How AI Rewrites Product Strategy | Romain Huet (OpenAI)

Good afternoon everyone. It's fantastic

to be back at Slush. I'd like to take a moment and ask you something in your head. Like think about the startups

head. Like think about the startups you're building now. How much of that startup would have been possible two years ago and how much are you rethinking your ambition, your road map

every day as the models get better.

When I started my first company, everything was about getting to the next fundraising milestone or hiring the next best engineer. The mode back then was

best engineer. The mode back then was simply can I build this product and how long will it take? But today the constraint has completely flipped.

building is actually much easier. So the

real challenge is figuring out what to build and what are the lasting modes today for the companies you're building.

And we've entered a new era for builders. AI is rewriting the product

builders. AI is rewriting the product strategy as we go. And that's exactly what I'd like to explore with you today.

I'm Roman and I lead developer experience at OpenAI. And my favorite part of the job is really spending time with talented founders like all of you.

uh and and really think about like what the future will look like for us.

Developers and builders have always been at the heart of OpenAI's mission. You

cannot bring the G into this AGI we talk about and that energy is a big part of why I'm here today. I actually started my career as a founder and CTO myself.

And I want to share a little bit of that story with you today.

So 17 years ago, yes, uh my co-founder Tariq and I decided to uh build an operating system for the cloud. And but

to give you a sense of that moment, this was 2008. So the iPhone had just come

was 2008. So the iPhone had just come out. The iPad did not exist yet, but

out. The iPad did not exist yet, but people were buying these very inexpensive small laptops called netbooks at the time. They were quite underpowered. They were meant to get you

underpowered. They were meant to get you online, but they were running Windows XP and some outdated software. And so given the pace of change at the time with HTML

5 and the web platform, we thought there had to be a better way. So we decided to build an OS from scratch. And the entire vision kind of fits here actually on

this sketch. We quickly discovered how

this sketch. We quickly discovered how ambitious that was and frankly a little insane to build that from Paris. At the

time almost no one knew what a startup even was. There was just like a handful

even was. There was just like a handful of us trying to will something into existence.

Capital was scarce, talent was scarce, and frankly even optimism was scarce, and everything about building was super slow. Like I spent a year heads down

slow. Like I spent a year heads down compiling Linux kernels and drivers just to get to the first prototype of this OS. And then when we got our 4 million

OS. And then when we got our 4 million euro series A, it felt like somewhat miraculous and gigantic at the time only to get us to the next step to achieve our ambition.

But every decision was quite heavy because you know you can't change course multiple times. Everything is going to

multiple times. Everything is going to take some time. And fast forward to today, it's how outstanding how everything has changed. I'm guessing

like a couple people in a few weeks could probably build what we've taken years to build. And that's the new reality.

And in this new reality, speed is not the mode anymore. It's kind of table stakes. We've shifted from a world of

stakes. We've shifted from a world of scarcity to abundance. Access to capital is much easier than it was. You now have powerful AI teammates at your disposal

for anything you're trying to to build.

And fundraising, connections, headcount, all of these old modes, they don't really protect you anymore in this new world.

So, we like to think that AI has changed the physics of building a company. And

speed is now only the starting line for all of you founders. And I'm sure the crowd here at Slush is very forward-looking, which is awesome spending time with all of you. But

sometimes I meet founders that are not quite yet on that starting line and you know being French as you can hear from my accent I really care about the success of European founders and all of

you setting the example. Sometimes I

sense some skepticism or hesitation as to the pace of change and the steep curve in those model capabilities. Some

developers might worry for instance that like AI is not writing quite perfect code and they could still do better or maybe it's going to take some time for them to review. That might have been true a year ago, but the progress since

has been radical.

So, as a founder, your job is to figure out how to create speed and turn that into an advantage. AI is no longer just a tool. It can really uh be the

a tool. It can really uh be the teammate, the the best teammate you've ever had. Frankly, it can help you with

ever had. Frankly, it can help you with anything about company building from refining your brand, finding prospects, obviously writing code, but also jumping into customer support when you need that

as well. And if you're not harnessing AI

as well. And if you're not harnessing AI today, you're not you're kind of competing with both hands tied behind your back because the competitors out there, they will know how to master AI.

