תעתיק 🔗
WEBVTT
00:00:06.719 --> 00:00:10.560
Hello and welcome to another episode of the Procamirba Podcast.
00:00:10.640 --> 00:00:16.800
I am your host, Jan Tego Engbetriebel, and with me today in the Saline studio is Dennis Becker.
00:00:17.120 --> 00:00:17.920
Great to be here.
00:00:18.239 --> 00:00:19.039
Hi Dennis.
00:00:19.359 --> 00:00:23.600
Our regular listeners will have noticed that this episode is somewhat unusual for us.
00:00:23.760 --> 00:00:30.320
It is only our third ever episode in English, and my first ever episode in English for this podcast.
00:00:31.039 --> 00:00:39.039
Because today we want to talk about how the developer experience is changing with the rise of AI.
00:00:39.119 --> 00:00:42.320
And maybe I don't know if rise even is the right term anymore.
00:00:42.399 --> 00:00:50.159
It feels less like a rise because that kind of like has already happened and more like a we're in the dominance area already.
00:00:50.399 --> 00:00:52.479
Maybe we'll see about that.
00:00:52.880 --> 00:00:59.039
And while we could have this conversation in English or in German, um, it's much more fun uh to have a guest on as well.
00:00:59.119 --> 00:01:06.640
And because our guest uh prefers to have this conversation in English, we're totally on board with that because uh the alternative would have been to do this in French.
00:01:06.719 --> 00:01:08.799
And trust me, you don't want to hear my French.
00:01:08.879 --> 00:01:12.319
Uh I I can't speak to Dennis's French, uh, but my French is like terrible.
00:01:12.640 --> 00:01:14.239
It wouldn't be enough for a podcast.
00:01:14.480 --> 00:01:15.120
Yeah.
00:01:15.359 --> 00:01:19.840
And uh with that being said, I would like to welcome uh Romain Huyet.
00:01:19.920 --> 00:01:21.040
Uh welcome to the program.
00:01:21.519 --> 00:01:23.120
Thank you so much for for having me.
00:01:23.200 --> 00:01:25.359
Uh guten Tag, ich heiße Romain.
00:01:25.439 --> 00:01:28.000
Uh, ich freue mich sehr illustr.
00:01:28.400 --> 00:01:33.519
Uh, you know, I studied German for many years, but I don't think we could carry the whole conversation in in German.
00:01:33.599 --> 00:01:36.319
So thanks for doing it in English.
00:01:36.879 --> 00:01:38.079
That's uh totally fine.
00:01:38.159 --> 00:01:42.640
We didn't invite you here for your German skills, but rather to talk about your actual job.
00:01:42.799 --> 00:01:53.120
And when we did the uh research on your like checked you out on what we could find on the internet, um you have done or you've been in developer experience for quite some time now.
00:01:53.280 --> 00:01:55.439
You've been at Stripe and at Twitter.
00:01:55.599 --> 00:02:09.280
Um, and maybe just as a small anecdote here, when I was doing developer relations at Shopify, Stripe was one of the like go-to companies that we would look at and see like how can you actually do developer experience on a like top-notch level.
00:02:09.439 --> 00:02:11.919
So uh kind of speaks to your work here.
00:02:12.240 --> 00:02:12.879
Thank you so much.
00:02:12.960 --> 00:02:13.280
Yeah.
00:02:13.520 --> 00:02:31.919
I mean, like I spent uh yeah, five years at Stripe uh building the developer relations, developer experience function, but also obsessing about the product, which was effectively for Stripe, the API design, the documentation, the SDKs, all of the tools that would make developers actually love integrating payments.
00:02:32.159 --> 00:02:38.000
And when you think about Stripe, um integrating payments had always been a pain for developers.
00:02:38.159 --> 00:02:50.560
It was always the thing before Stripe that they hated the most doing because it meant, oh gosh, I'm gonna have to leave my favorite Node.js and Ruby and Rails development all of a sudden to integrate with these very legacy APIs.
00:02:50.639 --> 00:02:54.960
And all of a sudden with Stripe, we made something very delightful and magical to use.
00:02:55.520 --> 00:03:02.080
And uh yeah, I spent a lot of my time making that uh uh as as great as possible with a fantastic team.
00:03:02.159 --> 00:03:14.240
And I think what I found the most exciting about Stripe is how everyone, from Patrick and John, the founders, would care so much about the quality of the API and the abstractions that we would put out there.
00:03:14.400 --> 00:03:17.759
So that was very much in the DNA of the company to have great DX.
00:03:18.639 --> 00:03:19.199
Interesting.
00:03:19.840 --> 00:03:23.439
And now you're head of developer experience at OpenAI.
00:03:23.680 --> 00:03:28.560
Um and what I want to say, uh thank you for the invitation to OpenAI's developer event in Berlin.
00:03:28.639 --> 00:03:33.360
We sadly couldn't record this on-site in Berlin, but we now get to do this uh remote.
00:03:33.439 --> 00:03:35.759
And uh you're in from Paris, right?
00:03:36.159 --> 00:03:36.560
Exactly.
00:03:36.719 --> 00:03:41.360
Well, we just uh hosted Berlin was the first of our stops in Europe.
00:03:41.599 --> 00:03:42.719
It was a great event.
00:03:42.879 --> 00:03:52.560
We did uh Paris uh this week uh after Berlin, and then we're going to London uh on on Tuesday as well to kind of wrap this European leg of DevDay Exchange.
00:03:52.639 --> 00:03:57.919
And of course, like people have probably followed DevDSF uh about four weeks ago now.
00:03:58.159 --> 00:04:00.719
It's like kind of like our flagship annual event.
00:04:00.879 --> 00:04:12.240
And we could stop there, but I think for us it matters also a ton to not just launch things from San Francisco and have a live stream, but also connect with the developer community around the world, right?
00:04:12.319 --> 00:04:17.680
And I think there are amazing developers in Berlin, um, and uh same in Paris and London.
00:04:17.759 --> 00:04:22.399
And we want to make sure we also spend time with them, to also listen to their feedback and what they need.
00:04:22.560 --> 00:04:25.439
Uh, it's important to like have that feedback for us.
00:04:25.519
Load full transcript
--> 00:04:29.120
That's very much the way we make the products better and how we build them.
00:04:29.279 --> 00:04:42.639
Um, an example of that, for instance, was uh every time I traveled to Europe the past year or so, people kept mentioning that they wanted to do more voice AI, but the accents and the voices in German or in French were not that great yet.
00:04:42.800 --> 00:04:43.759
And I agreed.
00:04:44.000 --> 00:04:51.360
And so that feedback was critical into making like GPT Live 1.1, which is now much better with the German accent and French accent, for instance.
00:04:51.519 --> 00:04:56.079
So yeah, we love spending time on the road and hearing from developers who are building great things in Europe.
00:04:57.040 --> 00:05:02.160
And what is your response to people who ask you why you still need this face-to-face time?
00:05:02.240 --> 00:05:08.800
Like lots of development work is becoming more remote, more asynchronous, um, teams are more and more distributed.
00:05:08.959 --> 00:05:11.839
Why do you care so much about that face-to-face time in person?
