Hi everyone.
Welcome to the Bright Signal Podcast, where we cut through the noise and bring you the
latest tech news and interviews.
My name is Marillo and I'm joined by my friend Bart.
Hey Bart.
And we just had a chat with Alexandre Pereira from twenty five one.
Yeah, it was a very, very interesting chat.
2501.
What they do is that they build autonomous SRE agents and bring them to uh large corporate
enterprises.
Why this is very interesting is that they, what their value proposition is, is that they
basically do root cause analysis on anything that goes wrong on their infrastructure, but
also try to automate the remediation.
which in these type of enterprises is typically not easy because they have a complex
infrastructure, a large part on-prem, part in the cloud, part of it modern, part of it
legacy.
So it's a very complex environment.
And if you can execute something like this correctly, then it becomes a very, very
interesting file proposition.
Indeed.
Yeah, was a very nice chat.
So yeah, let's go to the interview.
to the interview.
welcome everyone.
Hey Bart.
And hey Alexandre.
Alexandre Pereira.
Is that that that sounds good?
How are we doing?
Warm outside.
Or yeah yeah yeah.
thanks for joining us, Alexander, and welcome to the Bright Signal podcast.
maybe you can dive in already and like for the people that don't know you, like would you
like to introduce yourself, share a bit of your background and
Yeah, sure guys.
So Alex, I'm obviously French.
You will notice with the accent.
I'm like, let's say an entrepreneur since 10, 12 years now.
Working essentially on tech for enterprise.
Also spent the large majority of my starting career into corporate world, launching
e-commerce websites, working with like these big brands in Europe, in France especially.
Doing that like for the last 20 years now and...
It's very cool spot to be.
And since two years now, we work on AI, ah on agent-y AI especially for corporate as well.
So that's gonna be one of the topic today.
yeah, that's it.
So I would consider myself as a, I don't know, a mix between the tech guy.
I'm not an engineering by training, but I'm an engineering by love and trying to code when
I can.
But now, yeah, I'm not coding anymore so that much.
I'm just trying to make this thing move forward.
and do the best product we can.
So you don't you didn't study any like uh computer science or anything like that.
You just like on the side, you're just tinkering with it and and and learn by that by
doing.
Yeah, I was kind of a cliché nerd when I was a kid.
ah My parents and my family were like super nerdy as well.
I was playing video game like on Sega Master System with my grandparents.
So that was kind of fun.
I started to have my own computer quite early, coding, coding to pay, you know, I was
having like this 125cc moped and I needed to pay everything for it, so at some point I was
like, should you work in McDonald's or should you work in tech and do websites for the
groceries or for the bakery?
I started to do that and ultimately I landed in this company that is named Decathlon that
you may know.
It's a sportswear group.
The end of marketing of this brand there was completely crazy and it was like, okay,
you're 18 years old but you look fun and you know how to code websites so come to work
with us.
I was not having any other experience than that and that launched the thing, so yeah.
Super cool.
You also have a sim something similar, right Bart?
Also tinkering with the stuff and you you you went to officially to tech later, right, in
life as well.
There are some parallels indeed, yeah.
Okay, cool.
And also how did you so we're gonna talk about uh twenty five one or two thousand two
thousand five hundred and one?
I don't know how twenty five one, yeah.
Everyone is asking about it and I think we figured it out only two or three months ago.
We're like, everyone is asking, how do we pronounce it?
And we're like, we don't know.
So we have to choose.
So 2501 in English.
Okay, there we go.
And uh it's it's yeah, we mentioned agentic, right?
How did you how did you get in touch with uh agents and all these different things?
Did you just jump on it early on or?
So, yeah, if you want the genesis of the project...
Basically, we were trying to create LLM chaining technologies at the very, very start of
it.
I don't know, we were on GPT 3.0 or something like that, 3.5 maybe.
You started to have these guys online on GitHub doing what we call agent orchestration.
It was not named agent at this time.
It was name, I don't know, LLM orchestration or chaining LLMs.
And you had, I don't know if you heard about this kind of GitHub repositories like
BabyAGI, AutoGP.
Auto GPT went crazy.
I think was over to GitHub's fastest repository to get on.
And they were just doing something simple.
They were just chaining and looping on LLM response.
And so it was just very, very basic at the beginning.
But it was like the start of what could be Agentic.
I'll call it Agentic further.
So yeah, it started all of this kind of crowd, started to innovate a bit like in startup
garage mode, GitHub.
So we started to do the same with one of my colleagues.
just like trying to see if we can do something.
And I was not at all in this mood of, I don't know, creating a new company or whatever.
We're like just playing with the tech.
And I don't know, a few months after, early 2024, we started to see fundraising happening
in there.
Mm-hmm.
So I think the one that was significant and that was kind of the compelling event starting
everything on Agentic was Cognition Lab DevIn product launch um that raised, I think, 100
million from scratch at the beginning by...
um
US.
So everyone after that was like, okay, there is something interesting because he found
this fund, he's interested in this kind of thing that could be very good maybe in the
future.
So it started to get credibility to these GitHub repositories and all these innovators
that were trying to do something and trying to make sure that, I don't know, the VCs
getting there put a lot of legitimacy in the end.
And so I started to do the same.
was like, okay, maybe there's something to do.
And at this point I was a bit at the end of my previous entrepreneurial story, which was
in Japan.
And I was like, okay, let's look at what we can do.
Ultimately, like in a few months, we raised our first million euro that became like a two
million round precede.
And then we went into that.
