The Exchanges

Q And you, you pivoted a couple of times, right?

A Oh, back in the days, yeah, Gazian times, way too many times. Uh, we started with, uh, we started with, uh, digitalized coupons for, uh, local stores. Uh, it was kind of the Groupon days. So, uh, we were like, oh, the, the, the kind of, uh, loyalty cards should be digitalized on the phone. Uh, pretty tough to sell anything to, to, to local stores, um, above 10 dollars or something a month. So there was stuff, but we luckily stayed in that business only for four months. Um, then we, we did an app that was aggregating all the publicly shared photos in real time. So it was mostly Instagram Foursquare. Uh, Twitter and a bunch of other, uh, photo apps back in those days. Um, uh, it was a B to C app where you could teleport somewhere, seeing the live feed of photos, uh, that didn't work that well. Uh, but eventually we realized that there was a B to B interest. So we started building a B to B product around those kind of feeds, real-time feeds of photos, and eventually realized that Instagram was taking the line spots. So we pivoted on Instagram only, and then realized that the brands, which were the only ones that had money to Pay us, uh, events and medias didn't at the time. Uh, we're interested in not the, the content, but more the analytics. And so that's what we ended up doing, uh, an Instagram analytics platform solely focused on Instagram. And that's where we went from a long …

AI assessment note: “Oh, back in the days, yeah, Gazian times, way too many times.”

Q So does that mean that for now you find yourself, uh, which I guess it's pretty natural at this, uh, moment of the, of the cycle, but do you find yourself doing a lot of like effectively services where like you sit down and you help people define the problem and, uh, what could be solved?

A We obviously do some of that as an early startup should obviously do. Uh, but we, we really think that, uh, it's something that could, that can be scaled up through product, basically. What you really want is you want to, you want to have a product that where there is no question that the product is better than using TARGPD because you have, uh, GPT-IV, you have cloud, you have productivity features, you can collaborate and stuff and blah, blah, blah. And those early adopters can become the builders for the later adopters, meaning that those people that are early adopters, maybe engineers, they understand technology should be given a platform where they can easily create those tools for the other ones in a way that's at the end, we don't necessarily need to be involved inside the road of that product. The product by itself starts with the early adopters, start with the engineers and the engineers might build tools, uh, for the ones that Require a little bit more guidance. And I think all of that can probably scale through product and not necessarily handholding and kind of custom work with our clients eventually. But obviously in the beginning, we are to learn and to, uh, to, to, to, to, to build that product. We, we need to do that, uh, ourselves with them.

AI assessment note: “We obviously do some of that as an early startup should obviously do.”

Q Okay. Very good. And so that, that company, uh, as I was perving for this, uh, I think, uh, I read somewhere or heard somewhere you, you, you got into, uh, almost like a bidding war between Pinterest and Stripe. Can you talk about what, what, what happened? Well, at least you were in conversations with both.

A I wouldn't call that a bidding war. No. So we, uh, basically we, uh, we were struggling to raise the series A at the time. So time has changed since then, obviously, because when you're 10 people doing, let's say, a 1.5 million an hour, I think it's a much, uh, I mean, it was really hard at the time to raise the series because the growth was not explosive. The B to B on Instagram was not yet explosive. The B to C part was explosive, but B to B part was kind of a linear growth. Um, and, um, so we, we, we, we knew what we had to do to, to, uh, pump the growth, but it would probably would have meant to do the same trick on another platform. So start from scratch with fully verticalized product on, let's say, Snapchat. And since we put people together in time into that, uh, it wasn't kind of a, we weren't born to do Social media analytics and five years in kind of grew a little bit, uh, to be perfectly transparent, a little bit tired of, of it, uh, lost the, uh, kind of, uh, uh, uh, energy to push further. And that's where we started looking for a very, for a nice exit for the team, for the investors, for the company. And we, uh, we chatted a lot with, uh, Pinterest and the moment we got an offer from Pinterest, uh, we were like, okay, is there, is it, is it the thing we want to do? Is there a company that we're more excited about? And there was one where the, the biggest kind of, …

AI assessment note: “I wouldn't call that a bidding war. No.”

Q Yann Lacan, who's, uh, you know, obviously one of the godfathers of, of deep learning and, and, and French, although he's been living in New York for a very long time. So he, he, he launched his team, which was mostly fair and, and then, you know, head of, head of AI at Facebook, uh, he launches his team mostly in Paris, right? Is that, is that, is that right?