00:05:12.720 --> 00:05:25.120
I mean, the the sense of belonging and community, I think, is more important than ever, uh, especially after like some of the kind of COVID years where people had to uh work um, you know, uh remotely for a while.
00:05:25.279 --> 00:05:36.480
When you look at the rise of AI and things like coding agents, for instance, it's definitely true that now you can do a lot with those tools, a lot more that you could possibly accomplish before.
00:05:36.639 --> 00:05:42.639
But at the same time, the sense of like belonging, the learning from each other is becoming more critical, right?
00:05:42.720 --> 00:05:49.759
Like if you rewind just a few years ago, um, the technology you built upon was not changing that fast.
00:05:50.160 --> 00:05:55.439
Now, like all of the techniques that you might have practiced for the past like six months are already changing.
00:05:55.680 --> 00:06:07.600
In fact, I often tell like engineering teams like if you have, I don't know, your skills or your agents MD files that have been untouched for four months or something like that, they're probably far outdated by now.
00:06:07.680 --> 00:06:13.600
Like the frontier models are becoming so great that um you want to learn from each other.
00:06:13.759 --> 00:06:19.920
You want to like actually connect in person, share the best practices, see what the others are building.
00:06:20.160 --> 00:06:26.800
And uh, as I hinted at, also like um the needs from different people from different regions are also different.
00:06:26.959 --> 00:06:31.920
We mentioned languages, but uh when we go to APAC, that's even more acute, right?
00:06:32.000 --> 00:06:34.399
So, like in in Japan, in Korea.
00:06:34.560 --> 00:06:40.319
Um, the price points of of what we can offer these models for is also critical for some other regions.
00:06:40.399 --> 00:06:43.600
And so, yeah, the the community building is critical.
00:06:43.680 --> 00:06:52.000
And that's why I think we can do some of that work with our team that we've been growing around the world, but there's some limits to that.
00:06:52.079 --> 00:06:55.279
And that's why we created programs like Codex Ambassadors, for instance.
00:06:55.360 --> 00:07:07.279
And Codex Ambassadors are amazing to us because they are passionate builders using tools like Codex that are able to run events in regions uh that we cannot reach yet.
00:07:07.439 --> 00:07:11.360
You know, like there are cities around the world where we don't have OpenAI employees yet.
00:07:11.439 --> 00:07:13.759
Uh, it may take us a while to get there even.
00:07:14.000 --> 00:07:18.399
Um, but luckily we can rely on these codecs ambassadors building events.
00:07:18.480 --> 00:07:21.839
And yes, this sense of belonging is really beloved by the community.
00:07:22.319 --> 00:07:26.800
If we're talking about like developer experience, it already includes the word developer.
00:07:26.959 --> 00:07:33.600
Um, how quickly or at all is it at all changing your target group of people that you want to reach?
00:07:33.759 --> 00:07:40.000
Because I mean, historically, yes, it was very much a developer role and codex, you first think of uh developers as well.
00:07:40.160 --> 00:07:51.279
Um, but I guess we for us and also in the market see that like the roles are quickly merging between the different um yeah uh roles that we had before.
00:07:51.519 --> 00:07:55.199
Uh how quickly do you see that in, for example, your events?
00:07:55.279 --> 00:07:58.240
Like, did the target group already change, the visitors change?
00:07:58.560 --> 00:07:59.279
Oh, absolutely.
00:07:59.360 --> 00:08:01.600
I think like there's there's two dimensions, right?
00:08:01.680 --> 00:08:06.399
There's like um what is the definition of a developer?
00:08:06.480 --> 00:08:09.680
And that definition is kind of rapidly changing and evolving.
00:08:09.920 --> 00:08:19.040
Uh, and then there's also like um, who has now access to these tools and superpowers that would be uh otherwise limited to them in the past.
00:08:19.120 --> 00:08:19.439
Yeah.
00:08:19.680 --> 00:08:29.120
And so uh for us, like we um we don't want to draw like a very sharp line between like this binary decision of like, oh, you're either a developer or you're not.
00:08:29.279 --> 00:08:47.440
Um now what we've noticed is that we always want to serve, of course, the kind of early adopters, the people are kind of pushing the frontier of the tools, the people who are experimenting with the next model every few weeks, you know, and try to collect feedback directly from them because they are the tastemakers.
00:08:47.600 --> 00:08:55.759
But at the same time, what's amazing to see is that like in the past year alone, now like a product manager is way more technical than they used to be.
00:08:55.919 --> 00:09:05.039
Now, instead of writing a document to explore an idea with their team, I've seen many
PMs at OpenAI included building a prototype and actually reviewing that prototype with the team.
00:09:05.279 --> 00:09:09.200
Now, a designer can even test an interaction instead of just staying in Figma.
00:09:09.360 --> 00:09:24.159
Or uh at OpenAI, what we've noticed too, and the whole reason why we relaunched uh ChatGPT as one app combined with Codecs in a few a couple of months ago is because we had noticed the finance team, the legal team, the sales team, they were all on Codecs.
00:09:24.399 --> 00:09:31.120
Instead of having like uh the need for an engineer to build a dashboard, they could suddenly like build a live dashboard by themselves.
00:09:31.279 --> 00:09:37.519
Or um, if they wanted to like maintain that dashboard and deploy it every day automatically, they could build that with codecs.
00:09:37.679 --> 00:09:45.039
Or a salesperson, instead of having a deck for a customer, they could actually build a full-on interactive experience for that customer.
00:09:45.200 --> 00:09:54.960
Um, so that's been amazing to see how many more people have this like magic and this like superpower at their disposal, and they're all becoming more technical.
00:09:55.279 --> 00:10:00.240
Um and that gives us like a better feedback loop with the broader community.
00:10:00.399 --> 00:10:15.120
Um, and you could even argue that at some point we should almost rename our team, like because you know, we go beyond developers to with with tools like dots, if we talk about them at some point, like you know, you have proactive agents kind of like helping you with your goals, both at work and in life.
00:10:15.360 --> 00:10:26.960
And so we go far beyond developers, but we still think it's fundamental to obsess about the developers because at the end of the day, the engineers are still kind of the early uh adopters of that technology.
00:10:27.120 --> 00:10:32.240
And we want to make sure that like we're at the frontier with them, listening to their feedback, making the product really great.
00:10:32.320 --> 00:10:37.840
And if they love it, turns out I think the people right after them will also have a great time.
00:10:38.480 --> 00:10:52.159
And when you say you don't want to like have a hard differentiation between um like developers and non-developers, is there a differentiation between people who build with open AI and people who build on open AI?
00:10:52.399 --> 00:10:56.480
So, what I mean by that is there's obviously people who use all these tools that you've just mentioned, right?
00:10:56.639 --> 00:10:59.039
Dots, codecs, sites, whatever.
00:10:59.200 --> 00:11:03.360
And then there's people who just want to integrate with your API and your models.
00:11:03.519 --> 00:11:07.919
And is that two audiences for you, or do you see them merging as well?
00:11:08.240 --> 00:11:11.840
We see the interplay happening a lot more in real time.
00:11:12.080 --> 00:11:19.919
You know, in fact, if you rewind, like three and a half years ago when I was joining up on AI, all we had for developers at the time was an API.
00:11:20.000 --> 00:11:22.080
And we said, hey, we have ChatGPT.