So we like to think about how founders can really harness that AI the best they can. And today I'd like to zoom in on

can. And today I'd like to zoom in on the one part where the progress has been the most significant in the past year and that's software engineering itself.

We've talked a lot at OpenAI about 2025 being the year of agents and of course we've launched many of those like including deep research taggy agents but we've also launched codeex which I'll

talk about in a minute. Software

engineering is one of those agentic workflows that have shown tremendous success this year. So if you're a technical leader this next part is for you. It's really incredible when we look

you. It's really incredible when we look back just a few years ago at the pace of change. This was the state of software

change. This was the state of software engineering just a few years ago. A

single line autocomplete to kind of get you started faster as you wrote code.

Then we moved on to the ability to write some files, some functions, and that was quite good. It helped you save some

quite good. It helped you save some time. But this was still very much like

time. But this was still very much like the beginning. Fast forward to 25, the

the beginning. Fast forward to 25, the world looks completely different now, especially since this summer. We

launched GPT5 and just last week GPT5.1 two of the very best models out there when it comes to software engineering and coding and top coding startups like

cursor versel windsurf are using GPT5 to change how software gets written using their tools but I think writing code is essential but the best part the bigger

opportunity of all of this is to fundamentally rethink how we design and build software and I talked to a lot lot of engineering leaders, CTOs of medium to large

companies as well and sometimes one of the first questions they ask me is how many lines of code is AI writing and of course AI can now help you write code but I think that's no longer the right

question because AI can excel at understanding code refactoring code troubleshooting bugs and that all of that with remarkable precision like that

could be achieved by a software uh engineer that's very senior in their career And it's not really about the raw intelligence of the model anymore that you can see on benchmarks. It's also

about the ability for the models to perform really well in the real world and reliably use tools. And that's what we refer to as aic coding. The ability

for the model to verify their work and access these tools. And that's not some distant future, by the way. This is

really happening now. And AI can now take a lot of parts of the software engineering workflows that used to be very tedious or very costly. Whether

it's like understanding a codebase, tracking down some bugs, writing tests, big refactorings at scale, stress testing system ideas, and so on and so forth.

So I'm curious, raise your hand if you've heard of Codeex before.

Wow, many of you. That's awesome. Uh

maybe keep your hand up if you've used it in the past week.

Okay, we have a few fans over here.

That's awesome. Cool. Uh, well, earlier this year, for those who don't know, we introduced Codeex, and that's OpenAI's software engineering agent designed to work alongside developers. And it's

evolved, it's evolved quite quickly since we launched it back in May. And

now it's become kind of a teammate able to work alongside the engineering team for complex task, complex workflows. It

integrates in every part of the the workflow for developers. So, we like to think about being where developers work, whether it's in the ID, the terminal, or in the cloud or on GitHub.

And we've released a ton of features this year. I won't cover like all of

this year. I won't cover like all of these today. That's not the goal. But

these today. That's not the goal. But

just to say, it's been really impressive for me to watch the Codeex team actually build so fast using Codeex themselves.

They have a relatively very small team, but they were able to accomplish so much. and Codex is now running a new GBT

much. and Codex is now running a new GBT 5.1 Codeex model. A model that we actually uh purposely train for codecs and for agentic coding. So what that

means is that this model excels at taking very long and large tasks like we've seen it work for instance for 7 hours long on one big refactoring and

get it right in one shot.

And one of the key metrics we look at uh on the codeex team to make sure the product and the models are actually useful for developers is the number of daily messages kind of the conversations

and tasks that take place every day with the model. And since early August that

the model. And since early August that number is up 10x across all of Codex and that's not just individuals using codec.

The reason I bring this up is because we see a huge change in how much more useful codeex has become for developers really pairing with it every single day

to do their work.

