Olivier Godement: Head of Product, Platform at OpenAI

0:00Hello and welcome to Scaling Europe show. I'm Seb Johnson. I'm here with Olivier Godement, Head of Product Platform at OpenAI. How are you doing today? I'm doing great. Thank you for having me, Seb. It's my absolute pleasure. You're in London at the moment, is that right? Yeah, I'm in London this week. I usually live in San Francisco in the US. So what brings you to London? Coming to meet a bunch of startups, enterprises, cool businesses, building stuff with AI. are, you know, good way for me to stay in touch with the European scene. Yeah, absolutely. I've seen that you guys have been, like, running some amazing events across Europe.

0:35You had Dev Day Exchange recently. Can you touch a bit on that event, why you did it, kind of what was announced there? Yeah, so we have, like, every year our annual product conference back in SF, which is called Dev Day. And we've been trying to do, like, you know, smaller, like, more intimate, like, events, like, in key, like, you know, startup hubs across the world. so there was one in London and there were a few others in different cities and the goal here is to both showcase our new technologies so we made a bunch of announcements which I'm happy to discuss but also to just connect with Gelders I've been very impressed in the past couple of years frankly by the amount of energy and good ideas that are popping up in multiple hubs and so it's a good way for me and the rest of the team to stay super close to what are people building, where are they struggling, where are they seeing opportunities.

1:32And how does that, do you notice a difference in somewhere like London to San Francisco? Less and less, frankly. I'm from France initially. I spent a bunch of time in the European startup scene since 2012. Historically, there was a lag of business model, new technology adoption uh that's no longer the case frankly uh at that point i feel like the the startups coming out of london of paris or you know berlin are frankly like equally um mature when it comes to technology adoption new product design new product form factor um so uh yeah i wouldn't be able to tell that a big difference and is that the same across like different business types do you think?

2:22I think startups is good to hear, but in terms of maybe like enterprise adoption, you know, one of the things that I hear a lot from founders who are selling in San Francisco to selling in London is that selling in America is much easier because you have enterprises who are much more willing to try and test new technology. I'm curious to hear, is that something that's different or has that been your experience in OpenAI? Somehow a little bit, yeah. I think the main difference with the US is there are a bunch of what we call like digital natives which means like enterprises that were founded in the last like 15-20 years think about like companies like Salesforce for instance you know they're a pretty big company like they're an enterprise but their like DNA is fundamentally like very technical like very technology oriented and that's been like somewhat like less the case like you know in Europe I expect it will become the case like inevitably like you know some of the startups are going to become like large companies but I would say like the ratio of like enterprise between like commercial enterprise and like more like technical like technical native enterprise is somewhat like you know higher like in the US so that would be the main difference but that said frankly stepping back like you know we've had like some amazing like partnerships with like a bunch of enterprises across the Europe you know been working with companies like DBDA like you know which is one of the largest companies in Spain a bank and a bunch of others.

3:47And so we're seeing like a bunch of good adoption. That's amazing. And I guess that's very related to your role specifically at OpenAI on the platform side. Can you touch a bit on, you know, what it is that you look after and kind of your day-to-day job? Yeah, so I look after our business products, essentially, our B2B products. So OpenAI, of course, is well-known for ChatterPT, which is our consumer product, our consumer chatbot. but we have a bunch of business products that target essentially startups and enterprises. We've had an API, which was raw access to the models, which has been around since 2020 or 2021, 2020, I think.

4:32And we're building on top of that API more and more agentic product, agentic solutions for enterprises, for governments as well to adopt AI. and so yeah I look after the product teams building those products. And how do you decide you know what to build because your range of customers in this B2B enterprise space whether it's governments or big companies is massive you know like everybody's talking about OpenAI everybody's using it how do you know and decide what to build? It's a really good question I usually look at like two or three criteria. Criteria number one does it further the mission of open AI, as in, does it actually accelerate the distribution of economically valuable tasks?