00:11:22.240 --> 00:11:36.000
And but if you want to take these models and bring them to your application, you can take this API for the model, and you'll be able to do things like structured outputs, you'll be able to like um make your first agents.
00:11:36.559 --> 00:11:42.320
And um, you know, at the time we talked about RAG and fine-tuning a lot, but that was kind of the API that we offered.
00:11:42.480 --> 00:11:56.480
But over time, now that you have tools like Codecs, you can actually build software, include like any kind of software, even if it's not an AI-powered feature, but if it's like actually anything that you want to build, but assisted with Codecs and AI.
00:11:56.720 --> 00:11:59.759
And we've seen this interplay now going hand in hand.
00:11:59.919 --> 00:12:07.840
The idea of like, I'm starting building an app with codecs, whether it's an iPhone app, an Android app, maybe some web applications or something, I'm iterating quickly.
00:12:08.000 --> 00:12:12.000
But of course, I want that experience to be awesome for my customers too.
00:12:12.080 --> 00:12:15.919
And I want them to have the same feeling as I have when I'm using Codecs.
00:12:16.080 --> 00:12:20.559
So I'm I talk to a lot of developers who are like, I want to bring speech to speech to my app.
00:12:20.720 --> 00:12:22.639
I want to bring Imogen to my app.
00:12:22.720 --> 00:12:28.399
I want to bring all of these capabilities that I have in Codecs, but I want those in my app too.
00:12:28.639 --> 00:12:30.399
And we've noticed that already.
00:12:30.559 --> 00:12:45.600
And in fact, it's very timely because um just at Dev de Paris this week, uh, what uh what I announced was uh this new launch uh where you can now have, as part of your Chat GPT subscription, you now have API credits as well added on top.
00:12:46.159 --> 00:12:50.399
And we've also launched an open AI developers plugins for Codec.
00:12:50.639 --> 00:12:54.720
So, all in all, what that means is that I'm starting building something with Codecs.
00:12:55.039 --> 00:12:59.120
I want to have the most frictionless experience to bring these capabilities.
00:12:59.279 --> 00:13:05.200
I don't have to worry about setting up an ABI key because Codecs can do that for me automatically from Codecs.
00:13:05.360 --> 00:13:12.480
I don't have to worry about putting a card on file again at the beginning because I have credits automatically to start building right away.
00:13:12.639 --> 00:13:13.360
And guess what?
00:13:13.519 --> 00:13:20.799
I don't even have to read docs because I can just ask Codec, hey, like I would like you to bring uh speech to speech here with GPT Live.
00:13:20.960 --> 00:13:33.600
And when my customer is uploading these pictures of their uh home, for instance, I want you to generate an image of what the furniture could look like if you're building like a home
decor uh application, for instance.
00:13:33.759 --> 00:13:39.759
So uh back to your question, we really see this into play more and more, and we actually want to facilitate that.
00:13:39.919 --> 00:13:48.960
Uh, whether you're an API uh customer on OpenAI or a codex builder, we want that uh to be very like seamless to go uh between the two.
00:13:50.000 --> 00:13:52.879
And maybe one last question regarding that audience.
00:13:53.200 --> 00:13:59.360
Um in the past, addressing developers was probably not cheap but straightforward, right?
00:13:59.440 --> 00:14:05.600
There is like conferences, there is uh content sites where you could advertise, there is events that you could host yourself.
00:14:06.080 --> 00:14:08.799
Sooner or later you would somehow reach these developers.
00:14:09.120 --> 00:14:15.679
And now that like product people come in, designers come in every day, people from the street come in and just like start building.
00:14:16.000 --> 00:14:25.200
How do you actually reach those people and how do you make them aware of the fact that by some definition they are developers as well and they should feel addressed by what you do?
00:14:25.600 --> 00:14:29.120
Yeah, no, it's fascinating and it's a it's a great question, right?
00:14:29.200 --> 00:14:32.399
Because these tools are very much blurring lines.
00:14:32.480 --> 00:14:44.559
Um, that's why on our team, like we've been spending a lot of time trying to educate uh people with video content, uh different kinds of formats, like written guides on what you can accomplish with the platform.
00:14:44.799 --> 00:14:52.080
The fact that, for instance, you can not just ask Codex to build an app, but then say, hey, actually deploy it on the internet.
00:14:52.480 --> 00:14:56.159
And all of that, you can have a ChatGPT site up and running.
00:14:56.320 --> 00:14:59.919
Or of course, if you have like a Vercell or a Cloudflare account, it can use that as well.
00:15:00.080 --> 00:15:14.399
Like all of these things, like um ChatGPT sites actually having like a full-on uh database, uh, you know, synchronization, WebSocket support, all of these things is quite magical because you don't have to worry about infra as much.
00:15:14.639 --> 00:15:18.480
Um so yeah, it's a rapidly evolving field.
00:15:18.559 --> 00:15:23.519
We want to make sure that people know about these capabilities that are being added to the product.
00:15:23.679 --> 00:15:32.960
Uh, but meanwhile, that's also why we still care a lot about uh DX, because I think uh developer experience is also itself evolving.
00:15:33.039 --> 00:15:44.639
And what might have been for Stripe, like let's design the best API and the best SDK we can, uh, still very much is relevant, but also expands to other things.
00:15:44.799 --> 00:15:51.679
Like, well, now I'm fully expecting that my customers are using coding agents and tools to assist them in building software.
00:15:51.840 --> 00:15:55.840
And so that should be the obsession we now
have about developer experiences.
00:15:56.159 --> 00:16:00.000
Like, is it frictionless to ask the coding agent to build something?
00:16:00.240 --> 00:16:04.000
Will it find the right primitives, the right model, the right abstractions?
00:16:04.159 --> 00:16:11.519
Um, and we're effectively designing now both for humans and their agents uh to make sure that they can import very seamlessly.
00:16:11.840 --> 00:16:14.720
And in that sense, you're in a very privileged position, right?
00:16:14.879 --> 00:16:22.080
Because you also maintain the coding agent that most of these people will use to actually build the products and then integrate again with OpenAI.
00:16:22.720 --> 00:16:36.559
What would your message be for every other platform out there that doesn't host its own coding agent, that doesn't bring its own harness, that is like reliant on other coding agents to pick them up and to kind of understand, like Stripe, for example, right?
00:16:36.799 --> 00:16:38.000
We mentioned Stripe here before.
00:16:38.080 --> 00:16:45.360
So, what is your message to people like Stripe on how they should be doing developer experience in the age of coding agents?
00:16:45.679 --> 00:16:59.279
Yeah, um, well, I think one uh thing that's been amazing to observe, in my view, has been that the platforms that had the best DX for humans tend to also be the platform that have the best DX for agents already.
00:16:59.360 --> 00:17:08.559
Uh, you mentioned Stripe, Stripe is doing fantastic work, but also all the years of crafting great abstractions for humans have actually paid off now, even for agents.
00:17:08.720 --> 00:17:17.759
Uh, you see um products like Superbase, for instance, like, you know, agents actually have a great time using with Superbase, even though they designed these things before.
00:17:18.000 --> 00:17:28.160
My message to them would also be like, try to keep on building these great abstractions, but try to see what um tools do agents need to do their best work.