And when you think about what that means for a founder, it means kind of creating leverage to be more ambitious in what you can accomplish. It means like taking your 2026 road map and ideas you wanted

to ship and bringing them back to like the next few weeks uh before we close the year or potentially pulling 27 entirely to 26. So shipping these ideas

that maybe felt out of reach or and that's what you can do when you have 5x 10x leverage the whole tempo of your product building velocity uh is

obviously increasing. Imagine a bug gets

obviously increasing. Imagine a bug gets reported by a customer. Fire that off to Codex to fix it. Or imagine you have an idea that's kind of stuck in your backlog and it's going to take a few weeks for your engineers to get to that.

Just fire it off to Codex right away.

Codex will take the first step and at least your engineers will be able to iterate on that. And by the way, this is how we ship at OpenAI. We use codecs pretty much everywhere uh we write

software and 100% of the pull requests that we create at OpenAI are also reviewed by Codex and that's catching bugs before uh they hit production.

So anyways, if you have trouble sometimes keeping up with everything OpenAI is shipping recently, you can probably blame Codex for a lot of these.

So none of that will replace the judgment and creativity of great engineers, but it removes the drag and it creates the speed for your team. So

instead of telling you more about Codex, I'm going to show you a few live demos of what you can accomplish when you have aentic coding at your disposal. So I'm

going to go over to the laptop here and let's switch to the demo laptop.

Amazing.

So the first thing I wanted to show you uh imagine this is a startup I'm building called wonderless imagine I'm building a travel application for consumers very simple well here we have

a basic assistant and I can interact with it to say for instance where is slush >> slush 2025 me over here >> an exciting event for founders and entrepreneurs

>> imagine I jump over to slack and it turns out uh someone mentioned hey like the app actually does not work well on mobile and another teammate mentioned it's also not working in dark mode

either. Well, what could I do in this

either. Well, what could I do in this case? I can simply tag Codex on Slack to

case? I can simply tag Codex on Slack to ask to uh create that for us. And if I jump over to the task, this is what Codex accomplished in 3 minutes. Not

only did Codex fix that, but also took a screenshot of its work. So, you can see now the app working on mobile and dark mode beautifully. And all you have to do

mode beautifully. And all you have to do is click a button to create a PR. So it

really rewires uh how you think about coding, right? Because now a bug comes

coding, right? Because now a bug comes in, you can just send it over to Codex.

But now imagine like speaking of multimodality and this vision capabilities. Imagine I have an idea for

capabilities. Imagine I have an idea for a new feature for my startup. And let's

say that I want to go back to this app and I want to add a new um screen. So

maybe I call it like a travel log. the

ability to have like every trip I've taken, my upcoming trip, some interesting stats, and that's like a sketch that I've made very quickly to get the the ideas flowing, but I'm not

exactly sure where to go. Well, what I can do is uh take a a photo of this, send it over to Codex here in the cloud, and this is one of my favorite features

called best of end. What that means is that Codex can actually take a pass at this task multiple ways, in this case, four times in concert. And you can see it had different approaches to the

design. So as an engineer, I can now

design. So as an engineer, I can now like have a few ideas for where I want to get started. Pick the one I like most. Maybe one or two. I think I like

most. Maybe one or two. I think I like one a lot. And then I can pull that down to start coding. So that's really useful. That's codeex in the cloud with

useful. That's codeex in the cloud with best of now. One thing that I'm very excited about also is the integration with tools. So here for example you can

with tools. So here for example you can see the Figma design that were made for um this this app wonderust that screen exists but for instance we've never got

around to implement this screen which the ability to like you know plan a trip how would you go about that well the interesting thing with codeex is that it plugs with all the MCP tools I would

bear so here for instance I can select a component right for this one and if I go back to uh VS code here I have my project open codeex is running on the

right side of the screen and take a look at this. If we go back to this prompt, I

at this. If we go back to this prompt, I said, "Hey, go fetch the code of the selected component on Figma, make that a component and integrate it so we can see

it." And that's really the magic once

it." And that's really the magic once again of aic coding because actually the model went through u the chain of thought here, pulled the code automatically from Figma, gathered a

screenshot to check the work, and sure enough made a component. And so now if I go back to uh what it looks like, boom, we have a component that's already implemented. I didn't have to write a

implemented. I didn't have to write a single line of code. I can now have my tools directly fetching from Figma. And

that's true for any kind of tool.