Then I can explain later how the construction of the product, but the very origin of it
was just like pure innovation research with people trying stuff with this technology.
And I like that because we are entering a phase again with AI where we innovate a lot
based on pure, how to say, pure random ideas sometime.
And this chaining of LLM was not at all anticipated by guys.
OpenAI didn't anticipated it, Anthropeak didn't anticipated it.
It's just like two years ago.
that Sam Altman says, no, no, the chat is not what will be the end product, but there is
something else cooking.
And there is just one guy stupidly saying, but maybe it can generate code.
And if it can generate code, it can generate action itself.
And so here we go.
So.
So that means that when you created the company, you didn't really have this exact vision
that you have today, like what you want to do.
It really started from you see that there is a value in this and we're gonna see an
experiment and innovate and see what direction we can take this.
Yeah, that was basically like this kind of thing.
We were like, okay, how can we create something that is autonomous in a computer terminal
system, in a shell script system?
And how we can make something that will be, don't know, the next AIOS or the next
developer experience?
So everyone was kind of touching everything everywhere.
was like two years ago, so Cloud Code was nothing.
You had like bit of projects trying to do this with, I don't know, Copilot was trying to
do some stuff.
You had like some guys at Y Combinator starting to create companies similarly, like I
don't know, Pythagora, Marble, that has pivoted or died since.
And what I liked is that, yeah, it was very, very R &D driven and...
the need started to emerge after that.
And it's not a need that is a new need.
You don't create a market, and I will come back to our market later, but the technology is
so huge, so powerful, that you can apply it to a lot of stuff.
And this was what was the, know, kind of the guess, the gut feeling about it with the
first investors and with...
colleagues were like we're gonna figure it out it's just like so big so huge it's a
once-of-a-lifetime maybe once a system to every revolution right now that we're that if we
play well this thing it might work
And you said like we like your co-founders as well, like or was it like back then when you
s were being curious and like trying things out and tinkering, was it already the the
co-founders?
Was it like colleagues?
Was it just friends?
How like what what are the when you were sharing these ideas, who what was the team or who
are the people that you were discussing these things with?
So at beginning we were like kind of racing with one of the guys that is my friend.
He was having his project, I was doing my project.
We were like a bit competing every night trying to code what could be the next thing.
Ultimately, he joined me on the project because I was like, okay, now I think we have
something interesting.
I have a bit of pre-seed money arriving.
Join me.
We don't compete.
We continue together.
And he was not the co-founder, but the first employee of the company.
And then my co-founder Alex joined me a few months after from his current thing.
So yeah, it was a bit of a story of just...
I don't know, coding and seeing how it goes.
And Investor, we are like in the same way.
We are super lucky to have these guys that joined on as well because we got like, think,
what is the best configuration of Proceed Investors.
And we're just super supportive on, yeah, you're gonna figure it out.
And that also means that you got your pre-seed round, also really before there was a very
strong ID to sell.
Yeah, and that's what it should be, in my opinion.
This is the thing, know, the VC world is a bit like paradoxes because there is always a
reason to not invest.
There is always one million reasons to not invest in a company.
But in the end, if you look at pre-seed rounds, what do you look for?
You're looking like at co-founders, team, a bit the vision, but the vision can be a bit
blurry and gross still.
But you're not looking at any market trends.
You're not looking at, okay, market research can happen, but not if you are purely
tech-driven, R &D-driven.
It's different than just if you launch a service company.
Yeah.
Maybe to make a step to today, like what is the product that you're offering today as
2501?
Yeah, sure.
So if you want to summarize what we do in a nutshell.
we automate what is IT maintenance and cloud maintenance of infrastructure for large
corporates.
So our job is to make easier the way to manage this large hybrid cloud on-premise data
centers, IT corporation infrastructures.
So in a way that this thing has not changed for the last 30 years.
It has been a bit modernized.
They did a bit of automation in the job.
The arrival of the cloud changed a bit on stuff, but it's still very complex stuff that's
happening.
If you look at S &P 500, CAC 40 in France, these guys have large infrastructure, super
difficult to manage.
By design, they are not state of the art because they their product.
They are not a tech company.
We are not speaking like the NASDAQ companies.
We are speaking like the whole guys that are managing the main part of the economy.
so it has been like for the last decades a very kind of simple job because it was human
led.
A lot of shoring was done as well.
And right now this world needs to change.
what we've seen is that technology like agent I can kind of help this transition and help
these guys to modernize their park to modernize their processes to optimize what is like
their velocity of production release their maintenance responsiveness and Also reply to a
lot of topics that are like quite crunchy right now like sovereignty of data.
So at all so there is a
of trends also right now on trying to avoid offshoring, trying to get back on EU soil, on
the US soil, make the data private again, back to the data center, back to the on-prem.
So this is all this world that we're serving.
So we have a big agent-tki solution that is now able to manage all of these services that
are like not always AWS, GCP, Azure, but could be like some VMware server, some Fortinet
firewalls, stuff like that.
We do kind of every technology that we can connect to.
and serve the spark that could be sometimes complex because sometimes it's fully on-prem,
fully air-gap, fully privatized, no access to the internet at all and stuff like that.
So you need to create a product that is adapted to this kind of large variety of
technologies that is covering and also to the privacy concerns and the privacy
restrictions there because they are often like government-related, vital asset-related.
So when you work for banks, energy companies, government-related assets, et cetera, et
cetera.
So, $25.
one is all of that.
We serve our customers by pushing the most optimized solution for them.
That is completely agnostic in terms of technology.
