In this talk from Startup School Paris, AMI CEO Alexandre LeBrun talks about his journey building Vertos, selling Wit.ai to Facebook, and why the best advice for early founders is to pick an extremely narrow problem while holding an extremely large vision with the space.
00:00 — Intro 00:33 — Same YC batch, four companies later 01:40 — Zuckerberg's email in the spam folder 02:14 — The YC rule on acquisition talks 03:04 — Always 20 years too early 04:45 — Play to your strengths, not the market 05:43 — Why AMI is different 06:46 — What is a world model? 08:31 — Training like a baby, not a reader 09:22 — Robots with actual brains 10:27 — Why VLAs are a bad hack 11:44 — Where LLMs still win 12:39 — World models already in the wild 13:15 — Why big labs can't take the risk 16:12 — Is this still contrarian? 17:03 — How Jan recruited Alex 18:49 — Splitting research vs. execution 20:08 — Raising Europe's biggest seed 22:23 — Expectations are the real price 23:24 — Why not SF? 24:48 — Zuckerberg wanted 10,000, not 100 27:25 — Go narrow, aim massive
AI Generated·May contain errors
I told the team yesterday that the real cost of this $1.2 billion was not dilution.
The real cost is expectations, not only from our investors,
because I think our investors, they understood and they agree on the long-term vision of the company,
but the expectations you create in the outside world,
and if you raise $1 billion and you do nothing for two years,
or people don't see anything coming out of it for two years,
then it's very, very hard to survive.
So we actually met Alex, fun fact.
We actually did the same YC batch.
Back then, Alex was building WIT.ai, and we were building Algolia.
It was in winter 2014, quite a while ago, except Alex has done four companies.
One before, two after.
I've only done one, so I cannot compare.
So I'm super impressed by all your background and everything you've done.
What do you remember from WIT?
WIT.ai, was that the first AI company or something?
Yeah, it was the first .ai domain, actually.
It was very hard back then to buy .ai.
It's an island off Australia.
Did you know about it, or did you find it by chance?
No, I found this and I thought, oh, it would be cool to have this domain,
but nobody understood it was a domain back in early 2013.
That's how old we are.
I was surprised because you sold WIT shortly after to Facebook, right?
How did that go? Any memory of that time?
In 2015, I received an email from someone called Mark Zuckerberg.
After half a second, I deleted the email without opening it.
It's a scam, you know, it's a spam.
I took about half an hour, and then I thought, oh, but maybe it isn't.
I should check, and I opened my spam folder, and it wasn't.
So it started like that.
And so what happened after?
You met him, it was real, you met him, and he made an offer?
Is that it?
We actually followed the YC playbook.
You're not supposed to...
If you don't want to sell, don't speak with them.
Yeah, you're not supposed to speak to anyone.
As soon as you start a startup, you receive lots of offers to talk.
But usually, these are from the corporate development of large companies,
and they need to talk to everyone to understand the market,
and they need to know if you're desperate enough to sell for a low price.
And so it's a very bad idea to waste your time and your attention.
So YC says you never talk to anyone unless you are selling your company.
Unless you want to sell.
But that was not the corporate...
I mean, or maybe the corporate arm of Facebook was so smart,
they got Mark to send the email?
Yeah, actually, the playbook is if someone really has the power
to do something big, and you can talk.
So tell us what was...
In a way, it feels like all your companies were very early
in what they were doing.
Like Virtuos was already a conversational agent, right?
Before that name even existed.
So this one was really, really too early.
But WIT was also early.
WIT was not that early.
So Virtuos, my first company, almost out of school,
was a chatbot company.
I started in 2002, so you are not born.
20 years too early.
Yeah, 20 years too early.
But still, we made it, you know, we made something with it.
But then people ask you...
So I'm trying to sell chatbots for customer service in 2002.
And people ask me, but what is a chatbot?
And I say, it's like live chat, but it's automatic.
And they say, but what is live chat?
So it was really, really too early.
And we spent 10 years with this company.
We survived long enough to...
And then there was actually a market in 2010,
as a nascent market, but it was definitely too early.
You sold Gatros, you sold WIT,
then you spent a lot of time at Meta,
and then you created a third company, Nabla.
And then at last, Amilabs.
Was that earlier this year?
Yes, in January.
It feels like every single one of these companies
was kind of early.
Even Amilabs, right?