Imagine you're tracking your workflows in linear for instance. You can pull the context from there or maybe you have documents on notion and so on. So that's

the ability to pull things directly from tools.

Now the other thing I wanted to touch on which I find very exciting and I hinted at that earlier is code reviews. So here

for instance imagine I'm working on GPOSS our open source um repo linked to the open weight models and let's say that I want to add a new thing here very

simple task codeex was able to review it just a thumbs up emoji because there was nothing special to report but imagine here this more interesting PR I tackled

two things in this PR but right away Codex found out that there's maybe something here I should pay attention to because the streaming code might now no

longer work in every possible scenario.

Well, sure enough, I ask Codeex, hey, can you please fix this? And it can take a pass at that. So, for engineers who are just getting started with Codeex, I often recommend to them to use that as a

starting point.

So, that's Codex code review. Now, one

thing I wanted to show you and that's like very new and I think it's hiding somewhere here. Uh there we go. Uh,

somewhere here. Uh there we go. Uh,

so I previously made a demo when 01 came out to interface with one of those little drones, but um, this was the very beginning of how AI coding worked. And

what I did with 01 at the time was a very simple interface that looked like this to connect with one of those little drones. But before I explain everything,

drones. But before I explain everything, I want to go back to uh, codeex here and actually send a task right away. So, I'm

going to send this task and we're going to see what happens next. Uh, I want to interface with such a game pad which I have never done before and we'll trust codeex to hopefully figure this out. But

what it's working on this because I'm guessing it's going to take like a minute or so. I want to go back to the interface and talk a little more about multimodality.

So, what we'll do here is we're going to turn that on and I'm going to be able to connect to it over Wi-Fi. So, we're

going to take a little bit of a beat here for it to turn on.

And I think what's very important as you're building your company is not just to think about how to leverage AI as a tool to build, but also how to leverage the new AI capabilities within your

product. So in the case of Codeex, I

product. So in the case of Codeex, I showed the vision capabilities, right?

And those vision capabilities make the tools work better. But what if you were able to use vision and speech to speech in your own application? So that's the thing I wanted to show you today. So

here what we'll do is like now that this thing is on, we're going to let Codeex is still running over here. Perfect. And

I'm going to start the server to connect to the drone. Great. And we're going to start the video stream here. Great.

Perfect. And so now what I'm going to do, I'm going to enter voice mode and hopefully we'll hear what's happening.

>> Hello.

>> Hello.

>> Hi there. What would you like me to do today?

>> Yeah. Could you tell me a little bit about your capabilities?

>> Sure. I can take off, fly around, and help you with video analysis of the surroundings.

>> That's great. Actually, speaking of video, what do you see now?

>> I'm going to take a look for you. Give

me a moment.

I see a large indoor audience focused on a stage and there's a digital timer showing 9 minutes and 9 seconds. That's

right. I need to uh keep going. Exactly.

Actually, can you take a look at what you see and turn that into a money painting?

>> What a creative idea. I'm on it.

>> Great.

>> What?

>> Uh yeah, you can do it in the style of a money painting.

Got it. I'll transform the current view into a Monae style painting for you.

>> Perfect. I'll quit the voice mode and let work on it. Thank you.

>> All right. So, we'll let work on that.

It's going to take a minute. The more

interesting thing I wanted to look at now is what's happening with our codeex task.

