In this discussion at Startup School Paris, Dust co-founder Stanislas Polu reflects on his path from Stripe to OpenAI and why he and his co-founder decided to build their company in France.
00:00 — Intro 01:09 — From Stripe to OpenAI to Founder 02:49 — Why Stan Left OpenAI 03:46 — What Dust Actually Does 04:07 — What Work Looks Like in 3 Years 05:10 — Where the Thesis Was Right (and Wrong) 07:05 — The Horizontal vs. Vertical Debate 09:15 — Model Agnostic as a Moat 10:49 — How Labs Starve the Funding Market 12:19 — Raising Small on Purpose 14:39 — The Mistral Lesson 15:11 — Why Build in France? 17:08 — When Labs Enter Your Vertical 18:27 — The Margin Problem
AI Generated·May contain errors
Video Summary
AI Generated
Stan, a former engineer at Stripe and OpenAI, discusses the founding of Dust, an AI platform designed to redefine professional workflows. He emphasizes the importance of remaining model-agnostic, focusing on \"multiplayer AI\" for collaboration, and maintaining a disciplined approach to fundraising and growth to ensure long-term sustainability in a rapidly evolving landscape.
Model agnostic, which I think is a very interesting aspect of that and it's something that the labs
would not be able to do. If you think about buying your product and your tokens at the same place,
it's like if you were building a plant with machines in there and you would buy the machines
from the energy provider and the plug would only work with one energy provider and if your
energy provider was a Russian gas would be in a pretty bad place right now and you wouldn't be
able to connect to nuclear French, French nuclear. Makes no sense.
So you can give a round of applause to Stan. Thank you for having me. I feel like we're
super lucky to have him. He's extremely wise. He's gonna talk a little bit about building in
the shadow of Frontier Labs and then at the end I want to ask him about company building because
Dust has been a company that has been super thoughtful in how they've raised, hired, operated
and so I want to be able to share some of this knowledge with you. Maybe let's get started. Can
you tell us who you are? I'm Stan. I was lucky to join Scribe pretty early. Spent five years there.
That was an acquirer when we joined Scribe and then at the end of those five years,
got interested in into AI. I'm not a trained researcher but I was also lucky again to join
OpenAI at a time where you could do research as an engineer and it was all possible and
spent three years there working on the large language models and maths. And you decided to
leave OpenAI? Yes. Why? Crazy, huh? Right. Yeah. I think there's a fun fact on this one. I think if
I stayed at OpenAI, I think it's not true anymore, post-Series B, but the stock options that I gave
up were more than the whole company as of six months. That seems like a bad idea. Yeah. Still,
do you regret it? No, not at all. So why did you leave? So really, research is a very specific
exercise. Again, as I said, I'm not a trained researcher and so I spent those three years doing
the PhD that I never did, really enjoyed it. But at the end, I really wanted to go back to
building a product. When you do research, it's really, you're scratching a surface for weeks or
months and then you find something interesting, it gets you super high, super, super high,
but it lasts for five seconds, literally five seconds. And after five seconds, you're like,
oh yeah, that was a ghost. And you keep scratching again. So I think it's something
interesting and I really enjoyed it, but I wanted to go back to building something
more around the product. And so you decided to build something on top of the labs. So you started
Dust. Can you tell us more about what Dust is? Yeah, sure. Obviously, we live in a world where
the definition of the company is a moving target because we're operating on a technological
substrate that moves super fast. For the past 20 years, you've known that, I've known that,
the technological substrate was JavaScript and Postgres. It was kind of a rock solid. You could
see what you would be building in years and that's not necessarily the case. But I think the Dust,
really the initial motivation was very simple. It's like those models are ready to really change
the way we're going to work. And we want to focus on applying LLMs to the workplace, which in 2022
or early 2023 when we were raising was almost considered a niche, which was fun.
But yeah, that's the main idea. Obviously, the what itself evolves a lot. So you spend a lot
of time thinking about work and how work is going to evolve. So if you're right in how you're
thinking about it today, what do you think is going to happen like two to three years from now?
Wow, two to three years. It's like impossible to predict anything at this.
Just try to tell us what work looks like two to three years from now. Your best guess.
Yeah, I'm a bit of a Sam Altman. I mean, he said that at some point, I think I relate to that.