In a way, you are going after a big new approach to AI
that could feel very early today.
Is that something that's new?
You always want to really...
I think it depends what is your strength.
If your strength as co-founders is on go-to-market,
sales, marketing, then you shouldn't be too early.
You need a product to be ready.
You need the market to be ready to receive your go-to-market.
And so if your strength is to push a product to the market,
then being too early will kill you.
I think my strength, and I think usually it's a cliche,
but in France, probably,
we are most famous for engineering and tech
more than the genius of marketing and sales.
So in my case, I think what I like to do
is take very, very early tech
and think about how we can sell it through a product,
build this product, understand the early market.
So I think my skills are more on the early side
of the go-to-market,
so that's why I prefer to take early ideas.
But it really depends on what you want to focus on.
Let's continue with Amilabs.
It feels like it's the most ambitious project ever for you, right?
I mean, the first one were normal startups.
This one is not really normal.
What led you to actually take that leap?
So, I mean, it's a normal project.
It's just very, very expensive,
because if you want to build a financial model,
you need to buy thousands and thousands of GPUs,
and so you need billions of euros.
So this is what is not normal in this project.
So I've done several startups,
always on the applied AI space,
so I'm using machine learning research
on models that other people have done.
And after all this year,
we're having the opportunity to work upstream
on the model that, if we are successful,
millions of people will use.
It's very exciting.
So moving upstream is what I think is the most exciting.
How many here know what a world model is?
More than I would expect.
I had to research it myself,
just to make sure I understood what you were doing.
Because it's kind of like a very different take at the market
than LLMs are kind of like the one thing, right?
In a way, it's contrarian.
It's been contrarian for a while.
Why wouldn't LLM do everything?
Can you tell us a little more about what is a world model
and what's your bet on the market?
Yeah, so first, I'm happy to see that many people
have heard about it.
How many were honest? I don't know.
So my co-founder, Yann LeCun, speaks often, so maybe...
You heard about him.
So basically, as you know,
a large language model learns to manipulate language.
And so LLM will be trained by reading all the texts,
all the books, all the Internet in the world.
And so LLM is learning what we humans have written
about the world.
So it's a kind of proxy.
LLM doesn't learn directly from the world, the real world.
We humans have wrote books and articles about the world,
and then LLM learns these texts.
LLM is like someone who has never gone outside,
never left the room they were born in,
but read all the books every day for centuries and centuries.
And so this person seems like an expert on many things,
but with no direct experience of the world,
will have limited common sense and limited ability, actually,
to find new solutions or to know what to do
when facing a situation for the first time.
Lacking the direct world experience.
So a world model is a model where we don't cheat.
We are not learning through text.
We are learning directly, like humans, like animals,
from the real world.
So meaning, video, audio, sensory data,
playing with objects.
So we are really starting from scratch, like a baby,
without any language.
So what's the training data?
What kind of training data are you using?
We are using a lot of video data,
because it's easiest to find.
But it works with audio.
It would work with only...
If you were born without sight or hearing,
you can still live.
So it's possible to live through touch,
and it's the same in what we do.
Can you collect touch data? How do you do that?
Through robotics.
Okay, that's interesting.
So once you get that model,
what are the use cases?
What can it do that an LLM cannot do?
So in the short term,
there are lots of applications, especially in robotics.
Current robots are very, very narrow, vertical robots.
They do one thing in one position,
and it takes a lot of time to train them
to do only one thing, always the same.
We are not ready at all to have these robots
be in an open environment,
whether it's in your house or in the street.
And you see these videos of robots
that start to dance and break everything around,
or I don't know if you've seen the one
where a robot is doing a karate demonstration
and does like that, and there is a child,
and the child is...
So these robots are very, very dumb.
The hardware is incredibly evolved,
but the brain hasn't changed.
And you cannot take these robots
in an open environment.
It wouldn't be safe, and they wouldn't be useful.
So this is a first application.
How does that differ from VLA, for example?
VLA is a very bad hack where you have a hammer,
you have an LLM, and you see everything as a nail.
You are trying to use your solution
for things that it was not designed for.
So VLA is not accurate.
It's very slow, but you need very fast,
very short latency to do these applications,
and it's very expensive.
You cannot pay $100 of compute
for 24 hours of your home robot to work.
So there are many, many problems with VLA.
And so, like, what would be a better solution?