So, let's take a look at what happened here. I asked Codeex as I two minutes

here. I asked Codeex as I two minutes ago to create an interface to wire this thing up and literally I've never done this before. Uh but it turn sounds like

this before. Uh but it turn sounds like it it made a plan with three steps to review the code, move the command animation and as we hear like as we get here it sounds like in 1 minute and 53

seconds it got it done and finally did everything right. So we even have the

everything right. So we even have the mapping that it decided to do. So we

have to put this to the test and go back here. Let's see if I The game pad is

here. Let's see if I The game pad is connected. Let's see. There we go. We

connected. Let's see. There we go. We

can now control this thing with a game pad, which was not possible 2 minutes ago just because Codex was able to figure everything out for me. So, that's

pretty awesome, right? And we can just land.

And what's really cool here is that I didn't have to write any code for all of this. And sounds like the money painting

this. And sounds like the money painting just appeared for our lucky uh front row uh attendees. Awesome. So, let's go back

uh attendees. Awesome. So, let's go back to the slides, please.

I hope this gives you some ideas of how on one side you can leverage codecs and coding tools to build faster, but also on the other side how you can think of multimodality, speech to speech, vision,

image generation, all in concert to create great products. All right. So now

once you unlock that speed, the real question is what are you going to choose to build? When everyone can move fast,

to build? When everyone can move fast, the differentiation will come from your choices as founders. What do you decide to focus on when you what you cannot compromise on? And since AI is changing

compromise on? And since AI is changing everything so fast, I want to share some ideas with you today.

You know, I love the program, you know, startup advice that's now timeless, you know, make something people want. But I

also like how Guiamo from Verscel kind of like had I did a twist on it recently in an interview where he said it's like build what users want and then iterate towards greatness and raise their

expectations along the way. And I really like that framing cuz when you have a rapid iteration it's essential but you also have to bring your customers along uh to make sure that you actually make

progress towards their goals. And you

know at Amazon Jeff Bezos mentioned like people did not really ask him what would not change as opposed to what will change. And for him it was lower prices

change. And for him it was lower prices it was you know faster deliveries. And I

think that's important to remind yourself that great products aren't just built fast. You also have to build

built fast. You also have to build towards what will stay true for a long time. And some of the biggest mistakes I

time. And some of the biggest mistakes I think founders make is that they're building shallow features that the AI models could actually take on in a few months or years from now. So the way you

differentiate your product is by building parts of your product that the models can't copy. And for that I'd like to share uh from my talk with like hundreds of startup founders some of the

most successful some of the most successful ones tend to kind of fall into four categories of modes and I'd like to share some of that with you today. Let's unpack each of these on the

today. Let's unpack each of these on the screen. The first one is deep

screen. The first one is deep integration. Too many companies still

integration. Too many companies still treat AI like a chatbot feature, a bolt-on that they have on the side and it let them it let them tick the box like we have AI but it's not really

fundamentally changing the workflow.

They kind of expect users to like leave what they're doing go to something else that may or may not have the context and that's just surface level AI. It's not

really deeply embedded in the workflow but take ramp for example a leading fintech in the US. What I love is that they completely reimagine the spend management from the ground up with AI.

So when you open ramp, you're not just seeing expense reports. What you're

seeing is like an AI working behind the scenes constantly to make sure it proactively finds savings, benchmarks the cost and so on. So it's making every employee a little more productive. And

that comes back to this new mode, right?

Anyone can copy obviously your interface in the weekend but no one can really copy the deep understanding that ramp has about how building that product uh should be.

The next one is taste and point of view.

I think it's a critical one in an age of AI. Take clay for example. They've built

AI. Take clay for example. They've built

one of the most elegant products in the space and they give everyone even the one with no technical background the ability to uh actually go out and find

prospects things that would actually be requiring Salesforce admins or engineers now you can spin up with them like automatic prospecting flows and that's what it means to truly understand your

users right like removing that friction that you didn't know existed before uh and your users did not know that was holding them back and I I think taste is really critical in this era of AI and

that's really like your point of view using things that we just saw like speech to speech and how you bring that to your customers.