That is completely agnostic also in terms of LLM usage.
We can plug ourselves to kind of any LLM that is, uh let's say, agency credit.
And...
We deploy that in a few hours to our customers so they can automate everything that is
going down or that they need to maintain in their infrastructure.
if you are like, don't know, BNP Paribas, you have a server going down at 4 a.m.
in the morning, some customers will feel this impact because they don't know some part of
the application doesn't work, stuff like that.
Instead of waiting one hour that someone would wake up and take the ticket in the ITSM
systems, it will be automatically done or at least automatically started and investigated
by the AI.
maybe you made like last point how how far is the LLM or the agent autonomy goes?
Do they also rem like make fixes or do they just do the initial investigation?
what to which extent gets the does the agent have autonomy to interact with the system?
Yeah, so our philosophy is that agents should be autonomous.
This is what we are kind of obsessed with that.
And our design of products is quite unique because of that.
And this is maybe what's pushing us a bit ahead of the curve.
We don't build workflows.
You will not see in our interface some point and click complex things where you need to do
if else do that.
It's quite very simple.
It's a lot of prompt and securities that we build in the agentic system, but then
everything is autonomous.
You can put like restrictions.
So of course we have a human industry.
loop mode where the agent is restricted from doing action, but still all the commands are
generated by the agent and it's just an appreciation of command that will use, I don't
know, LLM as a judge technology or stuff like that to ensure safety of the product.
So yeah, we want the agent to be fully autonomous.
That's a bit our goal and we want to share that with our customers because otherwise ROI
will be lower and also technology will be less kind of innovative.
So we push forward this with our customers and this is quite important for us.
But at the same time, we build 50 % of our work is about safety.
50 % of what we do is human in the loop situations, how to be sure to inject the right
business rules, the right restriction rules, blacklist, how to measure performance of the
agents, et cetera, et cetera, cetera.
So it's not just about like the AI.
Mm-hmm.
products are just AI and there is nothing around it but it's 50 % or so of tool set that
we give around the product that is allowing the customer to feel safe to measure and to
not do things that will go wrong.
And it's an interesting approach.
I perhaps like crack me if I'm wrong, like classify this a bit as AI as RE.
so site reliability engineering, where there are a number of competitors, but I think a
lot of your competitors, their, their, their value proposition is more root cause
analysis.
It's like pointing at this is going wrong.
but not necessarily this remediation phase that you're discussing as well, right?
And I think what you're focusing on is the two.
And this is why also we kind of are feeling a bit six months ahead of the curve.
Or maybe we're just a crazy guy in turn, I don't know.
Because yeah, from the beginning, well, no, we should not just observe.
Of course, it's obvious that everyone can observe.
But the value of agentic is about doing things.
This is why we build this agentic.
My job is to just watch the agent work and not touching my fucking keyboard.
that's what it should do.
So because we are onto that since the beginning, we have built kind of a framework that
works super well on that.
Everything is in house built.
And this choice is a bit risky because in sales, you will have like...
more, let's say, it's taking more time to convince customers and to show that your product
is safe, et cetera, et cetera.
So it could be like a bit more difficult to sell.
But when you are in, the value proposition is insane.
And the ROI, when we send ROI matrix to our customers or leads, it's like, it's obvious.
This is more like, okay, in three years, course it will be the standard.
Because you put so much value back to the customer that they have no choice in doing that.
This is where it's interesting.
It's very early days of agentic.
Adoption is still ongoing.
But at the same time, it's unanimous that everyone says it will be done like that.
Maybe to double click a bit, like you I understand what I I think I can definitely agree
that it may be hard to convince people there's more scrutiny, right?
When they are when they have like agency, basically.
What are how do you how do you get people over the hump?
Like what are the what are the things that you can show them to say like okay, it's not
gonna it's not gonna mess up your infrastructure, it's not gonna make a bigger mess out of
the the incident.
What are the things that you can that you can share that you can say to them to get people
to get in, basically?
Yeah, so first, I'm not saying it's never gonna fuck up because one day I will have a very
bad call and it's gonna be a bad day.
But it's the same with if you hire some engineers offshore or even near shore, even in
shore, it's the same.
Someone's gonna fuck up someday and your infrastructure is gonna be down.
I leave that.
thousands of times in my life when I was working in big corporates or even when I was in
the startup, the bad call on Saturday night will happen.
But it will happen maybe less.
It will happen maybe in different stage.
so we are like...
saying that AI is not less intelligent than human.
It's actually on this kind of task more precise often, depending on the model of course.
We can debate on the model stuff and the model choice.
But let's say we are convinced with our customers that in the end, a human being will make
more mistakes than an AI that is well done.
And if we look at new AI technology like cloud.
these new models and even smaller models like if we speak about like the latest Quen
models, Zai models, GML, it's like these guys are crazy.
It's the quality of work that is done by the AI is very, very good.
So it's quite an optimistic view of what's gonna happen but I'm pretty sure that on the
statistic level once everything is on production, we have like, I don't return of
experience on thousands of hours of work, et cetera, the statistics would be in favor of
the AI.
I'm 100 % convinced of it and almost of our customers say so as well.
The second part is that you need to bring the tool set as I was saying.
You need to bring not only the AI, but you need the tools to control the AI, to measure
the AI, to make sure that we are not having something that is wrong going on.
so for that, for example, we built one of the first benchmark engine that is able not to
monitor only if the AI does the job as it succeeded the task or not,
also measure the compliance of the tasks that have been done inside the job.