Are they easy in terms of inference?
Are they costly? Can they be real-time?
So they are very costly to train, like LLMs.
Hence, you know, the GPUs...
More or the same?
It's comparable.
It's hard to compare because it's not the same structure.
But once they are trained,
they are supposed to be more lightweight
for several reasons, but you have less parameters.
So inference should be less expensive.
Okay.
In the world where world models are, like,
in production everywhere,
what is still good for LLMs?
Like, is that one replacing the other?
No, no, I'm not saying LLMs are not good.
LLMs are already very, very good,
as we all know,
for everything that is language-first,
you know, everything that is a sequence of symbols,
like discrete and low-dimension problems,
so anything with language, mathematics, programming.
These are all low-dimensional sequences of symbols.
LLMs are designed for that.
We will never beat, I think, the LLM at that.
But then you have lots of high-dimensional,
very noisy, very long-horizon problems
for which we think world models will be vastly superior.
So we're speaking a lot about the future.
Like, is there already world models in prod?
I got some flavor of it, at least.
There are some ideas behind world models in production.
For instance, some autonomous driving companies
are using some flavors.
Is Waymo using a world model?
Some flavor of, and Wave also in the UK.
Some of the ideas we are using, like JEPA,
have been around for at least five, ten years.
You know what happened in 2018?
The theoretical idea behind LLMs, like transformers,
the paper was there,
but Google did nothing with it.
Nobody did anything with it.
Nobody realized the power.
Is that the pattern?
And then OpenAI,
we can criticize OpenAI for many reasons,
but they took this and scaled it,
solved some engineering scaling issues.
DeepMind is meta today,
because it was developed there at first?
I think all the large companies back then,
you know, I was at Meta, so I know exactly.
There are many reasons why in a large company
there are things you cannot do
that as a startup you can do.
You can take lots of risks.
If you don't take any risks, you're dead,
because the big companies have more resources than you.
They have a brand,
and they have many, many things you don't have.
But the one thing you have that they don't
is the ability to make crazy things,
to take some risks.
And the reason, when I was at Meta,
for instance, we worked on a chatbot,
which was a language model, not large yet,
but the idea that we trained...
Just out of WIT.ai?
Yes, but trained statistically without rules,
but we trained from statistical data.
We worked a lot, we had a lot of fun,
and the first time we launched a chatbot,
in the lab at Meta,
we asked a question, how are you today?
And this chatbot answered,
oh, just chilling, watching porn on my couch.
And then, before that,
I wanted to release this chatbot.
That was not...
And then I thought, okay, the legal team,
they will never ever want to have this discussion with me.
And so we knew we cannot release it.
It's just an example.
And so OpenAI could take some risks with GPT-1, 2, 3,
you know, until...
And that's a GPT-1.
That no one else could...
So they have the ability to...
They have enough money and resources to scale the training,
but the ability to take risks...
Because they had this combination,
they made it what became a GPT.
So, just to finish on that,
what's the current bottleneck?
Is that training data, then?
So you need three things.
You need smart people, you know, talent,
to understand, because not so many people understand that.
And most of them are working on NLMs.
And most of them are NLM peeled, like Kian says.
You need a lot of data.
And you need GPUs, computes,
which is very...
Even when you have money,
it's very, very hard to secure GPUs.
So now you...
I mean, you've already hired quite a lot of talents,
and you've raised a lot of money.
So what's left? Data?
Data and compute.
Even with money, securing compute today
is difficult.
Okay.
It's kind of like a contrarian bet.
What convinced you to bet everything on it?
So I don't think it's contrarian.
It was maybe contrarian a year ago,
but I think it's becoming less and less.
A year ago, when we were saying
NLMs are nice and powerful,
but they are not AGI.
They don't lead to artificial, you know,
general or super intelligence.
Most people disagreed.
Now I think most people agree
that NLMs don't lead...
Maybe not most people,
but a good chunk of people agree
that NLMs don't lead to general intelligence.
So now you can disagree on what leads to...
I mean, the definition is also fuzzy.
The fact that the definition has changed.
So the very people who promised AGI
changed the definition
and stopped using the term.
I think it's a good sign
that something is not working.
And you're doing that with Yann,
and you met Yann at Meta, I guess?
Yeah, so yes, of course.
You were working with him before,
so you kind of knew him enough.
How did that go?
Did he coach you for the job?
