web content. So like, you know, how can we help you with that?”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“When it can surface ideas for helping you really rapidly, when it's right, you're accelerated. When it's wrong, it's not like that annoying. It can be annoying, but it's not that annoying. And so you can create this like mixed initiative system that's like contextually responding to what you're attempting to do.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“To be clear, we're not building that, but like, you know, it's a fun idea. I mean, in this example, though, like one of the things that it's doing is it's consuming external signals, right? I think the other really interesting thing is if we think about what is the most successful AI product to date. I would argue, it's funny actually not to confuse things at all, but like the first time we used the brand codecs at OpenAI was actually the model powering GitHub co-pilot. This is like way back in the day, years ago. And so we decided to reuse that brand recently because it's just so good, you know, Codex, Codex execution. But I think actually auto completion and IDEs is like one of the most successful AI products to date. And part of what's so magical about it is that”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“And one terrible idea for how it could look is that it's actually there's a mobile app. And every idea that the agent has to do is just like vertical video on your phone. And then you can swipe left if you think it's a bad idea and you can like swipe right if it's a good idea. And like you can press and hold and like speak to your phone if you want to give feedback on the idea before you swipe. Know, and in this world, like basically what your job is just to plug in this app into every single signal system, you know, system of record, and then you just sort of sit back and swipe, I don't know.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“It's just like stuff is happening on social media and like in your team communications tools. And then as a result, like code gets written and deployed. So yeah, I think I'm a little bit more oriented in that way of, you know, I don't even necessarily want to have to write a spec. Like sometimes I want to, only if I like writing specs, right? Other times I might just want to say like, hey, here's the customer service channel and tell me what's interesting to know. But if it's a small bug, just fix it. I don't want to have to write a spec for that. Have this Hypothetical future that I like to share sometimes with people as a provocation, which is like in a world where we have like truly amazing agents, like what does it look like to be a solo entrepreneur?”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“I think spectrum development is like an interesting idea. It's not clear to me that it'll work out that way because a lot of people don't write writing specs either. But it seems plausible that some people work that way. Bit of a joke idea, though, is like if you think of the way that many teams work today, they often don't necessarily have specs, but the team is just really self-driven. And so stuff just gets done. And so almost that is like, I'm coming up with this on the spot. So it's not a good name, but like chattered driven development.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah, I mean, I think there's constantly these levels of abstraction, and they're actually already played out today, right? Like today, like coding agents, mostly it's like prompt to patch, right? We're starting to see people doing spectraven development or like planned and driven development. That's actually one of the ways when people ask, like, hey, how do you run codecs on a really long task? Well, it's like often collaborate with it first to write like a planned.md, like a markdown file that's your plan. And once you're happy with that, then you ask it to go off and do work. And if that plan has verifiable steps, it'll like work for much longer. So we're totally seeing that.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“And it gets all the way down into like micro decisions if you're gonna have an agent capability to validate work and let's say you have like, I'm thinking of Codex Web right now, like you have a pain that sort of reflects the work the agent did, what do you see first? Do you see the diff or do you see the image preview of the code it wrote? And I think if you're thinking about this from perspective, like how do I empower the human? How do I make them feel like as accelerated as possible? You obviously see the image first, right? You shouldn't be reviewing the code unless first you've seen the image, unless maybe it's been like reviewed by an AI and now it's time for you to take a look.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Plays out all the time in like a ton of micro decisions. And so we as a product team are always thinking about, okay, how do we make this more fun? How do we make you feel more empowered? Where is this not working? And I would argue that like reviewing agent written code is like a place that today is less fun. And so, you know, then I think, okay, what can we do about that? Well, we can ship a code review feature that helps you build confidence in the I written code. Okay, cool. Another thing we could do is we can make it so that the agents are better able to validate its work.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“And so there's just going to be a ton more need for people with this, like humans with this competency. So that's my view. I think this is quite a complex topic. So it's something we talk about a lot and we have to kind of see how it pans out. But I think what we can do basically as a product team building in the space is just try to always think about how are we building a tool so that it feels like we're like maximally accelerating people rather than building a tool that makes it like more unclear what you should do as the human right like I think like to you know give an example right now like nowadays when you work with a coding agent um it writes a ton of code but it turns out writing code is actually one of the most fun parts of software engineering for many software engineers and so then you end up reviewing ai code right and that's often a less fun part of the job for many software engineers right and so i actually think like we see that like this this comes up”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“I think there's two sides to it. But the one we were just talking about is this idea that maybe every agent should actually use code and be a coding agent. And in my mind, that's just like a small part of this broader idea that like, hey, as we make code even more ubiquitous, I mean, you could probably claim it's ubiquitous today, even pre-AI, right? But as we make code even more ubiquitous, it's actually just going to be used for many more purposes.