00:17:28.319 --> 00:17:36.960
So, for instance, if you have an MCP server, that's even better because all of a sudden, like uh the agent can uh access that directly.
00:17:37.119 --> 00:17:44.319
If your docs are consumed by humans well, that's already a good enough step because the agent will be able to figure out.
00:17:44.480 --> 00:17:51.759
But if you're able to serve your docs for agents specifically in a more efficient manner, that can make them even more effective.
00:17:52.000 --> 00:17:57.279
Um and then trying to offer the tools too for like things like WebMCP, for instance.
00:17:57.359 --> 00:18:07.440
Imagine you're building not just an app for humans, but if you want the app you're building to also uh be well used by agents.
00:18:07.599 --> 00:18:13.039
Well, it's sure enough, agents like Codex have amazing computer use and act as humans.
00:18:13.200 --> 00:18:23.359
But if you integrate initiatives like WebMCP, then even the apps that your customers build with your
interfaces can have this amazing way to work with agents.
00:18:23.599 --> 00:18:33.519
So we really think about the whole flow, both for the building side uh side of things, but also for the user interacting with the end product.
00:18:33.759 --> 00:18:42.319
Um and so for us, for instance, the journey that we've taken is like, you know, we've been the open AI, we've been building the OpenAI developers plugin uh for codecs.
00:18:42.480 --> 00:18:50.079
I recommend like most platforms to also do such things because then you can just package up skills and functionalities right there.
00:18:50.319 --> 00:19:00.640
And so now if you're using Codecs or another coding agent, um, it's much easier to integrate with your API or your platform because you have this plugin that's like one click away.
00:19:00.799 --> 00:19:07.039
And I, as a human, don't have to read every API ref parameter anymore because my coding agent will do that.
00:19:07.519 --> 00:19:14.880
Do you anticipate a change towards technologies that are maybe less friendly to humans but more optimized for agentic workflows?
00:19:15.039 --> 00:19:19.920
For example, in the past, APIs basically always meant REST APIs, right?
00:19:20.160 --> 00:19:24.400
They're easy to read, they're stateless, um, they're easy to document.
00:19:24.559 --> 00:19:29.759
Um, but there's people out there arguing that technologies like GraphQL, for example, are way more efficient.
00:19:29.920 --> 00:19:35.839
They're just harder to use, which is why hardly anyone out there actually implements proper GraphQL APIs.
00:19:36.000 --> 00:19:45.839
But now that I don't need to read these APIs and these instructions anymore, and my agent and my machine take care of that, will we maybe see a shift towards technologies like that?
00:19:46.319 --> 00:19:58.319
It could, yeah, because like the you know, GraphQL was always fascinating, but a bit hard to reason about because you had to specify the field you needed, the field you did not need, and things like this, but the agent could quickly uh figure this out for you.
00:19:58.400 --> 00:20:00.000
So that's definitely one way.
00:20:00.240 --> 00:20:04.400
Um I think the same question goes for SDKs in some ways, right?
00:20:04.559 --> 00:20:10.079
SDKs were initially meant for humans to abstract away the complexity of interfacing with REST.
00:20:10.240 --> 00:20:18.480
Like, for instance, at every REST API request I'm doing, I want to also make sure that there is like some uh retries built in.
00:20:18.640 --> 00:20:21.759
In the case of Stripe, it's important the payment goes only once.
00:20:21.920 --> 00:20:25.599
And so there is like the idem potency key concept and things like this.
00:20:25.839 --> 00:20:30.400
You could one on one side, you could argue that like, well, coding agents could just write more code.
00:20:30.559 --> 00:20:34.799
At the same time, the SDKs can still be interesting there to package that functionality.
00:20:34.880 --> 00:20:45.920
And
if you're actively reviewing the code made by an agent, it's also nice to see that code quite short and easy to digest when the SDK is officially hiding some of that complexity.
00:20:46.240 --> 00:20:54.319
But sure enough, I think like there are um over time uh we're gonna read less code and we're gonna trust agents more and more.
00:20:54.559 --> 00:21:04.319
And I could see a world where GraphQL and and some other techniques that maybe the human developers were kind of like Barry Sean, um, could be great for agents.
00:21:08.880 --> 00:21:15.200
Yeah, at that point I always wonder like what happens or are new technologies still developed?
00:21:15.359 --> 00:21:22.079
I mean, GraphQL, for example, was developed in a sense that humans uh interact with it and have an easier way uh for it.
00:21:22.160 --> 00:21:28.880
And I I somehow imagine like, is all of this infrastructure that we have and that we use just come to a complete stop now?
00:21:28.960 --> 00:21:40.799
And that's just the current state of that agents know, because there's not really a need for new technologies, like that anybody comes up and says, okay, I have a new GraphQL, because the adoption will be very hard in the future as well.
00:21:42.079 --> 00:21:42.960
I think it depends.
00:21:43.039 --> 00:21:48.880
I think there's always gonna be like uh, you know, products that um could make agents even more effective.
00:21:49.039 --> 00:21:57.599
I think like uh you see, like we we try to come up with like something neutral for agents MD to kind of package up the instructions from any coding agent in one place.
00:21:57.839 --> 00:21:59.119
Same thing for WebMCP.
00:21:59.200 --> 00:21:59.759
I think WebMC.
00:22:00.400 --> 00:22:05.200
Try to like you know provide a net new interface for agents.
00:22:05.440 --> 00:22:12.400
Uh, we've also worked with the the folks who came up with MCP in order to introduce like UI components in MCP.
00:22:12.480 --> 00:22:16.240
That's how we build like the ChatGPT apps, for instance, and the Codex plugins.
00:22:16.559 --> 00:22:23.920
You have like um these UI primitives that you can expose, not just for us, but for any kind of client of that technology.
00:22:24.160 --> 00:22:37.119
So you can see that even if the models are progressing and the coding agents are being more and more effective, there are still ways also to kind of uh bring some standardization across the different agents and always kind of polishing that.
00:22:39.920 --> 00:22:47.039
I want to touch on um a kind of new set of challenges that you might have experienced in developer experience.
00:22:47.359 --> 00:22:53.599
Um so in the past, when working in DevX, you most often interacted with developers, right?
00:22:53.839 --> 00:22:56.799
Or at least with people who were from the same domain.
00:22:56.960 --> 00:23:07.440
So when you talk to them about like what are the constraints within that you operate, you could explain to them why maybe certain features aren't available or why they might take some more time.
00:23:07.759 --> 00:23:17.359
And I feel like AI has kind of opened Pandora's box here because people just assume that these machines are like super intelligent and almost to a human scale.
00:23:17.599 --> 00:23:31.599
And now when they approach you and ask for other features, or why doesn't this work, or why does it behave in this and that way, how do you explain to them the intricacies of like, well, this is an LLM, this is a transformer, it's it's not a human after all.
00:23:31.680 --> 00:23:36.640
It feels like a human very much and oftentimes, but you still work within certain constraints.
00:23:36.720 --> 00:23:38.480
And how do you communicate that?
00:23:39.200 --> 00:23:43.039
Yeah, um it's challenging, right?
00:23:43.279 --> 00:23:49.279
Because uh even the capabilities are changing so frequently now with with this space of AI.