Then there is trust and community. I

think the best spongers build products that get smarter the more people use them. And when people trust your product

them. And when people trust your product and really enjoy using it, what's beautiful is that they extend it. They

become part of your R&D, right? And you

see that really beautifully with lovable. They've built one of the

lovable. They've built one of the strongest brands for developer communities. They're running hundreds of

communities. They're running hundreds of hackathons in the world. And every one of these event is a great feedback loop because they learn from how users use it. They turn that into insights to make

it. They turn that into insights to make the product better. And when your users love and trust your product, they become co-creators, right? They scale your

co-creators, right? They scale your products in ways you did not think about and they become part of marketing for your company.

And finally, there's specialization. I

think that's a very important one as well. The best founders know that you

well. The best founders know that you can't win every battle. You have to pick precisely where context is the most. And

code and prompting can't get around yours of living the customer pain and having that obsession for the problem you're solving. And if you take Legora

you're solving. And if you take Legora for example, they've built one of the most trusted AI platforms in the legal world. But what's fascinating is that

world. But what's fascinating is that they did not come from a legal career.

However, it came from doing the work.

They spent countless hours inside law firms with partners, with associates to understand how do they work and why do good enough products typically not get

adopted as much in this setting. And

that's really specialization. They

really learned everything about liability, trust, how they operate. And

this product now is know like knows the domain so well. It's extremely hard to replicate by anyone who has not put the same level of work as they have. And so

when expertise is your edge, you're no longer threatened by, you know, a copycat or imitation because, you know, someone may copy your code, but they can't copy that obsession.

So zooming out, there are clear patterns that emerge across these companies. And

I think that means as a founder, you can kind of like rephrase them as questions to ask yourself, right? One is am I embedding AI so deeply that it becomes

really like reshaping my product? Do I

have a real point of view like a product that feels like it was designed by someone who truly knows the users?

Am I building enough trust and community so that it extends my road map? And

lastly, am I going deep enough in my domain so I understand really so well that no one has the same understanding?

And if you can check one of these, you're in a great shape to build the modes. If you tick multiple, you're just

modes. If you tick multiple, you're just compounding and harnessing the power of AI to go even faster. When I think back about my own journey with my first

company, Jolly Cloud, it's wild to think that like what we used to do back then could be done in just a few weeks, right? And we learned firsthand that

right? And we learned firsthand that having great technology alone was not enough. You also have to, you know, nail

enough. You also have to, you know, nail the timing and have this alignment with the real world use cases. And when I think back on companies I admire, Journal General Journal Magic is also

coming to mind. To me, it's one of the most talented collection of individuals in Silicon Valley. They spent years meticulously building incredible technology. They actually invented what

technology. They actually invented what became the iPhone two decades before and yet they were like so early and for them everything was so hard to build because they didn't have access to the same

tools. And so seeing their story kind of

tools. And so seeing their story kind of reinforce what I've learned from mine.

Groundbreaking tech alone is not enough.

You really have to iterate constantly with your users, with their needs, perfect your timing, and validate your direction constantly. So today, the

direction constantly. So today, the question is no longer how I build this, but how do I differentiate in this new world? And today, the smartest founders,

world? And today, the smartest founders, they're not racing against AI. They're

really harnessing it to build better and more resilient products. So, if there's one thing I want you to take away from today, it's this. The pace of change will not slow down. It's only going to

accelerate. And your job as a founder is

accelerate. And your job as a founder is to learn how to master AI, master the tools, create speed for your team, and use that page to your advantage. I truly

believe that in this new era, the winners won't just be those who build faster, but also the ones who learn quicker to compound that and compound

their advantages. So we at OpenAI love

their advantages. So we at OpenAI love founders. We want to be partner uh on

founders. We want to be partner uh on your journey. That's why we have a

your journey. That's why we have a global team and presence here in Europe.

Uh I would obviously personally love to help and keep meeting many of you. Uh

you can find me online uh right there on the screen. Uh our goal is to build uh

the screen. Uh our goal is to build uh you know amazing products alongside all of you and help you build long-lasting companies. So with that, thank you so

companies. So with that, thank you so much and I can't wait to see what you build. Thank you.

build. Thank you.