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Think so, I think coding agents are pretty great. I think And then I think agents outside of coding, it's still very early. And this is just my opinion. But I think they're going to get a whole lot better once they can use coding too in a composable way. Is it's kind of the fun part of when you're building for software engineers in my startup, we were building for software engineers too for a lot of that journey. And they're just such a fun audience to build for because they also like building for themselves and are often even more creative than we are in thinking about how to use the technology. And so like by building for software engineers, you get to just observe a ton of emergent behaviors and like things that you should do and build into the product.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“And so we start to get to this thing where, like, yeah, we have this agent sitting in front of a computer, but we need to make that configurable for the team or for the user, right? And let them like stuff that the agent does often, we probably just want to like build in as a competency that this agent has that it can do. So I think we end up with this generalizable thing that you were saying of like an agent that can just write its own scripts for whatever it wants to do. But I think that the really key part here is can we make it so that everything that the agent has to do often or that it does well, we can just like remember and store so that the agent doesn't have to write a script for that again, right? Or maybe like if I just joined a team and you are already on the same team as me, I can just like use all those scripts that the agents had written already.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“More about, well, how can we help the agent understand the context that it's working in and like the team that's using it probably has a way that they like to do things. They have guidelines. They probably want certain deterministic guarantees about what the agent can or cannot do. Or they want to know that the agent understands sort of this detail. Like an example would be, you know, if we're looking at a crash reporting tool, hitting a connector for it, every sub team probably has a different meta prompt for like how they want the crashes to be analyzed.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“For this, it's like quite early, but we're starting to see people who are using codecs for coding adjacent product purposes. And so as that develops, I think we'll just naturally see that, like, oh, it turns out we should just always have the agent right code if there is a coding way to solve a problem instead of, you know, even if you're doing it financial analysis, right? Like maybe write some code for that. So basically, like, you know, you were like, hey, is this like the two ends of this product for the super assistant, right? Of chat GPT. In my mind, just coding is a core competency of any agent, including ChatGPT. And so really what we think we're building is that competency. So here's like the really cool thing about agents writing code is that you can import code. Right. Code is composable, interoperable, right? Because if we, you know, one very reductive view we could have for an agent is it's just going to be given a computer and it's just going to like point and click and go around. But that is the future and then how we get there is difficult to sort of chart a path because a lot of the questions around building agents aren't like, can the agent do it? But it's”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Okay, so now we're like, okay, we need the super system that can use a computer, right? Or many computers. And now the question is, okay, well, how should it use the computer? And there's lots of ways to use a computer. You could try to hack the OS and use accessibility APIs, maybe a bit easier as you could point and click. That's a little slow and unpredictable sometimes. And another way, it turns out the best way for models to use computers is simply to write code. And so we're kind of getting to this idea where like, well, if you want to build any agent, maybe you should be building a coding agent. And maybe to the user, a non-technical user, they won't even know they're using a coding agent the same way that no one thinks about are they using the internet or not which is they're more just like is wi-fi on right so i think that what we're doing with codecs is we're building a software engineering teammate and as part of that we're kind of building an agent that can use a computer by writing code and so we're already seeing like some”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“A lot of this might happen sooner, but in terms of fuzziness, I think the one year. So I'll give you like. Contention and like a plausible way we get there. But as for how it happens, who knows? So basically, we're going to build a super system, it has to be able to do things, right? We're going to have a model and it's going to be able to do stuff affecting your world. And one of the learnings I think we've seen over the past year or so is that for models to do stuff, they're much more effective when they can use a computer.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“And, like, look at the code and work with the code. So, I think what we need to build as Open AI is basically this idea of you have chat, chat GPT, and not as a tool that's ubiquitously available to everyone. You start using it even outside of work, right? To just help you, you become very comfortable with the idea of being accelerated with AI. And so then you get to work and you just can naturally just, yeah, I'm just going to ask it for this. And I don't need to know about all the connectors or like all the different features. I'm just going to ask it for help and it'll surface to me the best way that it can help at this point in time and maybe even chime in when I didn't ask it for help. So, in my mind, if we can get to that, I think that's how we really build the winning product.