So we measure every part of the job done against scenarios that the customer can customize
and say, okay, am I seeing this task and this part of the task and this part of the task?
Is this forbidden things never happening because we don't want to see them, et cetera, et
cetera.
So the agent is able to work in full autonomy and the customer can monitor them and say,
okay, this type of task that I want to see in production now is stable.
is going well and we can continue to push it in production.
And if there is an anomaly, you go to this qualification environment, to the sandbox
environment, and you can see the anomaly happening because I don't know, you change the
model, you change some points here and there, and there is maybe a bad impact.
As soon as you will see that, you can stop your agent in production, find the fix in the
sandbox, and then replicate.
And so we bring the process of managing these things.
It's still early days.
We are building as we fly.
But at the same time,
It's the new way to work.
You were speaking about these SREs or C-SOPs guys in the corporate world.
Their job is going to change.
They are kind of changing.
Before they were piloting human work or they were doing themselves like production
engineering, for the run engineering.
Now they like evolving to be AI pilots.
And this is a new world.
This is a new game.
We are discussing with our customers every day.
And it's so cool to see these guys that were like before just doing installation by hands,
map notes by hands, that use this knowledge that they have accumulated to guide AI and to
build the best AI systems with us and that will do the job for like a fraction of the cost
and that's a fraction of the time necessary.
So, yeah.
Okay.
So maybe maybe to to rephrase what you said, put in my words to make sure I'm following.
Like you're saying, you're not promising that the the LEI will never make mistakes because
but just the same as people do, people have done for a long time.
And you really double down on also the traceability to make sure you can check what
happened.
Let's let's let's give the visibility and give that feedback loops to make sure that these
things don't happen again and all these different things.
It's more auditable as well, it's more maybe even predictable, I guess, 'cause uh
agents if especially if you have open models, it's uh it's easier to to achieve
reproducibility with people, maybe it's more blurry, right?
Does that does that sound about right?
Yeah, totally.
We cannot lie saying it's perfect, it's never going to do something wrong.
It's a lie.
Don't do that and no one will trust you if you say that.
But yeah, everything is in how to make sure that the technology will continue to evolve
and be more and more qualitative and now we are, think, in a kind of a murlot of model
accuracy and we cannot even release models anymore without thinking they are too strong
now with the mythos.
events are arriving.
So I think the model technology is kind of still evolving quite fast, even if it's maybe
not anymore the exponential curve that we've seen three years ago, but it's still very,
very impressive.
And there is a new generation of models arriving, a lot of research in world models, et
cetera, et cetera.
So it's going to be better and better every day we speak.
But yeah, the need of having products is that a lot of people think that AI is just AI,
end of the story.
And you can see a lot of companies that are just having PhD guys doing incredible job on
models, but then on the application side, there is something missing because you need to a
good harness, you need to build a good tool set, et cetera, et cetera.
So.
you, you, touched a bit on sales.
like, uh, it's, and I can imagine that you're a bit of this dichotomy because you're like,
assume that your typical company that you sell to, like, they are very focused on having
robustness, having reliability.
And because AI agents and especially autonomous agents are still so new, like the Chris,
this question, like, will this not
take our reliability down, even though your value proposition is actually to improve it or
detect issues more quickly or resolve them more quickly, right?
Like you have this, you're in this time period where we're still gathering data on to
prove, to basically make this argument it's definitely better than how we used to do it.
Yeah, so on our side, is a way that we have, this is something that we have implemented
directly in the release process with our customer.
So instead of saying, we are going to show you proofs, we build the proof together with
the customer.
So when we onboard the customer, the customer is entering what we call a sandbox phase on
which we work with the customer, like next to them, they have access to the tool, they
build things with us.
And this is important because this is when they see that by
to prove that it's working.
And they continue the product themselves.
We train them, we help them, et cetera.
But in the end, they become this AI operator.
So this is very interesting because as I was saying, you see a lot of top managers guys
that are like reading Gartner's or Forrester every day.
And these guys will, of course, be convinced that in the next few years, it's going to be
the trend.
It's going to be the default solution.
And everyone agrees on that.
But then you need to make sure also the entire teams, especially large corporate, agrees
with that vision.
And if you lose the acceptance of the vision under the...
the big C level people, it's a bit a mess also to implement.
You will have adoption issues and we've seen that also in the company.
We've seen in some customers real adoption issues because it's a change management for
them that is quite critical to happen, especially when you are an IT company.
Your job is completely upside down since a few months, now.
So the only way is not to have the best product on them but also to onboard them, part of
your team kind of, and use the product as quickly as possible to big
convince themselves that what they do is providing value and this is going to be the
default for them.
Yeah, and I think that's what you're also doing is you're uniquely positioned in the sense
that the...
the skill that you're productionizing, like you're at the frontier of what is possible
with autonomous agents.
And at the same time, you have a strong expertise in how these large corporates, how their
infrastructure looks like, which is typically not the most modern cloud setup, right?
Like they have a lot on-premise, they have a lot of old school hardware that they're
running on.
And I think that combination of skills and combining them in a product is quite a unique
position,
It's even harder to find just people, individual people that know about these two things.
So yeah, that was actually a fun thing.
yeah, you would see on the competition mapping that this enterprise, AA agent equal is
quite still a blue ocean, which is surprising.
Everyone is SaaS cloud based.
This part is overcrowded.
On this you have like half of white communities two years ago was there.
You still have like a lot of companies that raise actually a lot of money because there is
still a huge market there.