I mean, if you think about about our great grandparents, work was mostly physical. Great,
great grandparents, whatever is the number of grades. And if they were looking at us today,
they would be like, you're not working here. You're just sitting in a chair, looking at the
screen all day, chilling around. That's not work. And so I think like two to three years that I'm
pretty convinced, you know, and I'm pretty optimistic that we'll still be working, but that
work will feel like not work. We'll be looking at us in two, three years and today's us would look
at us in two or three years and we'll be like, that's not really work what you're doing here.
So the interesting thing, though, is, as you said, in 2023, when you started,
it was the very early days of agents. But you still had a thesis on how they would impact
work when you reflect back, where were you right and where were you wrong?
So I think we were deeply, deeply convinced that work would be incredibly disrupted by
the technology. And I think that's something where we were deeply right. And I think we're
starting to be right only now at the three years time scale. It's only during the past couple
months that work has started to change. Like as a dev, the way we work has completely changed
since November 2025. So it's been just a couple months, but the work has like radically changed.
And I think we're starting to see that in every other functions. And so that is that this
technology would be extremely disruptive and structuring for how we work. We were pretty
right there. Even if for three years, work was kind of the same. There were a lot of
verticalized AI product on some particular pieces of work. There was two years ago, it was all about
clean, which was mostly about asking questions and getting answers. But the shape of work was
not necessarily changing. And so that's the part where we were right in a sense is that
this is going to reinvent the way we work. The part where I was wrong,
I think I was more of a believer of an earlier plateau in a sense. Even if that technology had
plateaued like a couple months ago, we would still have like decades of deployments and innovation
to do for it to be adopted in every pieces of places where we do work. But that plateau is not
happening at all. And it's still, if you look at the way we work is being redefined right now,
it's clearly not happening. So I think that's maybe that part where I was not as optimistic
as I should have been. Yeah. I remember you telling me about this plateau and I'm still
waiting for it. Yeah, it's still not coming. Yeah. So I want to make sure we talk about
the elephant in the room, which is OpenAI and Anthropic are giants. They move fast,
they innovate and you're working right alongside, like right next to them. How do you feel about
this? Challenged, I guess. No, it's obviously challenging. We went for an horizontal platform
very early on. Again, for the past couple of years, everybody was saying, verticalize AI,
verticalize AI. And there were great arguments for that, is that when you have a verticalized AI
product, your go-to-market is much simpler because you know who to target, you go talk to them,
you watch, rinse, repeat. It's much more obvious. When you have an horizontal platform,
it's super well documented. We're not the first to do that. Notion, Airtable, all of that. The
go-to-market is much more finicky because these platforms are obviously extremely powerful, but
people are like, ah, yes, but what am I going to use it for? I don't understand, blah, blah, blah.
Working in the, as you said, in the shadow, shadow is a bit harsh for us, but working alongside,
obviously with tiny, tiny spec compared to Anthropic and OpenAI, that's fine.
Working is both beneficial and challenging. It's beneficial because they educate the market
very efficiently. I think as we're seeing today, it's all about horizontal platform. It feels like
every product is converging towards the same thing, which is some sort of new productivity
suite like Microsoft has been doing or Google Drive has been doing, but for human-agent
interactions. The product, they used to be pretty deep product pockets, those verticalized AI,
but it seems like everything is converging towards much less depth in terms of product and much more
coverage in terms of where it integrates on everything. It's a bit like an industry-body
problem novel or TV show. You see that thing that goes around the earth. You almost don't see it,
but it's super powerful because it covers it entirely. I think we're competing for that.
There's a lot of convergence in the product space, I think, in that aspect.
There'll be some different sensibilities in the way you build that bubble of product,
which will create differentiation. We are going a lot towards collaboration and multiplayer AI.
We could spend minutes and minutes about explaining why it makes sense. We also model agnostic,
which I think is a very interesting aspect of that, and it's something that the labs will not
be able to do. If you think about buying your product and your tokens at the same place,
it's like if you were building a plant with machines in there and you would buy the machines
from the energy provider and the plug would only work with one energy provider. If your energy
provider was Russian gas, it would be in a pretty bad place right now and you wouldn't be able to
connect to French nuclear. It makes no sense. I think it makes a lot of sense for those products,
for the pro player, to be agnostic in terms of where you bring your intelligence from.
You're model agnostic today and you'll remain this way?