5:20Like, are we automating, like, economically valuable tasks? Are we increasing the GDP, essentially, of the whole economy? Would be number one. The second, like, data point I usually look at is how much do those businesses, like, rely on, like, data, in particular computer data? some businesses are really much in the physical world manufacturing and it's somewhat harder for LLMs to be leveraged but some businesses you think about financial services life sciences some retail, telecom operate a lot on their computers bits and therefore it's more tractable to leverage LLMs and then finally we're being problematic as well when we come across like amazing design partners like i-credit companies that really want to transform like the whole industry we sometimes take you know paths so you know i was working back in the us i've been working a bunch with amgen who is a pharmaceutical company amgen is a pretty big company in the life sciences industry they design drugs especially in the in for oncology like cancer and for inflammatory diseases as well and they've been one of the most like you know the fastest like adopters of AI I've seen both across like their you know R &D like scientific clinical team and also like you know operation teams and so you know when you have that kind of you know opportunity and eagerness home enterprises like to to work with us you know we take bets.

7:00Now that's really interesting it's a pharmaceutical company is not the first company that you'd have in mind as being a really fast moving, quick to adopt AI company. So that's really fascinating. I'd love to know, you know, what makes, or if you were a large enterprise looking to kind of accelerate with OpenAI and AI more generally, what should they be doing internally to kind of get their teams, their users, and also their systems ready for AI? The way I think about it is you want to have like top-down and bottom-sept option. Bottom-sept option is fairly simple in the marketplace, in the AngelPrize.

7:36Basically, you buy like ChedPT, world-to-world for all your teams. Like, you know, you let them tinker instead of you with ChedPT. Some of them will just be using it like to write emails, like, you know, search like information. Others will be way more creative and be like, oh, that process is, you know, that process which is quite boring and, you know, I have to do it like over and over again. Like, you know, maybe there is a way, like, you know, like to automate it. So that's more bottoms up. And then you want to do some top-term adoption. And by top-term I mean like, you know, what are the most important problems that the CEO, they care about, you know.

8:09Give you an example, like we're talking about like Amgen. Amgen is like a company that designs drugs. Of course, the top problem is like how do I design like better, more drugs, right? And so here essentially you identify like a few like key like processes within, you know, that big use case, which could be automated or you're augmented with AI. and then you essentially work like super closely with like open air research team, engineering team, like forward deep engineering team to make that happen. And usually like we see like, you know, essentially that bottom's up and like top down, like, you know, approach, like meeting somewhere in the middle.

8:43And that's the holy grail. But yeah, stepping back, that's how I've seen essentially like, you know, those projects like work well. Maybe a question which is also interesting is like what projects do not work well adoption wise? when the leadership does not really believe in it, I've seen that, like, you know, like, executives who are, like, not really showing the example and not using it, like, not, you know, like, offering, like, bad tools essentially internally. Like, for sure, you know, that slows down adoption. But, yeah, overall, frankly, I've been pretty surprised in a good way by, like, the rate of adoption of other technologies, like, in the workplace.

9:23No, that's great. And when you talk about, you know, senior leaders or enterprise companies or clients, who are your stakeholders in those businesses? Who is it generally who's got to be on board and be pushing this agenda forward? Is it CEOs? Is it CTOs? Is it sometimes like a head of innovation? What's that kind of key stakeholder that you're targeting in a business to try and help them to kind of, yeah, adopt ChatGPT, OpenAI, your platform, your APIs? So we have three stakeholders usually. AI is very top of mind for like every single board like in the US, in Europe. And so usually the CEO is on the lookout, as I said, to leverage AI to solve some of their most important problems.

10:07That could be like, you know, increase the pace of innovation, like in software, in like, you know, R &D. It could be like making their operations like much more efficient. It could be reinventing their products, like you know for the product to be more customized, their customer experience. So that's usually like where the CEO comes in and then you have like two like different stakeholders which are really interesting. You usually have like the CTO or like let's say like the technical team, the engineers and then you have like the operations slash subject matter teams, subject matter experts.

10:42So give you an example. Let's say you want to use AI to automate like customer support you know I've been working a bunch with T-Mobile in the US, which is a major telecom company, to automate some of their voice, customer support. You have a really interesting setup here, which is the technical knowledge, of course, is with the CTO, right? They have the developers, the relationship with the consultancy, et cetera. But of course, if you ask any average developer at an enterprise, like, hey, how does customer support work? They don't really know. That's not what they do on daily basis, they're like code.

11:19Whereas if you ask a customer support manager or user officer, how does it work? They have all the procedure, processes, like, you know, in their brain. And so I found those projects to work really well, essentially, when we can pair together those two skills. Like, you know, when the engineers are capable, essentially, to really deeply understand what's going on. And second, when the operations people are able to provide feedback on samples, logs, every LLM output, essentially, in order to tune progressively the accuracy of that agent. So yeah, that's essentially the stakeholders. No, that's super interesting.