How did you end up working together?
So Yann was an investor and advisor
in Nabla, my company prior to AmiLabs.
So I saw him very often.
And one day he came to me and said,
hey Alex, I think world models
are ready to become very big.
And I think we could go faster
if we do that outside of Meta
as an autonomous startup.
So I want to do it, and I need a CEO.
So...
Ah, you're a poet.
Yeah, I mean, I spent...
My first company was doing chatbots,
as I said, and after 10 years
trying to teach a machine
to understand human language and concepts
for customer support,
my conclusion was we cannot cheat.
We cannot...
A machine can speak about a cat,
but if the machine has never seen a cat,
you know, touched a cat,
played with a cat,
maybe be hurt by a cat,
then it will never be very, very smart.
You need, when we talk about a cat,
it actually reactivates in our brain
some past experience we had with cats.
And even if you read a book
about some animals
that you've never, like a dinosaur,
like never seen,
the way you imagine a dinosaur
is built on basic objects
that you had experience with.
So if you go, you know,
what we call grounding,
you need real experience of things
to be smart.
This is my...
It was my belief before, 10 years ago.
Just need to see, just bark,
and you're good.
Okay.
And with Yann,
how do you split responsibilities then?
Because he's still the inventor,
like the researcher
who kind of like came up
with that JIPA thing.
Are you more like on the execution side,
making sure we can build
the product out of it?
How do you split the roles?
So, yeah.
Yann is the executive chairman of the company.
He has a scientific vision.
And so he's leading the direction
of the research.
And then my job is to make sure
that there are people,
we have electricity,
GPUs, data, office.
So if it fails, it's because of you.
Everything is working, of course.
And I love this.
And also,
it's very hard to manage researchers.
You know, there is a joke
that it's like herding cats.
If you have to leave,
you cannot tell researchers what to do.
They hate that.
They will flee if you try.
But on the other hand,
if you leave researchers alone,
they will just do very random things
and run in all directions
and you will never get anything of them.
And so,
finding the very, very tight balance
between direction,
enough direction with enough freedom,
I think,
in order to get the product one day
is what I like to do.
Excellent.
You also ended up raising
the biggest ever seed round in Europe, right?
Like 1.2 billion.
Is that in dollars?
Close.
Like 1 billion in euros, right?
Walk us through that.
First,
could you ever imagine
raising that much money?
No, I could never imagine.
I think no one could imagine
a few years ago
that any startup would need
so much seed money.
The real craziness is the cost of compute.
And because you need a lot of compute,
you need a lot of money,
and then everything has shifted.
So, yeah,
this is a new territory for everyone.
What's difficult is
when you start a company,
I think having a culture
of not spending too much money,
be very careful,
spend money as if it's yours,
is very, very important for the company.
Very hard to change later
if you don't start like that,
being very cautious with money.
But it's hard
because if you spend 1 billion for GPUs,
then everything else looks like very cheap.
So it's very tempting to...
And how about finding the money?
Was that easier than expected?
It was...
not that hard, actually.
Here you go.
You should aim for billions,
not millions.
The rule is not the amount.
The difficulty doesn't depend on the amount,
it depends on if it's your first time.
Yes, of course.
I don't think anyone else
could have done that.
First time you raise money
for your first startup is hard.
And we've been all through...
I remember your story.
You have to go through this
unless you're very lucky.
And then second time is easy.
Just the YC, that helps.
Second time is easy,
and then after third time,
people just beg you to give you money
to do anything you want.
But you have to go through this journey.
And because the first few times
were successful,
that helps too, I guess.
I mean, with that size,
how do you set expectations
before you even start working?
I mean, does that money buy you
passions, or does it put you on the clock?
Actually, I told the team yesterday
that the real cost of this $1.2 billion
was not dilution.
The real cost is expectations.
Not only from our investors,
because I think our investors,
they understood and they agree
on the long-term vision of the company,
but the expectations you create
in the outside world,
and if you raise $1 billion
and you do nothing for two years,
or people don't see anything
coming out of it for two years,
then it's very, very hard to survive.
So the expectations is much harder.
And do you have a timeline?
You see expectations coming?
Yeah, of course.
You raised a lot, now you have to...
Internally, we have a timeline, but...
Are you going to see anything this year?
We don't communicate about it.
I tried, I tried.