A I, to be honest, I don't know the details of all meta. I, I believe there was a lab in India. Of the obviously in New York, but there was like, yep. There's a very big lab. And what I know is that there, there, there, there is a very big lab in Paris and there has been a very big lab in Paris for a very long time around AI and a lot of tenants, uh, grew into that lab in the same way that some talent grew into the deep mind Parisian lab and the deep mind London lab. And so as a consequence, the reason why Paris is so attracting today, I believe that's my hypothesis is that it's a, it's a pretty deep pool of AI research talents. But one that is not too coveted yet, meaning that if you want to build a strong AI research team today, it'll be 10 times easier to do it in Paris than it is to do it in SF, uh, with OpenAI, uh, as a lab, as a competitive lab to, in the same hiring markets. Um. Still hard though, but. Easier. But much, much, much, much easier. And so I think that's why, uh, so obviously Mistral is started, started from by French people and they have their connections in the kind of French and London ecosystem. So that's kind of natural to them. The much more interesting case is the poolside case, uh, because they elected their, their U S funders, or I don't know if they're U S but non-French funders, uh, multi-time entrepreneurs in the U S and they decided they elected th…

AI assessment note: “what I know is that there, there, there, there is a very big lab in Paris”

Q you know, maybe at a high level and taking a step back, what's your current take on OpenAI as a, as an organization? And, um, you know, they're going in lots of different directions and doing lots of different things. Are you, uh, are you super bullish on it as a, somebody who spent years there? Like, do you have any reservation? I'm just, uh, just curious if you take.

A Yeah, so to answer the question, to be very clear, I think I had my insider hats for the previous questions, because we're talking about a period when I was inside OpenAI. I know it's been a year that I'm outside of OpenAI, so I'll take my, my outsider hats to answer the question with my knowledge as an insider, obviously, embedded in my brain. Um, no, I think, uh, I'm, I'm, I'm still extremely bullish on OpenAI because I think this is a unique organization that was able to raise a very large amount of capital to poorer research, research organization and a research project that is, uh, uh, very hard to reproduce anywhere else. Um, obviously you have Google and DeepMind, Meta to some extent, Anthropik, but still remains kind of a uniquely positioned there. And they, The potential is humongous, so I understand why some people may decide that the expected value is humongous because the potential is humongous. I think the expected value is humongous because it's, it's not if, but it's more when, but you obviously if the when is very long in the future, then the value starts decreasing because you apply it kind of a rate to it, but It's hard to believe that it's going to take a hundred years to achieve something spectacularly, uh, transformative for, for the society. And today, It remains, it remains clear that the most efficient way we know how to push the technology is by scaling…

AI assessment note: “I'm still extremely bullish on OpenAI because I think this is a unique organization”

Q under, under the hood, you have, uh, multiple models. You have, uh, GPT, uh, turbo, uh, cloud instance, uh, and you can have custom models. How does that work? Are you, are you, do, do people pick which model they want to use for, I don't know, cost reasons or whatever? Uh, or do you, are you building like a routing layer that, uh, pushes certain queries into certain models?

A No, so we're trying to stay pretty transparent there. So if you are a free user, you only have access to a small model. So the ones you mentioned, if you're paying customer, you have access to GP four and Claude, and we foot the bill, uh, really want to be in a business where you're not thinking about tokens. And we, and as we build a product, we are fine using way more tokens if it increased the quality of the answers. Uh, because the cost will go down. So that's fine. Even if you're losing money today, you'll be winning some, I mean, earning some money tomorrow. Uh, so we, we're pretty, we, we, we exposed the models that are available and we let our users pick the model, uh, they want when they build their, their assistants. Uh, so we try to be quite model agnostic. And today we indeed have GD for Claude. We probably rolling out just for, for fun because we're in Europe, uh, Mistral, uh, very soon. Which will be a much smaller model, so it's gonna be, I mean, it's, it's, the risk is that people start comparing GPT-FOR with Mistral-Seven-B, which is an affair comparison, because it's a very good model, Mistral-Seven-B, but obviously not at the scale of GPT-FOR. But we, we, we, we really want to focus on the product layer, and so, um, we quite model agnostic, and we actually quite transparently expose which models are available, and you can pick the one you want to use, uh, you…

AI assessment note: “we exposed the models that are available and we let our users pick the model”