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“AI teammate or super assistant that you just talk to, and it just knows how to be helpful on its own, right? And so you don't have to be reading the latest tips for how to use it. You just like you've plugged it in and it just provides help. And so that's kind of the shape of what I think we're building. And I think that will be like a very sticky winning product if we can do so. So the shape that in my head, at least I have, is that we build maybe a fun topic is like, is chat the right interface for AI? I actually think chat is a very good interface when you don't know what you're supposed to use it for. In the same way that if I think of like I'm like on MS Teams or in Slack with a teammate, chat is pretty good. I can ask for whatever I want, right? It's like it's kind of the kind of the nominator for everything. So you can chat with a super assistant about whatever topic you want, whether it be coding or not. And then if you are like a functional expert in a specific domain such as coding, there's like a GUI that you can pull up to go really deep.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Again, comes back to this idea of like building a teammate and not just a teammate that participates in team planning and prioritization, not just the teammate that really tests its code and helps you maintain and deploy, but even a teammate, you know, like if you think again, an engineering teammate, they can also schedule a calendar invite, right? Or move stand up or do whatever, right? And so in my mind, if we just imagine that every day or every week some like crazy new capability is just going to be deployed by a research lab, it's just impossible for us as humans to keep up and use all this technology. And so I think we need to get to this world where you kind of just have like”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Way that we built codecs is that it just uses the shell, but in order to make that safer and secure, we have a sandbox that the model is used to operating in. So I think one of the biggest accelerants to go all the way back to your answer to your question is just like we're building all three things in parallel and like kind of tuning each one and constantly experimenting with how those things work with like a tightly integrated product and research team.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“And so, if you want to train a model to be good at all the different ways it could work, maybe you have a strong opinion that it should work using semantic search, right? Maybe you have a strong opinion that it should call bespoke tools. Or maybe you have, in our case, a strong opinion that it should just use the shell, work in the terminal. You can move much faster if you're just optimizing for one of those worlds. And so.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“And so for a model to work continuously for that amount of time, it's going to exceed its context window. And so we have a solution for that, which we call compaction. But compaction is actually a feature that uses all three layers of that stack. So you need to have a model that has a concept of compaction. And those like, okay, as I start to approach this context window, I might be asked to prepare to be run in a new context window. And then at the API layer, you need an API that understands this concept and has an endpoint that you can hit to do this change. And at the harness layer, you need a harness that can prepare the payload for this to be done. And so shipping this compaction feature that now just like made this behavior possible to anyone using codecs actually meant working across all three things. And I think that's increasingly going to be true. Another maybe like underappreciated version of this is if you think about all the different coding products out there, they all have like very different tool harnesses with like very different opinions on how the model should work.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“One of the things they're really proud of is you can have GPT 5.1 cut specs work for really long periods of time. That's not like normal, but you can set it up to do that or that might happen. But now routinely we'll hear about people saying like, yeah, it ran like overnight or it ran for 24 hours.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Some of how we're thinking about this is evolving a little bit from being like, yeah, we're just going to think about the model and let's just train the best model to really thinking about what is an agent actually overall And I'm not going to try to define agent exactly, but at least the stack that we think of it as having is it's like you have this model, really smart reasoning model, that knows how to do a specific kind of task really well so we can talk about how we make that possible. But then actually we need to serve that model through an API into a harness.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah, so there's like a few components here. I guess you were mentioning models and the models have improved a ton. In fact, just last Wednesday, we shipped GPT 511 Codex Max, a very accurately named model. That is awesome. It is awesome both because it is for any given task that you were using GPT 511 codecs for, it's like roughly 30% faster at accomplishing that task. But also it unlocks a ton of intelligence. If you use it at our higher reasoning levels, it's just like even smarter. And that feedback or that tweet you were saying, like Carpathi made about like, hey, give us your gnarly bugs. Like, you know, obviously there's a ton going on in the market right now, but like Codex Max is definitely like carrying that mantle of, you know, tackling the hardest bugs. So that is super cool. But I will say it's like.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Totally. And this was, it was quite interesting because we dog food products a ton at OpenAI. So, you know, dog food, as in we use our own product. And so Codex has been accelerating OpenAI over the course of the entire year. And the cloud product was a massive accelerance of the company as well. It just turns out that this was one of those places where the signal we got from dog fooding is a little bit different from the signal you get from like the general market because it open ai you know we train reasoning models all day and so we're very used to this kind of prompting and like you know think up front run things massively in parallel and uh you know it would take some time and then come back to it later asynchronously and so you know now when we build we still get a ton of signal from dog footing internally but uh you know we're also very cognizant of like the different ways that different audiences use the product”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“But if as you work with them side by side, you can be like, oh, you don't have a password for this service we use. Here's the password for this service. You know, yeah, don't worry. Feel free to run this command. Then it's much easier for them to then go off and do work for hours without you.