Still have all these retail brands that are completely cloud-based, AWS still.
These guys are doing an extremely good job at doing this on this perimeter.
But on our side, it's very interesting because you are on this whole kind of world that is
still the default for the majority of the companies.
I was super surprised when we are in New York often we speak with guys and a lot of people
were saying, okay, how much on-prem is JP Morgan Chase?
60, 70%.
Goldman Sachs, 90 % on-prem.
All of them.
All of them.
I was in the Office of Societe Generale a few weeks ago.
It's still majority on-prem as well.
oh And this is what no one has understood is that there is still a large spectrum of IT
infrastructure that is not based on the cloud, that will never move to the cloud.
It's not designed to be in the cloud.
How do you move ABM, AS400?
machines, yeah, no, it's not possible.
It's not possible.
And it's not the work of AWS, of GCP, of Azure to do that.
They are not able to host these things.
So how do you serve this market?
And this market is well underserved.
It's still like all these actors that are still here forever, SAP, IBM, Fortinet, yeah,
ServiceNow.
These guys are here.
They are not moving.
And this is like a huge chunk of the IT services in the world.
So yeah, when you are speaking about the unique skill set and the unique understanding,
this is quite true because we were lucky to be surrounded by people from this world from
time ago and I was at the Bayer site a few years ago, so I could have understood both
sides of the game.
But I was so shocked that it's still the case.
The on-prem majority component is still like a huge thing everywhere.
And no one is like kind of interested in that part because it's super painful to interact
with.
Connectivity is not super easy.
Sales cycles are super long.
We are speaking about like six to 12 months in sales to sign a deal.
This is why we kind of raised our second round because we needed to be strong enough to
close our deals.
So yeah, and you need to speak to some people that have kind of a unique culture as well.
When you speak to these IT services persons, it's not like the usual startup client that
you're gonna have.
These guys are super, super critical.
They know their job super well.
They are here forever.
You speak to guys that are here since.
30 years more sometime, they have seen everything.
They have seen like the dot-com revolution.
They have seen like the mobile revolution.
They have seen everything and they're like still here doing data center stuff.
okay, so it's very.
all the time, the mainframe just stayed the same mainframe.
These guys are like, yeah, I've been here when we did the first automation for Sokira's
code with Terraform and I'm still here doing now the agent-y AIOps transition.
ah And so that's very cool because you speak to some people that have like a large
understanding of what is tech, but not shiny tech always, but very useful tech for the
world.
Maybe I have a question also.
Like uh I imagine that for for LLMs it's much, much harder to to debug the scenarios.
And I mean, 'cause just thinking of documentation, like the the the the data that it was
trained on, right?
I'm sure there's it's very disproportionate.
There's way more talking about AWS, G C P, all these things.
Have you do you have any comparisons, views or anything?
Like, yeah, like like I don't know, um let's say minimax.
without much fine-tuning all these things does very well on debugging AWS and on-prem we
see that it drops a lot.
Do you have any any more numbers, any more details on that?
Yeah, I have a good example on this, which is, so.
For the last, let's say, 12 months, and especially until very recently, our best player of
recommendation was Quen, because Quen models are small, light, efficient.
They have kind of the right balance between size, accuracy, cost, running on custom GPUs.
So we were all in Quen.
We are completely agnostic.
We can work on any model, but this was our top one favorite model to push on.
So our customers were like, which model we should choose if we have the choice?
We said, start with Quen.
Up to 135B, 80B, next.
Now we have new ones arriving since a few months.
So these guys are the best.
But when we tested something, which was like managing Windows server, we noticed that Quen
was not that good.
It was hallucinating a lot in, I don't know, versioning of PowerShell and stuff like that.
And we tested some alternatives and we found that GPT-OSS, the open source, open weight
LLM of OpenAI was performing well better on PowerShell commands, Windows Server commands
than Quinn for our use case.
Maybe it's not absolutely, but our use case was the case.
And even if it was a smaller model.
So we were like, okay, maybe there is like some interesting conclusion to have that maybe,
I don't know.
It was from that gut guessing because we are not in this
companies doing the training with them but potentially GPT and OpenAI being closer to
Microsoft are having more data set on Windows than the rest of the world.
Maybe Chinese companies that are training mostly on public resources or on own inside
resources have less training on Windows server documentation and so that was some
assumption we're having and now we see that
GPT models are good for some certain type of task, et cetera.
I don't know if you heard about the skateboard benchmark.
um It's quite a niche thing online.
You should search for it.
So there is something named the skateboard benchmark.
And if you compare a Chinese model to an American model, the American model will know all
the tricks of skateboard, where the Chinese doesn't really know it.
Yeah.
because the training corpus is not the same.
And that's cool because I think the English models and the American models especially are
like training on everything that's online about skateboarding and stuff where the Chinese
maybe social media is not that much on that part.
Exactly.
And what we've seen is kind of equivalent of this skateboard benchmark.
We've seen that Chinese model will perform maybe less on Microsoft products.
Mm-hmm.
But aside of it, if you go to the extreme position of this, as long as the model has,
let's say, public corpus to train on...
it's quite okay to work on and you can work without that much fine tuning, just like a
good prompting engineering that we do inside the products is enough.
But when we see that some products is exotic, and for example, there is a very famous
firewall in Europe named Storm Shield that is used mostly by government in France and in
Europe.
This one has almost no documents online.
And engineers working on these products are super niche.
There is only like a 600 page product on their website that is uh the documentation of the
product.
And so on this, you can try every model you want.
Even Claude cannot work on that.