And you do that here in Paris,
where pretty much nearly everyone else
is building these labs in SF.
Why Paris?
So we do this in Paris,
we also have a big team in New York,
one in Montreal, and one in Singapore.
We have these four cities
because we had existing teams.
I didn't hear SF.
And no SF, so it's a kind of...
It's a contrarian bet too?
A little bit, yeah.
I think it's...
When people join AMI,
we hire lots of people in SF,
in the companies you can imagine.
They have to move to New York.
It's like a commitment,
and it's very important at this early stage
to have people who join with a real commitment,
and not just because they want to try.
So it's half.
One day we'll have an office over there,
but it's kind of on purpose.
Do you feel you may miss anything
by not having a foot there?
No, because we are over there very often.
We live over there, we...
So you're still super connected to SF,
and so you kind of benefit from the ecosystem anyway?
Exactly, so I think it's...
Okay.
AMI Labs is probably one of the most ambitious projects,
probably the most ambitious project you have ever done,
but it sounds very ambitious.
What would you say to this room of talented
kind of like builders,
who probably are picking on less ambitious ideas than yours?
What would you tell them about how you learned
about being ambitious?
And should they be that ambitious?
I think we are never enough,
especially in France, in Europe,
we are not ambitious enough.
It's a cliche, but it's true, we don't take risks.
I was telling you,
just one month after we were acquired by Facebook,
I wanted to hire 100 people
to do some kind of data annotation for machine learning.
You mean like for WTI?
For WTI, yeah.
So I wanted to hire 100 concierges
who work for real customers,
you know, to book travels, for instance,
and record everything they do,
in order to train my model.
So it was in 2015, 10 years ago.
And I prepared the memo
to ask permission to Zuckerberg
to hire 100 people.
They were not planned in the budget of the company.
Back then, there were 6,000 people at Facebook.
Yeah, so 100 was not nothing.
It's not nothing, you know,
and 100 concierges in Menlo Park is expensive.
So when you meet Zuckerberg,
you have to send a memo 24 hours before, at least,
or your meeting is cancelled.
I arrived there,
and I was ready to argue,
to give every scientific argument,
why what I wanted to do was a great idea,
and I arrived, met him,
and he had read the memo before,
and he tells me,
Alex, I love the idea.
That's great.
I think we will kill Google
because we will capture the intent
with advertising and everything.
So let's hire 10,000 concierges, not 100.
No, no, don't do that, actually.
I had to tell him,
no, no, no, Mark,
actually, I'm not sure it will work.
It's very dangerous.
There are lots of unknowns in this project,
so we should not...
To this day, I don't know
if he was really willing to hire 10,000 people
or if he thought it was the best way
to get the real information
about what could go wrong in the project,
but I love the ambition.
And we hired 200.
200 instead of 400, okay.
More than you expected.
Well, Mark has been well-known
for taking wild, crazy bets on stuff.
Sometimes it works, sometimes it doesn't.
So I'm not that surprised he would go there,
but yeah, kind of probably erased the bar
from what you think is ambitious.
What should the audience remember from that?
Like, how high should they aim at their stage?
How high should they, like...
At one point, people seem too small
when they're starting,
like they're trying to find a safe idea.
Or should they actually aim
for something a lot more high-potential, high-risk?
If I could go back,
I think when you start,
you should look at a very narrow problem,
but be very, very ambitious in this narrow problem.
I think, including myself,
there's one thing I did wrong,
and I wanted to do a chatbot
that does everything for everyone,
so not narrow.
Very, very hard to provide value
as a very small team, small startup.
But if you choose a very narrow problem,
like doing one thing for one industry
or for one set of customers,
but in this narrow tunnel,
have a very big vision.
Don't be afraid to share your long-term vision.
Okay, I'm doing this today,
but this is what we want to do later.
This is what I would do.
You have only one life.
You're young.
But if you're very ambitious in a very wide field,
then you're not credible.
All right.
Just to conclude,
if AMI works,
no, when AMI will have worked,
what will the world look like,
like five, ten years from now?
Can you give us a glimpse of the future you envision?
Yeah.
First, I think we'll have robots,
you know, helpful robots,
actually helpful robots.
I think most jobs will still be here,
but not as dangerous or hard to do as they are today.
And yeah, I think machines with some sort of common sense
will make a big difference.
Thank you so much, Alex, for joining us today.
Thanks again.