11:59And you also, I know from my place in the ecosystem, you see that the startups who are leveraging AI to kind of do, like sell products that, you know, automate customer service or whatever, they often have that amazing operational experience themselves. I think Fixer AI is a great company where they like set up an EA agency and now they're building an AI, EA essentially. And, you know, you need that context, that operational process knowledge to be able to like really well automate it and leverage AI to the maximum capacity. Completely. Something that first time founders, especially building enterprise software, usually do not have a good appreciation for is in the enterprise, the vast majority of the knowledge is institutional.

12:45It's in people's brains. There is not a perfect Google Doc that explains what you support at X enterprise. you usually go ask essentially like the five like most tenured like highest performers and then you sort of average out and you get essentially the operating procedures and that's that case you know across every single like you know functions and so unless you build like tools unless you are one of those people and if you're not one of those people like you essentially have to build tools that make it super easy to sort of emulate like those people like extract like the knowledge you know um from them uh yeah and i was gonna say that's one of the really it's an interesting challenge in in every business but especially now we're trying to automate processes it is taking that knowledge from people's brains and putting it somewhere where it's easy to understand so that both you know the tech team but potentially agents can automate it and run that process exactly exactly and you know you mentioned fixer uh fixer works with us open ai I think they've done that particularly well.

13:47Like I think it's a good example of, you know, you could take a very naive approach and say, the solution is technical. And that's not the case. I mean, of course, the technique like you know is important, but you have to really understand like the depth of that job, like the depth of that use case. And I think the Fixer team did it really well. Yeah, and it's like the user experience. You've got to be so deep in the experience of the users who are going to be using that tool. But I also wanted to touch on like the UK Data Residency and the new government partnership. Can you touch on that a little bit?

14:16Sure.

14:21AI across the globe is becoming more and more of a, I would say, like, you know, survey topic for governments. And by that, what I mean is that, you know, governments both want to build, like, local compute capacity, so local data centers. and number two is governments have some very sensitive operations like defense national security medical research and for that kind of data they do not want that to leave essentially their borders and so we've been working pretty closely with the UK government for a while and we announced Stargate UK a few weeks ago. So essentially what we deal is two of those components.

15:13We worked with a couple of partners to build data centers capacity in the UK. I think we're starting with something like 8 ,000 GPUs and we'll ramp that up over 2026 and beyond. And second, we've been working essentially on sort of a data residency, which basically means that our businesses who care a lot about their data like not leaving the UK borders, like we can just process everything in-house, I mean, in the country. So we've done that in a bunch of other countries in Europe, like the US and like Asia as well. So yeah, it's a fairly like regular, essentially, I think that we do. And it's essentially a way to like secure data sovereignty within those individual countries and the security of that data itself, right?

16:02Exactly, yeah. Amazing. And what's it been like, I guess, working with the UK government to get this up and running? They're super forward-thinking. Frankly, that's been the case for most of the governments we've been working with. Of course, we've been working a bunch with the US government. We've been working with the Singapore government, the UK government, Norway, Greece, UAE. I think all the leaders of those governments have truly, I would say, appreciated what is at stake right now and therefore are looking for new solutions to leverage best that new AI technology. And I think the UK has been one of them.

16:48We've been moving pretty fast and they've been really good partners. Amazing. That's great to hear. It's always hard when I cover a lot of European tech and UK tech. it's very hard to see the reception from governments, which I think is really interesting. But it's great to hear that, I guess, from your side, it's going well. Can you touch on a bit about how you, like how the system is set up to ensure safety, security and sovereignty for this data? Yeah. Oh man, it's going to get technical. There are like multiple ways. If you want to really protect your data, there are multiple layers or options.

17:32One of the reasons why we launched open source models, like open rates model, such as DPT OSS, was to essentially cater for that use case, which is for some specific industries, you just want to process everything on premise. like if you're running I don't know like a security command center for you know some like fire operation or something like you essentially have servers like you know in your premises like in your office and you just want to run the models on that so that's like the most easy right option the benefit is like you know complete like security privacy like control evaluate the The cost can be pretty complex to run to sample large language models at scale.