So on this, we start to feel that, okay, we will need some fine tuning.
We need to put more adjustment, reinforcement.
That sense.
Maybe one last question about model comparison.
You mentioned Claude as well.
How from your view, because I imagine you have a very good view on this, what is the gap
between open and closed models?
Like I'm thinking uh GDPT and and Claude versus Chinese models.
given demo accuracy.
Yeah, yeah.
So for example, I don't know, uh you have this client and then you say, Okay, you have
this type of problems and you say, Okay, Cloud I'm pretty sure is gonna do well, but maybe
the open models not sure.
I imagine the open models are still behind, right?
But like is the gap still big?
Is it is it is it noticeable or like you think it's it's the we're closing the gap, like
open source and open models.
So I would say the gap gets thinner and thinner, but still super visible.
You still have these flagship models, foundations, rock stars that are above the crowd
because their model size is like, we don't know them.
And they have this kind of weird architectures that are super, super complex to make this
happen.
When you speak to Opus, when you speak to Fable, these kind of models, it's not
comparable.
You're not speaking to a 200, 400 billion parameters model that can sit on a few GPUs in
the data.
center you're speaking to some stuff I think people don't realize the level of complexity
that this flagship model has just running on something compared to what is done on the
small open source open wide models and this is where it's impressive because if you you
compare I don't know the Concorde plane with the small touring plane that your grandfather
is driving on the weekend you know
Yeah.
It's still flying, both of them, but there is one that is doing Mach 3, 3 hours to New
York, and the other one can just fly you 200 kilometers to the lake.
I think this is the kind of thing.
It's still a plane.
It has an engine and two wings.
But we are not speaking about the same level of complexity at all.
But this is where I think it's super impressive to see these kind of models going shrinker
and shrinker.
And there is now two different ways to do it.
That is...
Do you need a flagship usage?
And on this case, you pay per token, you pay per intelligence, and this is a use case you
want on some case.
For example, code development.
All the guys now doing coding are using Cloud by default almost.
And you don't want to pay less.
You're very happy to use Cloud.
It's super intelligent, does the job well.
When you use your Open Cloud virtual assistant or digital twin, same.
You want to use Cloud because it does the job super well end end.
When you are in more guided space and when you're looking for efficient
when you're looking for cost performance, when you're looking for privacy, which is a
huge, huge topic also.
Main reason why we don't use Cloud is privacy first and then cost after.
And this is where you can start to build on open-weight open-source model that are
smaller.
On this game, people are trying to make the most compact models.
for me, it's two different games.
It's not anymore competing together.
I don't think it will ever compete anymore.
Interesting.
I can imagine that you're maybe you can shed a bit of light on that like your typical
customer like do they prefer open weight models that they can host locally on their
premises?
do you see a preference at these large corporates?
Okay, so there is first rule that is universal nothing goes outside the corporate world We
don't use any public LLM for none of our customers We do it in demo we do it in some test,
but that's it at the moment We are installing the corporate we use either their inference
partnership with Amazon Bedrock
because they have a private contract with them and there is like a deal or we use on-prem
GPUs on the customer side.
So that's the first rule.
Nothing goes outside because by design it's not secure for their data.
And if you want to stay like socked to and stuff, you need to respect that.
And especially if you work like in, don't know.
health situation, defense, corporate related or critical assets and all of them are almost
critical assets.
this is mandatory.
And then you see the planning now that we have with majority of our customers that year
one, we use an inference partner because it's easier to deploy.
don't lose time.
Year two, we move from OPEX to Capex and we move like investing in GPU.
So in the end, the vision is to go into open-source models.
And also what we do when we use the inference partner is to use this type of model as
well.
We don't use Claude, almost never.
Yeah.
use Bedrock, we will use a bit of mix of everything, GPT-OSS, et cetera.
On Azure, we use Quenolot.
So it's more like we start paying the token on this market that we target.
It allows us to show to the customer that it's working well, that it's not that difficult.
We don't need to invest day one in GPU, et cetera.
And then we move into more long-term situation from here too.
so you're also saying that long term for, you have this also this vision to build your own
data center to host your models.
How do you see this?
So yes and no, because it will not serve the customer needs that much, because we will
become a liability for them.
Yeah, yeah, see what you mean.
oh
No, our vision is more that customer environments are quite complex and sometimes even if
you are in a big company that has announced a partnership with Mistral, these business
units are very split and separated.
So if you work with the business case, they have already Mistral, GPU and stuff.
But if you work, for example, with the IT infrastructure, this guy maybe doesn't have a
contract with Mistral as yet.
And so sometimes we kind of fail in this situation and we try to fill the gap.
Yeah, okay.
So our job is to provide first the agentic solution and to make it work well.
This is 90 % of our job.
But we have also this capacity to bring hardware if necessary and to deliver to the
customer the capacity in GPU to host our models that we need.
Interesting.
And maybe to briefly touch upon the open weight models of today, because like you and I
think a lot of people are leaning on those very heavily, especially for these repetitive
tasks, guided tasks, where they are very cost efficient basically.
How do you see the future of these things?
When I hear this, always a bit worried in the sense that it costs a lot of capital to
train these open weight models and they're...
more like most of them are run by for-profit companies that release this today for
marketing purposes because we don't trust DeepSeek for whatever reason to BMP will not
open a contract with DeepSeek tomorrow, but they might with Mistel.
So there is some marketing that needs to be done.
Maybe there are geopolitical reasons to be competitive towards the States or something.