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“In the environment. And if it's a command that doesn't work in the sandbox, you can just ask you. And so you can get into this really strong feedback loop using the model. And then over time, like our team's job is to help turn that feedback loop into you sort of as a byproduct of using the product, configuring it so that you can then be delegating to it down the line. And again, Naji, keep coming back to it, but if you hire a teammate and you ask him to do work, but you just give them a fresh computer from the store, it's going to be hard for them to do their job, right?”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Never allowed to get on a call with them. And you can only go back and forth asynchronously over time. Like that works for some teammates. And eventually that's actually how you want to spend most of your time. So that's still the future. But it's hard to initially adopt. So we still have that vision of like that's what we're trying to get you to, a teammate that you delegate to and then it's proactive. And we're seeing that growing. But the key unlock is actually first you need to land with users in a way that's like much more intuitive and like trivial to get value from. So the way that most people discover, like the vast majority of users discover codecs today is either they download an IDE extension or they run it in their CLI and the agent works there with you on your computer interactively. And it works within a sandbox, which is actually like a really cool piece of tech to help that be safe and secure, but it has access to all those dependencies. So if the agent needs to do something, like it needs to run a command, it can do so within the sandbox. We don't have to set.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“We have this strong sort of mission here at OpenAI to basically build AGI. And so we think a lot about how can we shape the product so that it can scale. Earlier I was mentioning like, hey, if you're an engineer, you should be getting help from AI like thousands of times per day. And so we thought a lot about the primitives for that when we launched our first version of codecs, which was Codex Cloud. And that was basically a product that had its own computer. It lived in the cloud. You could delegate to it. And the sort of the coolest part about that was you could run many, many tasks in parallel. But some of the challenges that we saw are that it's a little bit harder to set that up, both in terms of like environment configuration, like giving the model the tools it needs to validate its changes and to learn how to prompt in that way. And sort of my analogy for this is going back to this teammate analogy. It's like if you hired a teammate, but you're”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“And then what we've started to see more recently actually is that other major sort of API coding customers are now starting to adopt these models as well. And so we've reached the point where actually the codecs model is the most served coding model in the API as well.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah, it's been Codex has been growing absolutely explosively since the launch of GPT5 back in August. There's some definitely some interesting product insights to talk about as to how we unlock that growth if you're interested. But again, the last stat we showed there was like we were like well over 10x since August. In fact, it's been like 20x since then. Also, the codex models are serving many trillions of tokens a week now. And it's basically like our most served coding model. One of the really cool things that we've seen is that the way that we decided to set up the codecs team was to build a really tightly integrated product and research team that are iterating on the model and the harness together. And it turns out that lets you just do a lot more and try many more experiments as to how these things will work together. And so we were just training these models for use in our first party harness that we were very opinionated about.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“It's basically this idea that we want the way if you're a developer and you're trying to get something done. We want you to just feel like you have superpowers and you're able to move much, much faster. But we don't think You need to be sitting there constantly thinking about how can I invoke AI at this point to do this thing? We want you to be able to sort of like plug it in to the way that you work and have it just start to do stuff without you having to think about it.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Think this is critically important to achieve the mission of OpenAI, which is to deliver the benefits of AGI to all humanity. You know, I like to joke today that like AI products, and it's a half joke, they're actually really hard to use because you have to be very thoughtful about when it could help you. And if you're not prompting a model to help you, it's probably not helping you at that time. And if you think of how many times like the average user is prompting AI today, it's probably like tens of times. But if you think of how many times people could actually get benefit from a really intelligent entity, it's thousands of times per day. And so a large part of our goal with Codex is to figure out like, what is the shape of an actual teammate agent that is sort of helpful by default?”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Going on, and like that's not only true when it's performing a task, but again, if you think of the best human teammates, like you don't tell them what to do, right? Like maybe when you first hire them, you have a couple meetings and you're like, hey, like you kind of learn like, okay, this is these prompts work for this teammate. These prompts don't, right? This is how to communicate with this person. Then eventually you give them some starter tasks to delegate a few tasks. But then eventually you just say like, hey, great, okay, you're working with this set of people in this area of the code base, you know, feel free to work with other people in other parts of the code base too, even. And yeah, you tell me what you think makes sense to be done, right? And so, you know, we think of this as like proactivity and like one of our major goals with codecs is to like get to proactivity.