Yes, definitely. It's funny because we model agnostic but at the same time,
at any point of stage, a lot of the usage goes towards one provider. The point is that for a
given period of time, that provider is the best for doing work. At another given point of time,
this other provider is the best for doing work and that still changes very dynamically. We don't
see any moment where we would not be model agnostic. Yeah, that makes sense. When I think
about these giants and what they do to the ecosystem of startups, they impact them in many
ways. One of them is the funding markets. There is a risk that I think is an actual thing happening,
which is they pull out the oxygen out of the funding market because you're a VC and you'd
rather spend $1 billion in one lab than spend 10 times $100 million in 10 different companies.
You've raised in this environment, so you have experience with this. How do you think these
giants impact the funding market? We were lucky to have a very easy seed,
very, very easy seed. We were lucky to have a very, very easy Series A and the Series B was
the first time we raised for real and it was complicated and hard as it should be.
I think part of that is because we're building from France in a moment where it's less,
from the American perspective, slightly less fashionable than in 2023.
The other part is that I think a lot of the energy in the market is being absorbed by the labs.
Indeed, as you said, if I was a gross investor, it would make total sense for me to put $1 billion
in Entropiq, get $2 billion in six months and be done with that. Hopefully, those companies will
become public, which will put them in a no-go zone for VCs, which means that hopefully they'll
bring liquidity to the market. Hopefully, they'll get better. I want to go back to being in the
shadow of Frontier Labs before because you talked about fundraising. I want to talk about this
because one of the things that, when I talk about dust, one of the things I say is, I think
Stan and Gabrielle, who's the founder, are some of the most thoughtful founders I've seen, specifically
on company building and specifically on raising money. You could have, coming from OpenAI Stripe,
being a second-time founder, you could have raised at a crazy valuation and yet you decided to raise
at reasonable valuations for your seed and then again for your Series A. How did you make this
decision? I think it felt very natural to us. We've seen some friends get a little bit burned
in the 2020-21 years, raising a little bit too much. I think there is a true coffin corner for
companies. If you raise too much and do not realize the potential of the money you raised,
you end up in a coffin corner where the only outcome is a downturn. I've never done that,
but I think a downturn is really, really bad in many aspects. We knew from the get-go that we
wanted to focus on the product layer. We had a mantra that is still almost true, no GPU before
PMF. We raised our seed at a moment where it was fashionable to raise money for training. Everybody
was running $100 million here, $200 million there, and we raised $5 million, which seems silly.
But at the same time, we knew that building at the product layer, we would not have been able to
use that money and it would have made no sense. We would much rather raise less, raise at a smaller
valuation, but be very aggressive into building into that and creating the momentum and raising
another round when we're ready, etc. That's the positive view of that. The negative view of that
is that we're being a little bit too French. I think you want to strike the right trade-off.
I'm always impressed by Mistral in that perspective. I've met Arthur Mensch because I
knew Guillaume Lamb very well when he decided to go for building a company. He's extremely smart,
but he knew nothing about building a company, like nothing. He was just a researcher, no clue,
and the way he grew into that kind of role of CEO of European Lab, with all its imperfections,
the fact that they managed to build a company that is valued at $12 billion, is extremely
recommendable, is extremely impressive. In that, there's a bit of, let's try to kick the ball
very aggressively. I think it's a fine line to consider. Both aspects are important,
and you need to find your own. It's a grey zone. Yeah, that makes sense. You talked about your
series being maybe more challenging because you're building in France, and it's true. I knew
you from your San Francisco days, so you knew what San Francisco was like. You built there,
you hired there, and yet you decided to come back to France to build the company.
How did you make this decision, and what are your thoughts today on having made this decision?
Everything would have been easier from a company building standpoint, going in the US from the
get-go, I think. But it doesn't mean it being easier, it doesn't mean that it's the right
thing to do, and I think we were really... I mean, I was, and Gabriel as well. We really wanted to
build something back in France because it was important to us. There's an aspect like national
preference or sovereignty to it. That's an additional constraint, so that's one that you
have to be very aware of. But I remain convinced that when things are going super well, that's no
problem. I mean, if you look at the Lovable of the world, the Spotify of the world, everybody says
SF, SF, SF, the founders need to be in SF. And the same VCs, they don't give a ---- about the
localization of the Lovable founders. They could be in whatever country. When the things are working
well, the things are working well, nobody cares. So it's an additional friction that you take in
for the greater good, maybe, but it's a trade-off. That makes sense. I want to talk about the labs.