But like there's, I don't see a very long-term future for these.
these models that require so much capital to stay quote unquote open source.
So you're right, there is like something still to dig on that part on how the...
small, middle-sized model business will continue to exist?
Is it something that will continue to stay sponsored for the sake of it?
You have also a lot of sponsorship happening when NVIDIA sponsors a lot of people with
free GPU usage, etc.
It's actually marketing that is done also for these inference partners.
If you're like Base 10, for example, you want to sponsor some guys on a game face to
deploy the models.
Because if one of them is going to be a hit,
it's going to be good marketing for you because you're going to train these models as
well.
So I would say that there is always an interest somewhere to do it.
ah And I would say that it's how good research should be done.
It should be independent.
Look at what has been done with DeepSeq.
DeepSeq is a very good example of a model that has been a bit oh cheating.
But at the same time,
of a distraction of the ghost.
Yeah, it was a proof of concept that distillation is sometimes better than training.
And that's actually something super interesting because ultimately it was not the right
way to do it.
But in the end, I don't know what they did is not, they showed that we should be able to
train models more efficiently without kind of spending that much money.
And so now,
People don't do that this way, they did it with Anthropic, but they tried to find a way to
replicate this philosophy of distillation.
So I would say.
It's a balance.
We should have people innovating.
We will have newcomers.
We will have guys stopping to do it.
And probably maybe at some point you're right.
The new wave will be Mistral, OpenAI, Anthropic, just releasing small models that are not
very time expensive to run.
And this is just part of their product lineup.
That would be maybe the new standard, but I don't see that for the moment opening.
If you look at the only one that has succeeded to do that is GPT OSS with OpenAI that is
how to say super small market share still.
If it was not Windows Server, we would not use it on our side.
Mistral is, I think, trying to find their business model because a lot of the revenue of
Mistral is services and engineering model tokenization.
Let's say selling API is not the majority of the revenue of this file.
So yeah, let's see how it goes.
It's still, think, building as we fly.
Sure, yeah, there as well, yeah.
Talking building as we fly.
Can you give us a bit of a view on, say, where do you see 2501 going in coming two years?
What are your big focuses?
That's a good question.
Now we see things a bit everywhere.
when we...
So our value proposition and the core thing we're doing right now is remediation.
It is to do maintenance and fix of incidents automatically.
So we focus 100 % almost of our time on this.
our job is to do remediations to the incidents and maintenance request management.
Observe, remediate, report.
This is what we want to do in a nutshell.
and we try to be the best at doing this because there is a lot of ROI for the customer.
Then you can extend in a lot of different spaces.
There is naturally the observability world that is super, super linked to us on which we
have some part of the product that start to compete with some people there.
For example, we start to have a releasing quite soon CMDB auto exploration, which is like
a huge pain point for our customers.
And there is no natural lineup of product that is solving this issue.
So which we know that we're going to like go a bit upper layer on observability.
if we're gonna go full scale there, or we're just gonna just touch it a bit.
But there is a lot of values there, because you have companies like Dynatrace, have
companies like Big Panda, Datadog, all these guys are here.
So we don't want to compete with them, but there is like maybe a sub layer of their parts
that is interesting maybe to absorb.
You have like the CyberSec world.
Enormous and with these new models arriving the cybersec is something super super super
big to consider budget in cyber security is increasing like I don't know we never seen
that the new mythos release has put in the highlights some big problems and now cyber
security team have a problem that there is Every day too many threats arriving that they
don't have like any more capacity to follow up and patch everything oh
It's gonna be something interesting and our technology is perfect to do that as well, like
cybersecurity audits, cybersecurity compliance check.
So see sock sock remediation everything that you try to do on cyber so We don't go there
yet because we feel we don't have the credentials to do it It's not our focus and we
should show some battle at some point So we want to stay on the remediation of IT
infrastructure, but we are for example some customers already I was discussing with one
yesterday that are like okay inside the use cases.
I'm gonna use two five zero one I'm gonna put some vulnerability vulnerability tests we
feel that cyber is gonna be a big topic for us as well in
coming years and and on the below I was saying before the hardware also there is a lot of
difficulties around this world how to build the hardware that is in capacity to host your
models for the usage of the models what is the hardware looking like there is today a lot
of
I don't know, under-optimized infrastructure on AI usage.
You will buy too many GPUs, you will not use that capacity, you will not optimize your
model correctly.
So there is also a lot of stuff.
On this, it's a very small part of our job because we don't do it for every customer.
We do it for very few customers that are using it.
But we feel that there is a small need there.
But I would say our big topic is observability and cyber next to what we're doing right
now.
But it makes me think, because again, you have a very interesting position there where you
can traverse a bit into different fields because you start from this, you're at the
infrastructure level.
You're doing observance there, you're doing root cause analysis.
It's not a big step to say from there we're also going to do more in the cyber sec, we're
going to do firewall monitoring.
It's not a big step to make.
on the other side, if you would start with a firewall anomaly detection and then say, I'm
going to go the other way, I'm going to do now, do infrastructure failure detection,
that's a way bigger leap.
It's an interesting point to start from and to further build up.
Exactly.
It's way more natural to start from infrastructure and then to go on these things that are
observability, cyber and audit.
Because doing the opposite is more like you need to go back to the central core problem
that is infrastructure itself.
And in terms of sales, it's better as well because you're already working with the key
decision makers that are like the CIO or the CTO of these companies.
And if you're doing well your job on the core infrastructure part, naturally you're going
to have opportunities elsewhere.
Maybe a last topic, fundraising.