00:23:49.519 --> 00:23:56.559
But ultimately, I think that's where the we're kind of reinventing also the the work of the developer, right?
00:23:56.720 --> 00:24:07.119
What you used to spend time on with all this toil, all of these migrations, now you're spending that time being more creative, be taking an idea and like actually turning it into a real product.
00:24:07.359 --> 00:24:29.359
And I think like even if you're a fantastic PM and you're able to prototype something and you can drive like excitement among your team to build a feature because you can actually interact with the prototype, it doesn't mean necessarily that the PM should actually go deploy that feature themselves with you know low understanding of the infrastructure components and things like this.
00:24:29.519 --> 00:24:33.200
But at least you can start speaking like more of the same language, right?
00:24:33.279 --> 00:24:42.400
Instead of having a PM that speaks documents and the engineers that speak code, at least we're getting closer and closer to like operating on shared surfaces.
00:24:42.640 --> 00:24:48.559
But you're right in terms of uh the challenge of like, okay, why did why did the coding agent behave like this?
00:24:48.720 --> 00:24:50.400
I told it to deploy my app.
00:24:50.480 --> 00:24:53.440
Why is it not working at scale for like 100 million users, for instance?
00:24:53.599 --> 00:24:56.480
Well, there's there's there's different parts there to the story.
00:24:56.559 --> 00:24:59.680
You have to have infrastructure to support that, you know, things like that.
00:25:00.000 --> 00:25:10.799
But um I think that's also part of our um role on you know, developer education and trying to understand like uh how do we teach more users about all of these tools?
00:25:10.960 --> 00:25:21.519
And when you zoom out on the number of users of ChatGPT today, like we're super fortunate that we have like a billion and 200 million, like 1.2 billion effectively, people using ChatGPT every week.
00:25:21.759 --> 00:25:31.680
I'm like, how many of those have not had their kind of like AGI
moment feeling yet of like, wow, I had no idea I could actually do that by myself.
00:25:31.839 --> 00:25:33.200
Like, I can actually.
00:25:33.359 --> 00:25:44.400
And so we have so much more potential to kind of lift up everyone with the power of these tools and be like, wow, I can actually be very creative here and turn my ideas into something tangible.
00:25:44.720 --> 00:25:47.759
Um, so I think to me that's that that's the exciting part.
00:25:47.839 --> 00:25:52.319
Uh, and goes back to also like the lines blurring between technical and non-technical people.
00:25:52.799 --> 00:26:04.000
Talking about lines, like, where do you draw a line within OpenAI when you talk about developer experiences, or is all what is user-facing in some way uh developer experience?
00:26:04.160 --> 00:26:06.480
I mean, we we mentioned uh dots, for example.
00:26:06.640 --> 00:26:06.799
Yep.
00:26:06.960 --> 00:26:10.240
Uh talked about them in the podcast uh um often.
00:26:10.640 --> 00:26:19.920
Um is that a product you are involved with and where you also talk about developer experience, or is it just somewhere outside your space?
00:26:20.160 --> 00:26:24.160
Yeah, no, it's a great question because even that for us is blurring a little bit internally, right?
00:26:24.319 --> 00:26:38.640
By definition, dots is not for just developers, because it's this idea that you can give anyone uh using ChatGPT or Codecs a proactive agent that's always on, that's very capable, that's doing things for them.
00:26:38.880 --> 00:26:44.079
Uh, even on this trip, my dot has been amazing almost every single day, doing something I would not have perceived.
00:26:44.160 --> 00:26:45.519
Like the model is really amazing.
00:26:45.599 --> 00:26:55.680
Like, you know, you land at the airport in a different city, and it tells you in three sentences like something critical that you have to know for like, I don't know, your route to the hotel or the venue for Dev Day or something.
00:26:55.759 --> 00:26:56.720
That's quite magical.
00:26:56.880 --> 00:27:01.599
But what's even more interesting, and where my team is involved, is that well, dot is for everyone.
00:27:01.680 --> 00:27:04.480
It's not just for developers, it's an agentic product.
00:27:04.720 --> 00:27:18.400
But as an engineer or as a developer, it's actually also giving you superpower because now you can start thinking of a dot as like a very high-agency senior engineer that I can actually delegate tasks to.
00:27:18.559 --> 00:27:30.799
So I can be on a plane noticing a tweet from you guys about, let's say, like an issue in codex, and I can be texting my dot now, and I can say, hey, can you look into this?
00:27:30.960 --> 00:27:34.880
I think there's an issue that was flagged by Dennis, like over here.
00:27:35.119 --> 00:27:47.680
Uh, and it will actually spin up codex task to look at the code base, introspect the issue, prepare the PR, ship the PR, and be like, all right, looked at the logs, found out the issue.
00:27:47.759 --> 00:27:52.640
This was isolated to this and that kind of customer, prepare the PR, prepared the fix, it's now in review.
00:27:52.720 --> 00:27:53.920
Uh, do you want me to look at it?
00:27:54.079 --> 00:27:55.839
Ping someone, do you want me to merge it?
00:27:56.000 --> 00:27:58.640
And that's kind of like my dot acting as an engineer, right?
00:27:58.720 --> 00:28:06.799
So it's this interesting thing where, like, even though the dot is for everyone, it will also transform the way we work as engineers.
00:28:06.960 --> 00:28:18.160
Because now you can think of your team as the humans around you, but also their dots as well in the picture, to be able to take on the work for them, take on the work for yourself.
00:28:18.319 --> 00:28:26.400
So internally, we have a team working on docs, uh dots, obviously, like product, engineering, research, marketing, and all of these functions.
00:28:26.480 --> 00:28:35.440
But we're also very well plugged in with them to uh to help them educate the world, especially with the more technical use cases of dots.
00:28:35.599 --> 00:28:48.960
Where I think, in my view, like dots and codecs are gonna start even blurring a lot more when you start not just prompting codecs to create a feature, but your dot is starting to proactively suggest things.
00:28:49.119 --> 00:28:53.440
It's like, hey, I've looked overnight at this feedback from customers.
00:28:53.599 --> 00:28:58.559
There's one theme that keeps coming back, which is this feature is not working the way they anticipated.
00:28:58.960 --> 00:29:02.960
So I just went ahead and proposed the first version of that feature for you.
00:29:03.039 --> 00:29:03.759
What do you think?
00:29:03.920 --> 00:29:05.440
It's like, wow, this is magical.
00:29:05.519 --> 00:29:08.480
This is already like thinking three, four steps ahead for me.
00:29:09.519 --> 00:29:09.759
Yeah.
00:29:10.000 --> 00:29:12.880
So yeah, can't wait for everyone in Europe to have access to this.
00:29:14.799 --> 00:29:20.559
Well, we certainly do have uh quite a big number of dots on our uh Slack uh channels already here.
00:29:20.880 --> 00:29:21.119
Right.
00:29:21.279 --> 00:29:25.440
And also, of course, because like business, enterprise accounts um can already access it.
00:29:25.519 --> 00:29:29.839
And then like hopefully in a few weeks, everyone, uh even consumers will have one.
00:29:30.000 --> 00:29:30.240
Yeah.
00:29:30.480 --> 00:29:32.640
Yeah, but also looking forward to these.