There is one specific question I want to ask that may be useful for the audience, which is,
we've seen the labs going after new verticals. So example of this is Anthropic going after
Legora in the law space. So if you're a founder and that happens to you, how should you think about it?
So being verticalized, great advantage in terms of go-to-market. Being verticalized made a ton
of sense because the models were not that good and so you had to build a bit of scaffolding
around that to make them useful. That second value just disappears as the models are getting better.
So I think that doesn't mean that it's not a good idea to be verticalized, but I think
we're almost in a world where intelligence is commoditized in the way it's like
the old good SaaS and internet days. It feels to me like the value that a verticalized product
will have is if they manage to create a different type of mode, which is probably the networked
effect you can find in your vertical. I think that will be the mode because that's the thing
that is defensible even in place of perfect intelligence coming in. So if you're building
a verticalized product in a given industry, I do think that we're going back to the advices that
we used to receive in 2010 almost, which is networked effects. Find a networked effect.
Find a thing that is defensible because you'll have value coming from having multiple users
interacting with the same platform. If you don't have that, if it's just you're delivering
intelligence to one and intelligence to the other and there is no networked effect, I think in that
case the defensibility is changing. How do you think about margins? Margins are tough in those days.
Tough, very tough. So and maybe the question is, so you know when you're a lab you're capturing
the margin at the token level and then you build a product and you don't usually add any margin on
top of this and so then you're pressured when you've already been taken some of this margin. So
for you, how do you think about charging and what feels like the right way to think about it?
Yeah, so we used to have a seed-based pricing which made sense for three years which has become
because we wanted to favor usage and value creation. So it was a blanket price you pay to
use as much as you want. Obviously over the past six months the usage has exploded, the agent loops
are getting longer, models are just using more tokens all over the place and so our margins
have compressed and we are in the middle of transitioning to a credit-based pricing and
I think that today if you are building at the product layer there is no way around credit-based
pricing because people want to use more AI and they'll just max use as much as you give them
and if you have a flat price that just doesn't fly. So that is for kind of how to think about
pricing in terms of a current product in AI I think. It's the only way to maintain your margins
because if you have a flat price the margins get compressed and they'll keep getting compressing
at least under the current conditions of models getting better, people are using more and more.
And then the question about competition with the labs. I mean there are precedents where companies
manage to sell a product on top of an infrastructure that captures the margins and resell
the same product at zero price. Many examples, Box, Dropbox, One Passport. I think there is
obviously value in creating a very well-thought-out, very well-crafted, very well-differentiated
product and that will enable to capture a fair amount of margin still. And finally the third
piece is that hopefully we have the open sources catching up to some extent and this will create
some pressure on the labs on the margins. If you think about it, when you look at the numbers
and you can look at the S1 from SpaceX as well, the labs and margins are humongous.
They're humongous. So I'll try to hold it in 10 seconds. But if you think about the cost of a
model, you can look at the latency. The latency is a measure of the cost of the model because
for a given model infrastructure, the latency gives you the size of the model, the energy
it consumes, etc. And if you think about the latency of Cloud Code versus GLM 5.2, it's roughly the same.
It's the maximal latency acceptable to run in Cloud Code. And so you can compare the margin, you can
compare the price of GLM 5.2 on base 10 or whatever fireworks to the price you pay for an equivalent
frontier model with the same latency. And the difference is almost 9x. So it means that there
is a very good chance that those labs are creating a margin of something like 70 or 80 percent
serving us those frontier models. So hopefully with open source coming up, it will recalibrate
a little bit the market and those margins will go down and you'll have a better chance to compete
even if they still capture a lot of margins. That makes sense. We're out of time. I'll still
ask you one last question, because as I said, you're very wise. So what is there any last word
of wisdom that you want to share with this crowd, this crowd being entrepreneurs or people that want
to be entrepreneurs or young in their career? Yeah, I think the only thing that really matters
at the end is find the... I mean, it's going to sound so typical, but I think it's really true.
Find the thing you want to fix, the vision you have, because building a company is shitty on a daily
basis, like shitty as hell. There are some very highs and a lot of lows. And so if you don't have
that thing that you can really wake up in the morning and say, but this is the reason why I'm
doing it, it's not sustainable. So that's the first thing. If you don't have that in your mind,
find it today. It's fine to reinvent it while you build. It's OK. But find that thing that will keep
you driven towards an objective despite all the glass eating that is involved in the job.
Yeah, I agree. Thank you so much, Sam. Please thank him as well. It was amazing to have you guys.