18:21You basically have to understand how to operate a fleet of GPUs. And that is a tricky kind of expertise. That's one option. I would say there are some industries that go down that track. But usually most industries prefer to process in the cloud for good reasons. like it's just much more convenient you can scale like much more quickly whenever there is a new model like you can literally like you know get access to the model like you know in a few lines of code um and so for those like you know cloud like um models access we basically have a couple of options like either um startups usually like you know just like process like wherever frankly like the data centers are based and you know some enterprises like prefer to have essentially the data uh being processed like you know in their borders and here that means a couple of things that could mean either like the data is being stored in their borders the data at rest like you know is literally like in storage like you know on the uk borders and then the second thing is like um the actual processing of the the llm like the the actual weights of the model and like the gpus like being in the border as well um so uh multiple options.

19:39Yeah, we covered all of those. Amazing. No, it looks great. I think it's really exciting for us here in the UK. I think there's always skepticism about new technologies and AI and security and safety. So it's good to see, I don't know, it being taken seriously and our government being kind of at the forefront of it, which is amazing. I also wanted to talk about Atlas. This is a big new release this week. Can you just give a quick intro into kind of like what it is? and I guess like how it's been since it came out. It's been, what, a day or so? A day? Not even. It's been not even a day. Not even a day, yeah.

20:16It's been under 24 hours. I don't know. I'm losing track. It's been 24 hours. Atlas is a browser. It's basically the ChatGPT browser. So what we did here is that we invented the browser that basically had like ChatGPT, not just like latched on, but like deep integrated into everything a browser can do. And so essentially I mean people should give it a try. It's a pretty wonderful application. You can instantly pop up a TEDxPT window on the side of any site, any page to ask questions about the content of the page. You can trigger an agent to go to perform stuff on your behalf within the DOM, within the website, including logged in services, if you want, I don't know, to buy a mouse for your laptop and you want to scroll to Amazon yourself.

21:18Atlas does it. We've also talked quite a bit about the design of the nav bar, the navigation bar as well. You have a tight integration with JGBT, which is pretty cool. It's the beginning, you know, but essentially the intention for us was hey people spend a lot of time in their browser we saw them doing like copy paste basically between like you know whatever they were doing like in the browser and chedp all the time and so we are like you know we should just meet people where they are essentially uh and so yeah i've been using it like you know uh for work like you know for several months at that point uh it's hard for me to envision like going back to a at that point no amazing it looks great and it's not just yeah i mean it's got like memory built in agentic functions as you mentioned i think i read that it's also going to be rolled out in business plans i think as well you're going to be offering it to business clients as well yeah totally um we are building all of that for knowledge workers frankly for people who spend their day consuming information reading writing texts writing emails writing spreadsheets writing code um and so the intention is very much like to make it work like you know for like businesses enterprises um so yeah we have a long list of like features that we want to build one of them that we need to make work you know for the business setup is a multi-profile you know personal and work profile yeah uh but you know we have like a pretty cracked like team working on it and so we'll expect to see a lot of updates in coming weeks amazing in the coming weeks i was going to ask you know do you have a timeline is this period of time just like get it out see how people are using it try and test and then iterate or have you got a roadmap in mind of okay within a few weeks we want to launch this next feature then this next feature then this next feature there are some features that we know we need to launch yeah uh but that said just like everything we do the reason why we ship like so early and often like such products is to iterate with customers.

23:22Those are new technologies for everyone. And so until you get them in the hands of users, it's hard to envision exactly what's going to stick or not. The other example I have in mind is Codex, which is the coding agent of OpenAI. We shipped it early. The team is iterating multiple times a week with updates based on customer feedback. And so that's very much like the OpenAI way to develop products. Yeah, I think that's great. I remember the first time I applied for a product job, the head of product who was interviewing me said that if you are not slightly embarrassed of the product that you're shipping you're shipping it too late and he's just like you've just got to get it out in the world it's not perfect it's not complete but you just got to get feedback and iterate quickly it's a it's a lesson of humility like you know as a product person yeah you know see your work of love like you know launching out there and you can see like all the 90 like you know different glitches but still you have to ship something you have to conform to users.

24:21I mean, that's the only way, thank you, I know, to get product market bait. Yeah, it's the only way of doing it. Well, Olivia, I think we're out of time, but thank you so much for joining me. I've loved chatting. It must be such an exciting time to be building at OpenAI, and I love seeing the constant kind of like train of new products that are coming out. So thank you so much. I can tell you one thing, we'll keep shipping, Seb, so expect more from us in the coming month. Can't wait to see you. Thank you.