You've been quite successful in fundraising.
Can you share a bit your approach there and also the potential for the needs in the
future?
Yeah, so we raised 10 million in two years, especially it was like last year almost
everything because our first round was a bit in two stages.
So yeah, last year we confirmed to 10 million.
So now the fundraising topic is about how we can accelerate into different markets, into
different needs.
So, yeah, the fundraising strategy is something that is always also very evolving day by
day.
So, try to also measure what is the capacity in terms of capital you need to deploy to
make your mission a success.
Because if you raise too much, you're then stuck in certain way if you don't achieve your
purpose.
We see a lot of startups, I will not name them, that have raised first round 50, 100
million and then are stuck in this position.
are so high that uh...
how do you evaluate a company that is like already evaluated at 500 million, half a
billion dollars when there is no revenue?
And it's not foundation, it's applied AI.
we need to do stuff.
Foundation, we can understand.
Yeah, your model performs well, it does good benchmark.
Maybe you have a few rounds before it starts to be mainstream.
But on applied, you need to have like results on revenue, signing customers, et cetera.
So right now our focus is...
commercial deployment.
We are 200 % on sales.
It's going well for us.
We're welcoming a lot of new customers this summer.
It's very seasonal because it's a long sales cycle, but we are quite happy.
yeah, for the moment, the big focus is on sales, but we are expecting probably to
fundraise again by the end of the year or mid next year maximum.
So.
not out of need, but out of creating this legitimacy.
And I will be transparent with that.
is, you know, in fundraising in VC world, you have at some points a kind of tacit election
that's happening where you raise Series A or Series B and the investors and the community
and the customers are like, okay, these guys are gonna be the default for a few years now.
And this is what we try to be in the next coming round.
Yes, see.
How did you experience the ecosystem for fundraising in what you're doing?
Because if I'm correct, the first round you did in Paris, It's a Galleon, I think it's
Persian.
we started in Paris, in everyone there's been like a majority of European actors.
Actually, right now everyone is European based investors.
So yeah, we started with Galleon.
That was our first crazy follower.
And last year what happened is also Cast Capital from Germany joined and...
the same vision with us and Galleon continued to follow us and we also made a very
interesting fund, a follower named Axelio and we have also K-Fund in Spain, which is part
of our captable.
So yeah, it's a very pan-European captable that we have right now, but our vision is to be
global, so we are completely open-minded to raise elsewhere.
Right now, I'm a lot in New York.
for customers but also for potential investors.
So it's not a thing.
But the feedback is, if you want my honest feedback on pre-seed seed stage, and this is
something that is, I would say, yeah, everyone says it but no one really confirms it is
that there is only a few actors that are pre-seed, that are really, really invested in
pre-seed.
I would say right now, yeah, only two, maybe four VCs in Paris.
our pure press player that will play the game.
though there are a lot of funds that claim to be pre-seed, right?
But they are still expecting you to show revenue and early market fit and that's what
you're saying, right?
Yeah, it's a very formal world.
So sometimes these funds that are expecting more metrics will just follow the crazy guy.
But yeah, the very precede logic, which is like purely founders led, vision led, I would
say there is less than five in Paris that are doing it properly.
I will not name people, of course, Gaglion is part of them.
Kima is part of them.
But this philosophy is still...
Not, would say, very French and very European.
We have seen that more in Silicon Valley.
It's more like, okay, I like what you do, you look fun, you look like a crazy founder.
Here is one million dollar and figure it out.
This is not something that we are used to do in Europe.
And in seed, it's a bit kind of the same because you're still on pre-market fits, kind of.
You start to have some spark and something is starting, but...
the conviction is still what's driving the investment.
yeah, having funds like Cusp helping us on this and having the same conviction with us,
putting kind of a good amount because oh it's still a good fundraising that we did, is
very, yeah, I would say it's very motivating because you're like, okay, people are also
thinking about it and people that does like good experience in the market are thinking
about it.
So yeah, but you know, as long as VCs follow the power law, this is what I say to people
fundraising.
Find VCs that are like extremely following the power law.
They need one investment to return the fund.
The rest is more or less just running.
And if they follow this logic, they will be able to also invest in kind of gut feeling and
conviction.
Yeah.
Interesting point of view.
Yeah.
point.
Okay.
Very impressive story, Alexander.
Exactly.
Yeah.
Is there maybe if people I wanna hear can follow up this story, how can they reach you?
How can they stay up to date?
The website is twenty five one so two five zero one dot AI.
But uh how can people find you, reach you, yeah, stay up to date and
My LinkedIn is the best way to say hi to me and I reply to every message so that's
something that I continue to do every day and I'm doing enterprise so LinkedIn is my first
social network now.
Yeah.
We'll put it also on the show notes for people that wanna follow you.
Yes, thanks a lot.
Is there anything else you wanna you wanna say?
Any any less less words before we call it?
Now I would say, yeah, people continue to grind.
It's been like 30 years that we didn't see something like that happening.
So AI revolution is the moment to create, it's the moment to grind, it's the moment to do
some code.
I think I'm very optimistic person.
I'm more like pro.
innovation than afraid of it, it's gonna change a lot of things.
Let's not be realistic.
But at the same point, we are potentially seeing what is one of the first step to, I don't
know, a new kind of economy, a new kind of humanity.
Yeah, our life is gonna change, and I think in quite interesting way.
Yeah.
Cool.
Thanks a lot.
And yeah, thanks for sitting with us and chatting with us.
Pleasure having you.
you