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“We want to get to the point where it can work just like a new intern that you hire, you don't only ask them to write code, but you ask them to participate across the cycle. And so you know that even if they don't get something right the first try, they're eventually going to be able to iterate their rate there.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Software and then further downstream in terms of like validation, deploying, and like maintaining code. Make that a little more fun, like one thing I like to imagine is like if you think of what codex is today, it's a bit like this really smart intern that like refuses to read Slack and like doesn't check data dog or like sentry unless you ask it to. And so like no matter how smart it is, like how much you're going to trust it to write code without you also working with it, right? So that's how people use it mostly today is they pair with it.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Totally, yeah. So I had the very lucky job of living in the future and leading product on Codecs. And Codex is OpenAI's coding agent. So super concretely, that means it's an IDE extension, a VS code extension, that you can install or a terminal tool that you can install. And when you do so, you can then basically pair with Codex to answer questions about code, write code, run tests, execute code, and do a bunch of the work in sort of that thick middle section of the software development lifecycle, which is all about writing code that you're going to get into production. More broadly, we think of codecs as like what it currently is is just the beginning of a software engineering teammate. And so when we use a big word like teammate, like some of the things we're imagining are that it's not only able to write code, but actually it participates early on in like the ideation and planning phases of writing.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Just super rising Was just like, again, surprised or even shocked when I arrived at like the level of individual drive and like. Autonomy that everyone here has. So I think the way that OpenAI runs, like many, you can't read this or be listen to a podcast and be like, I'm just going to deploy this to my company. You know, maybe this is a hard thing to say, but I think like, yeah, very few companies have the talent caliber to be able to do that. So it might need to be like adjusted if you were going to implement this.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Like, what product will we build and therefore, how will people use that product? That's the place where we're much more like, let's find out empirically.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah. It's like, okay, to use this analogy a little bit, I feel like there is an aim component, but the aim component is much fuzzier. You know, it's kind of like roughly, what do we think can happen? Like someone I've learned a ton from working here is a research lead and he likes to say that like an open AI, we can have really good conversations about something that's like a year plus from now. And, you know, there's a lot of ambiguity in what will happen, but that's a right sort of timeline. And then we can have really good conversations about what's happening in low months or weeks. But there's kind of this awkward middle ground, which was like as you start approaching a year, but you're not at a year where it's like very difficult to reason about, right? And so as far as like aiming, I think we want to know like, okay, what are some of the futures that we're trying to build towards? And like a lot of the problems we're dealing with in AI, like such as alignment are problems you need to be thinking out like really far out into the future. So we're kind of aiming fuzzily there. But when it comes down to the more tactically like, oh yeah.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“That now It'll be interesting if I ever work at like, I don't think it'll ever make sense to work at a non AI company in the future. I don't even know what that means. But if I were to imagine it or go back in time, I think I would like run things totally differently.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“But I think there's a lot of sort of counterintuitive things that surprise me when I arrived as far as how things are structured. One example that comes to mind is like when I was working on my startup and before that, when I was at Dropbox, it was like very important, especially as a PM to always kind of rally the ship. And it was kind of like make sure you're pointed in the right direction and then you can like accelerate in that direction. But here, I think because we don't exactly know what capabilities will even come up soon and we don't know what's going to work technically. And then we also don't know what's going to land even if it worked technically. It's much more important for us to be very like humble and learn a lot more empirically and just try things quickly. And like the org is set up in that way to be incredibly bottoms up. You know, this is, again, one of those things that like, as you were saying, everyone wants to move fast. I think everyone likes to say that they're bottoms up, or at least a lot of people do. But Open AI is like truly, truly bottoms up. And that's being a learning experience for me.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source
“I mean, so one thing is just the technology that we're building with has just transformed so many things from both how we build, but also like what kinds of things we can enable for users. And we spend most of our time talking about like the sort of improvements in the foundation models. But I believe that even if we had no more progress today with models, which is absolutely not the case. But even if we had no more progress, we are way behind on product. There's so much more product to build. So I think like just like the moment is ripe.”
2025-12-14 · Lenny's Podcast · Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead) · IDENTIFIED FROM THE TRANSCRIPT · source