Alexander Embiricos: recommendation

Transcript context

…You hinted at this, what unlocked this growth, I'm extremely interested in hearing that. It felt like before, I don't know, maybe this was before you joined the team, it just felt like Cloud Code was killing it. Just everyone was sitting on top of Cloud Code. It was by far the best way to code. And then all of a sudden Codex comes around. I remember Karpathy tweeted that he just has never seen a model like this. I think the tweet was the gnarliest bugs that he runs into that he just spends hours trying to figure out nothing else has solved, he gives it to Codex, lets it run for an hour and it solves it. What'd you guys do? We have this strong sort of mission here at OpenAI basically to 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 thousands of times per day," and so we thought a lot about the primitives for that when we launched our first version of Codex, which was Codex Cloud. And that was basically a product that had its own computer, lived in the cloud, you could delegate to it. And the coolest part about that is 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 environment configuration, like giving the model the tools it needs to validate its changes and to learn how to prompt in that way. My analogy for this is, going back to this teammate analogy, it's like if you hired a teammate, but you're never allowed to get on a call with them and you can only go back and forth asynchronously over time. 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. And 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 is 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 much more intuitive and trivial to get value from. So the way that most people discover, the vast majority of users discover Codex 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 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, it needs to run a command, it can do so within the sandbox. We don't have to set up any environment. And if it's a command that doesn't work in the sandbox, it can just ask you. And so you can get into this really strong feedback loop using the model. And then over time, our team's job is to help turn that feedback loop into you as a byproduct of using the product, configuring it so that you can then be delegating to it down the line. And again, analogy, keep coming back to it, but if you hire a teammate and you ask them 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. But if as you work with them side by side, you could be like, "Oh, you don't have a password for this service we use, here's the password for this service. 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. So what I'm hearing is the initial version of Codex was almost too far in the future. It's like a remote in the cloud agent that's coding for you asynchronously. And what you did is, "Okay, let's actually come back a little bit, let's integrate into the way engineers already integrate into IDs and locally and help them on ramp to this new world,"…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.