00:29:32.799 --> 00:29:38.480
I don't know, uh as far as I know, it's planned to have like enterprise dots as well, right?
00:29:38.640 --> 00:29:42.400
That can serve us in in specific roles for example monitoring and so on.
00:29:42.480 --> 00:29:44.640
I think that's that's gonna be really interesting as well.
00:29:44.960 --> 00:29:48.799
So yeah, I think it's gonna also change the yeah, it's gonna also change the developer experience.
00:29:48.960 --> 00:30:04.000
You could imagine
having on the developer experience side of things, uh, even on-call being reinvented, because the first on-call page is actually a dot that's able to actually decide what's going on and ask a human to wake up when the thing is ready, you know?
00:30:06.480 --> 00:30:15.440
So the kind of scenario you're outlining here is is very futuristic to some people out there, right?
00:30:15.599 --> 00:30:28.319
I feel like we see a certain fragmentation in the developer space where some people are very like AI positive, AI first, let's try all the things, let's early adopt everything we can and see where it gets us.
00:30:28.480 --> 00:30:37.839
And then there are some people who either buy company decisions, regulatory decisions, financial decisions, kind of feel like left out from much of this.
00:30:38.160 --> 00:30:50.799
And how does that impact your work where you basically want to share like all the goods with everyone, but there is a huge part of your audience that is not directly accessible or perceptible to that?
00:30:51.119 --> 00:31:03.039
Yeah, I mean, like um the adoption rate goes um slightly differently from like either like company size to company size, industry by industry, or even sometimes country.
00:31:03.200 --> 00:31:13.440
But what we've noticed is that the positive here is that unlike any previous technology change is being adopted across the board by everyone much faster on average, right?
00:31:13.519 --> 00:31:20.240
Like, for instance, like in the banking sector, how many years did it take for banks to have a mobile app?
00:31:20.559 --> 00:31:25.200
For some of them, it took a decade to have an iPhone app after the iPhone came out.
00:31:25.440 --> 00:31:30.880
Now, what we're seeing across the board is like adoption of AI is much, much faster than any previous tech.
00:31:31.119 --> 00:31:38.319
Now, you're right that like some startups and digital native companies, they are all reinventing their workflows with AI every week.
00:31:38.480 --> 00:31:48.160
For some others, they're still kind of like earlier in the adoption phase, kind of like um concerned maybe by like IT safety, security, deployment, or cost.
00:31:48.319 --> 00:31:50.799
And we're very sensitive to all of these chatter.
00:31:50.960 --> 00:31:53.119
That's why we're here to help them adopt.
00:31:53.359 --> 00:32:04.240
And I think sometimes like what I've noticed by all the companies I've spent time with is that success for AI adoption to me comes from two things that have to be true.
00:32:04.480 --> 00:32:17.440
One is like each company needs to have like one person at the highest level that gives that freedom and that top-down mandate to be like, hey, we're gonna reinvent the way we work on this company with AI.
00:32:17.759 --> 00:32:22.640
And giving that sense of direction and freedom that it is gonna happen.
00:32:22.880 --> 00:32:30.720
It also needs, though, like people that are curious, that are passionate, that are trying things, because you're
not gonna get everything right on day one.
00:32:30.799 --> 00:32:44.799
You have to reinvent the way you work with your engineers, the way you submit your PRs, the way you review each other's code, the kind of sensitivity to risk you have before merging to production, depending on the industry you're in, and so on.
00:32:44.960 --> 00:32:50.640
You need these bottoms-up of like passionate AI engineers that are kind of trying things out.
00:32:50.720 --> 00:32:53.279
And I think if you have both, you're absolutely gonna succeed.
00:32:53.440 --> 00:33:04.240
Because the challenge is like if you um if you don't have the top-down mandate, the the few engineers who are passionate might not have enough leeway to make change happen.
00:33:04.880 --> 00:33:15.119
But if you don't have the passionate, curious people and only a top-down mandate, you're not gonna try enough things and test different tools and actually be ready to reinvent the way you work.
00:33:15.200 --> 00:33:16.640
And so you're gonna need both.
00:33:16.720 --> 00:33:26.079
Um, but what I also noticed is that when I share or we share the stories of how we use our own technology at OpenAI, that actually inspires people a lot.
00:33:26.319 --> 00:33:32.000
Because, like just two weeks ago, for instance, I had a meeting with um a large company from Europe.
00:33:32.240 --> 00:33:36.559
And they were asking me, like, okay, well, you're clearly writing more code.
00:33:36.799 --> 00:33:44.000
Sure, we hear you, you're writing more PRs, but let's be honest here, do you feel actually more productive or not at OpenAI?
00:33:44.240 --> 00:33:46.400
I was like, well, let me give you some examples.
00:33:46.480 --> 00:33:56.640
Like, I mean, one of the biggest challenges we had recently was that ChatGPT.com was a code base that we started roughly four years ago now, right?
00:33:56.720 --> 00:33:58.640
Like before the launch of ChatGPT.
00:33:58.799 --> 00:34:01.359
And you can imagine it's like millions of lines of code.
00:34:01.519 --> 00:34:04.880
And by the time you start four years ago, you're not exactly sure where you're headed.
00:34:04.960 --> 00:34:08.480
You have no idea about the exact sequence of model capabilities.
00:34:08.639 --> 00:34:15.280
But meanwhile, we've seen this outstanding momentum of the ChatGPT codec kind of desktop app this year.
00:34:15.519 --> 00:34:20.159
That code base is much more modern, was made for like agentic delegation.
00:34:20.639 --> 00:34:24.880
And we knew we had to do something about these two code bases merging.
00:34:25.119 --> 00:34:30.559
But if we had tried to do this like three, four months ago, this was effectively impossible to do with AI.
00:34:30.639 --> 00:34:36.320
And we would probably have like taken like a year or so with our great engineers to do that change.
00:34:36.559 --> 00:34:42.480
It turns out with the help of Astra and Ultrafast, we were able to get this done in less than a month.
00:34:42.639 --> 00:34:43.280
And this is done.
00:34:43.360 --> 00:34:44.639
It's now in it's now in production.
00:34:44.719 --> 00:34:49.360
And we have also done this right before Dev Day, in the midst of everything else that we had to build.
00:34:49.519 --> 00:34:58.719
And when I tell people, like, well, if you don't believe this is like productivity, it's millions of lines of code, of very large code bases being merged and now deployed in production.
00:34:58.800 --> 00:34:59.360
And it worked.
00:34:59.440 --> 00:35:01.199
We've done that in less than a month.
00:35:01.519 --> 00:35:04.159
Could not have been done without Astra and Ultrafast.
00:35:04.559 --> 00:35:14.719
And they start thinking about their migration projects that they've been putting off because it's always something that you want to put off because you always have more important features to build.
00:35:14.960 --> 00:35:24.079
The best engineers are needed for that migration, but they're also much more needed for the other uh more uh business important features to build.
00:35:24.320 --> 00:35:39.760
And so when you start telling these stories, I think it really helps for them to reconsider wow, okay, so we need to take these things seriously because if we had considered some tools six months ago, the status has completely changed.
00:35:39.840 --> 00:35:44.000
Like what you can now accomplish versus six months ago is drastically different.
00:35:45.440 --> 00:36:00.239
Given that you already touched on model capabilities, I have a question about that because a few months, years ago, maybe, I was at a conference and I heard someone else from another lab, we don't want to mention names here, but I heard someone else speak.
00:36:00.480 --> 00:36:12.239
And what they said was basically we could stop research into future models right now, and we would still see months and years of new products being introduced that only just start to leverage what we already have.
00:36:12.400 --> 00:36:19.199
Because the products that we see out there today don't even begin to kind of like fully utilize the models that we have today.
00:36:19.920 --> 00:36:24.880
Now, at like approaching the end of 2026, would you agree with that?
00:36:24.960 --> 00:36:28.480
Or do you think we've gotten better at utilizing the models that we have?
00:36:28.559 --> 00:36:35.840
And is the the gap between utilization products and lab capabilities closing or expanding?
00:36:36.320 --> 00:36:36.559
Yeah.
00:36:36.880 --> 00:36:37.760
It's a great question.
00:36:37.840 --> 00:36:50.480
It's a feeling I have as well, because um, on one side of the of the picture, you can see there's safety that's paramount for companies like OpenAI.
00:36:50.559 --> 00:36:52.639
We care like immensely about that.
00:36:52.800 --> 00:36:54.000
This is our first priority.
00:36:54.079 --> 00:37:05.679
This is why there's this like conversation in the public about like pacing the frontier of research, doing it the right way, pausing a training run if need be,
because of that safety approach.
00:37:05.840 --> 00:37:24.159
But means in the meantime, you also see on the other side, on the product side, you see the like all of the amounts of ships that we've had at Dev Day in San Francisco, but also what the industry overall is able to accomplish, their velocity, the number of product experiences that they're able to deliver for their customers.
00:37:24.320 --> 00:37:27.760
And that fee, that that pace is only accelerating, right?
00:37:28.079 --> 00:37:30.880
Um and I believe that's still true.
00:37:30.960 --> 00:37:44.480
That like if you take a model like Astra, I mean, the incredible capabilities that are latent in there, I'm sure there are so many more product experiences we could still build today, even if those models were not changing.
00:37:44.719 --> 00:37:58.239
But meanwhile, what is also very interesting is that when you take the very best early adapters of models, I'm always surprised by how quickly they get used to this novelty and the next frontier of capabilities, right?
00:37:58.400 --> 00:38:08.480
When when Astra came out, this was this outstanding moment where like the community, for instance, realized, wow, it's so incredible at 3D modeling.
00:38:08.559 --> 00:38:15.920
And then for a week or two, we saw everyone experiment with Blender and all of these like 3D worlds, which was amazing to watch.
00:38:16.079 --> 00:38:22.239
And then, like, you know, this has been what, five, six weeks now since Astra, and people are ready for the next one.
00:38:22.320 --> 00:38:23.840
They're like, okay, let's bring the next one.
00:38:24.000 --> 00:38:27.119
We want to see what the next Filaquino frontier looks like.
00:38:27.360 --> 00:38:42.960
And so I'm also surprised by those early adopters and how quickly they're able to kind of poke at the model to find what it's able to do and discover new things that sometimes we were not aware of, like this, this like 3D uh capabilities that were emerging.
00:38:43.199 --> 00:38:47.119
Um But yes, at the same time, you could pause everything.
00:38:47.199 --> 00:38:52.880
And there are so many latent capabilities in these models that we have not fully grasped yet.
00:38:52.960 --> 00:39:02.320
I was seeing a tweet, I think, just this morning, about like multiplayer AI, for instance, and how we interact with each other with our own dots or equivalents.
00:39:02.400 --> 00:39:07.360
I mean, like, this is not quite just AI capabilities for that to happen.
00:39:07.519 --> 00:39:07.920
They're there.
00:39:08.079 --> 00:39:08.880
They're already there.
00:39:08.960 --> 00:39:13.199
But it's all about like reinventing the product, the UI, the UX.
00:39:13.280 --> 00:39:14.639
How do we work together?
00:39:14.880 --> 00:39:20.079
When does it make sense for my dot to write to you, or where is it going to be like poorly perceived?
00:39:20.159 --> 00:39:24.079
Because my dot should be telling something to your dot instead.
00:39:24.239 --> 00:39:36.079
And
like all of these little UI, UX workflows to figure out that's all product questions that have to be, you know, reimagined with trial and error a bunch of times with the current capabilities we have.
00:39:36.159 --> 00:39:39.920
So I think there's like really both uh happening at the same time.
00:39:40.239 --> 00:39:48.000
And even if we pause, or rather, we pace the frontier because we're never going to pause, the progress is very rapid and will continue to be rapid.
00:39:48.239 --> 00:39:54.719
But even if you pace the research, you will continue to see an acceleration of what the products can deliver.
00:39:55.760 --> 00:39:56.159
Yeah.
00:39:56.400 --> 00:39:58.559
So we're approaching the end of our time here.
00:39:58.639 --> 00:40:03.519
And before I ask my last uh question, I want to ask Dennis if he has any questions remaining.
00:40:05.039 --> 00:40:05.360
No.
00:40:05.679 --> 00:40:08.079
Um I mean I mean, yes and no.
00:40:08.159 --> 00:40:12.000
I I think we could talk endlessly about those topics because everything is interesting.
00:40:12.159 --> 00:40:16.000
Uh, but yeah, no, I think uh we got some got some great insights.
00:40:16.559 --> 00:40:17.440
Thank you for having me.
00:40:17.679 --> 00:40:20.239
Okay, then my last question would be this.
00:40:20.400 --> 00:40:25.119
Um From what I understand, you've been with OpenAI for like three years now.
00:40:25.840 --> 00:40:34.239
And you already mentioned in the very beginning like how little of uh developer experience there actually was when you first started and how much has changed since.
00:40:34.559 --> 00:40:40.880
What are some of the remaining opportunities that you still see for improvement in developer experience?
00:40:43.119 --> 00:40:44.000
That's a great question.
00:40:44.159 --> 00:40:44.719
So much.
00:40:44.960 --> 00:41:00.480
Um ultimately, I think developer experience is like everything that needs to happen for developers to have this like magic feeling when they're using our tools, right?
00:41:00.559 --> 00:41:18.960
And to me, um maybe in the past, developer experience meant this very precise playbook of like API design, SDK, docs, sample applications, demos, you know, all of that had to be ticked, you know, and we've tried to do as great as possible for all of that playbook in the in the era of Stripe.
00:41:19.360 --> 00:41:24.079
But now it's like everything that um takes your idea into a product.
00:41:24.239 --> 00:41:28.239
So that's why, even to me, ultrafast is part of the developer experience, right?
00:41:28.320 --> 00:41:35.119
If I'm pairing with the model and if it's a frontier model and I'm in the flow, I'm feeling like I want to iterate on something.
00:41:35.199 --> 00:41:40.400
I'm not sure what the end state should look like, but I have this nebulous idea.
00:41:40.559 --> 00:41:42.480
I need the model to pair with me.
00:41:42.639 --> 00:41:46.159
Well, now working with ultrafast is part of having this amazing DX.
00:41:46.239 --> 00:41:49.920
It's like, wow, I I can in just a few seconds I see something appear on my screen.
00:41:50.079 --> 00:41:53.679
I want that to be like now the way I interact when I'm in the flow.
00:41:53.920 --> 00:42:00.400
Uh, but sometimes where I have to do a complete rewrite of a back end, I want to delegate that to codex in the cloud.
00:42:00.559 --> 00:42:04.559
I no longer want to have my laptop open when I walk around for four hours.
00:42:04.800 --> 00:42:07.519
This should be completely seamless in the cloud.
00:42:07.679 --> 00:42:09.840
I can check on that later from my phone.
00:42:10.000 --> 00:42:12.000
I don't need to worry about any of this.
00:42:12.239 --> 00:42:15.519
The harness that's the codex harness that's open source.
00:42:15.920 --> 00:42:27.360
We can let uh developers introspect it, look at it, but it's also the same harness we run locally on your codecs as in the cloud and giving Codex Cloud its own computer.
00:42:27.519 --> 00:42:29.679
That's I think what we've reimagined recently.
00:42:29.760 --> 00:42:31.119
We're in the early phases of that.
00:42:31.199 --> 00:42:33.760
And I think that's kind of also part of the future of DX.
00:42:33.840 --> 00:42:47.760
It's like maybe my best development machine in the future is just my phone, because I have Codecs in the cloud with the same harness, same tool, same context, same code base, and it has its own computer in the cloud to check its own work.
00:42:48.000 --> 00:42:49.760
That's also part of the DX, right?
00:42:49.920 --> 00:42:53.599
And then on top of that, it's the proactivity of the dots.
00:42:53.840 --> 00:43:07.920
The dots actually being an extension of me, being the teammate that I have always on in the cloud, working with me and for me on these applications that I'm building, coming up with the new ideas, making the market research about what competitors are building.
00:43:08.159 --> 00:43:20.320
And so to me, it's it's fascinating to see that the DX is like basically the model, the speed at which we're delivering them, also the cost and the token efficiency at which we're delivering them.
00:43:20.480 --> 00:43:27.840
Because imagine we have these frontier, outstanding models, but they're expensive to run and you're not able to afford them.
00:43:28.000 --> 00:43:32.719
That's impactful for the development experience because it means the DX cannot be the same for everyone.
00:43:32.880 --> 00:43:36.320
So we're trying to also push on token efficiency for that reason.
00:43:36.559 --> 00:43:51.119
And so it's all of the playbook we knew with the coding agent friendliness on top of that, to make sure that like everything that was in your way before, writing code by hand and all of this, that's all gone because it's taken for granted.
00:43:51.199 --> 00:43:54.000
Then the coding agents will take care of most of that for you.
00:43:54.159 --> 00:43:59.760
But it's also making sure that you can work effectively, either by yourself when you're in the
flow or with your team.
00:44:00.400 --> 00:44:08.320
Everything going from like how do you do codec codec code review or how do you harden your infrastructure with codec security?
00:44:08.400 --> 00:44:29.119
Like all of these pieces to me have to all fit together in the sense of like, as a developer, it should be as seamless as possible to not just take my idea into a product that my customers love, but also making sure that I have all of these tools to make it better every day, more secure every day, more cost effective every day, and all of that.
00:44:29.280 --> 00:44:36.079
So so much more progress to do on each of those, but we can start seeing the convergence of all of these bricks together now.
00:44:37.199 --> 00:44:39.039
Well, uh, thank you very much.
00:44:39.199 --> 00:44:40.559
Thank you for taking the time.
00:44:40.800 --> 00:44:44.559
Uh a bummer that we didn't get to talk about Jolly OS and Jolly Book.
00:44:44.639 --> 00:44:47.840
I saw that on your LinkedIn and I didn't even realize that was you too.
00:44:47.920 --> 00:44:50.559
So maybe we need to have a follow-up at some point in the future.
00:44:50.880 --> 00:44:53.119
We could, we could talk about that one for a long time.
00:44:53.280 --> 00:45:04.960
But um, you know, it's funny because like you now that you mention it, uh I've actually been reflecting on that recently because I've worked so for context for the audience briefly, this was my first company back 20 years ago.
00:45:05.119 --> 00:45:09.920
I was building an operating system for the cloud, and at the time, like HTML5 was very nascent.
00:45:10.239 --> 00:45:18.079
Um, the Chromebooks did not exist yet, but this was this idea that you could buy an inexpensive laptop and just boot it to the cloud.
00:45:18.159 --> 00:45:23.199
And the cloud would effectively be like where your apps live and where your data lives and everything.
00:45:23.599 --> 00:45:26.159
Kids ask your parents about netbooks.
00:45:26.559 --> 00:45:27.599
Yeah, yeah, yeah.
00:45:27.840 --> 00:45:35.920
Uh netbooks are, yeah, we we used to uh they were like all the rage for maybe two, three years because the iPad did not exist.
00:45:36.239 --> 00:45:40.079
Multi-touch tablets were not really a thing yet.
00:45:40.320 --> 00:45:45.679
But we had these small laptops for 300 or so euros, so very inexpensive.
00:45:45.840 --> 00:45:52.719
They were a very small phone factor, but they were shipping mostly with Windows XP, which was already no longer maintained at the time.
00:45:52.880 --> 00:45:57.119
And so clearly there was the need for a new OS for this new generation of cloud.
00:45:57.360 --> 00:46:11.280
But uh the reason I was reflecting on that recently is because when you when you see the progress with AI and the fact that like anyone can create software so quickly, but also this era of like personal software.
00:46:11.360 --> 00:46:14.000
Like imagine you have an idea for an app for yourself.
00:46:14.239 --> 00:46:21.519
Well, you would never build such an app back in the day because it would be like too much uh work hiring some developers and such.
00:46:21.679 --> 00:46:29.440
But now, like a few prompts of pairing with the model like Sol or Astra, and you can just get your app up and running just for yourself.
00:46:29.679 --> 00:46:36.400
So the idea of uh an OS that's kind of reinvented for each individual is now in sight, right?
00:46:36.480 --> 00:46:43.840
You you could re-reimagine this like Joliclad, Jolie OS version again, uh, but with AI now.
00:46:44.000 --> 00:46:47.280
But yeah, we can save that for the for our next discussion.
00:46:47.760 --> 00:46:48.480
Will do.
00:46:48.719 --> 00:46:49.840
Okay, uh, thank you so much.
00:46:50.000 --> 00:46:51.679
Thank you for taking the time today, Omen.
00:46:51.840 --> 00:46:53.199
It was really nice talking to you.
00:46:53.360 --> 00:47:00.559
Um, and maybe we can repeat this at next year's uh developer event in whatever cities uh will find you next.
00:47:00.880 --> 00:47:01.360
Amazing.
00:47:01.440 --> 00:47:03.119
Well, thank you so much uh for having me.
00:47:03.199 --> 00:47:07.840
Sorry, we meet each other in per in Berlin, but I'm glad we could uh we could do this today.
00:47:07.920 --> 00:47:10.159
Thank thanks so much, and thanks everyone for listening.
00:47:10.480 --> 00:47:10.960
Thank you.
00:47:11.280 --> 00:47:13.280
Thank you, everyone, and bye bye.