Opening Keynotes - AIE Paris 2025 (Day 1)

Channel: aiDotEngineer

Published at: 2025-09-24

YouTube video id: d6dp_dwgpYQ

Source: https://www.youtube.com/watch?v=d6dp_dwgpYQ

[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
Heat. Heat.
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
[Music]
Heat. Heat. Heat.
[Music]
[Applause]
[Music]
[Music]
[Applause]
[Music]
[Applause]
[Music]
Ladies and gentlemen, please join me in
welcoming to the stage your MC for the
AI engineer, Paris, developer
experience. experience engineer Ralph
Jabri.
[Music]
[Applause]
[Music]
Yes. Hello, AI engineer Paris. So, I'm
Raou, your MC for the next two days, and
I'm super happy to be here with you
today. And I would like to start by
saying thank you. Whether you're tuning
in online or here with us today in
Paris, we couldn't do this without your
support. So, let's hear it from you
guys. Woohoo.
We have an amazing event and lineup for
you guys in the next two days. You're
here for a treat. Believe me, I've seen
some of the talks and they're they're
just fantastic. You're going to learn
and hear from European and international
experts and leaders in the AI space. And
we couldn't be better be we couldn't
feel better about having the event uh in
Paris and in Station F.
So we see so many European labs and
startups like Mistral AI, Black Forest
Labs and Qout who are redefining the
state-of-the-art in open models in
generative media and beyond and Paris is
at the center of all this innovation and
this is why we are here today.
This event is building on the successes
of AI engineer worlds fair and summits
in San Francisco and New York that
gathered thousands of engineers from
around the world and who shared their
best practices and experience building
with AI. And at the World's Fair in San
Francisco this year, we had over 150
sessions across 18 tracks discussing
topics like software engineer agents,
MCP, generative media, robotics, graph,
and AI infrastructure and more. And
we're bringing all that energy here to
Paris today.
Speaking of graph and AI infrastructure,
shout out to our platinum sponsors,
Neo4j and Docker. A huge thank you also
to our gold sponsors, Sentry, Arise AI,
Deep Mind, and Alolia. And of course to
all our sponsors and partners. Without
you, this event wouldn't even be
possible.
So many of these folks actually brought
their teams of engineers and PMs and
execs and even founders to meet you
guys. So, please take the time to go and
check out the expo, ma make new
connections, and also maybe you're going
to land your next partnership, job, or
even customer. Speaking of upstairs, we
also have registrations upstairs. I see
many badges here, so which is a good
which is good news. But make sure to get
your ticket for tonight's uh for
tonight's welcome party. That's going to
lend you a free drink. And um if you see
badges, you're going to have we have
many colors. You can we have blue,
black, orange badges if you are an
attendee, but we will also have a great
team to support you throughout this
event wearing green and purple badges.
So if you have any questions, please
feel free to ask them.
Um, so what to expect from today? Well,
we have amazing guests and a talk by
Mistral AI. But please make sure to
stick around for tonight's party. And we
also uh we're gonna be very happy to
have you there and you're gonna meet
great engineers, founders, and and
leaders in the AI space. But before
that, let's hear from our first guest.
Our next guest is a person that I'm
lucky enough to call a friend who and
who I truly admire. He's been in the
developer community space for years.
He's behind Reactathon,
behind gems.com, and he also co-founded
AI engineer with Swix. Ladies and
gentlemen, please join me in welcoming
to the stage co-founder of AI Engineer
Benjamin Dumpy.
[Applause]
[Music]
All right.
When Swix and I founded AI Engineer more
than two years ago, we both had a strong
desire to extend the brand beyond the
events that we produce ourselves.
Inspired by great conference series like
JSCON, we envisioned a future where
community members were to organize
events around the world and we would
show up as attendees.
And while this is a nice theory, it's
not so easy in practice, especially if
you have a high bar for the quality of
content, brand, and experience because
it takes the right partner who has the
vision, the drive, the grit, the
professionalism, and the risk tolerance
to organize a successful, high quality,
and high signal conference.
So, we had to be very selective with who
who we partnered with. especially for
our very first community event. And
that's why we're happy that we selected
our friends at COB. Let's hear it for
COB.
So tonight, thanks to our wonderful
partners at KOYB and on behalf of the
entire organizing team, staff and
volunteers, I have the distinct honor to
welcome you not as a host, but as a
fellow attendee, as a fellow member of a
community that has grown beyond its
founders, beyond San Francisco, into
something bigger. A community that has
taken its first step to becoming a
global movement.
Ladies and gentlemen, welcome to AI
Engineer Paris.
>> Ladies and gentlemen, please join me in
welcoming to the stage co-founder and
CEO at Coy, Yan Leger.
[Applause]
[Music]
[Applause]
[Music]
[Applause]
Hello
and I'm thrilled to take my turn to
welcome you today at the first AI
engineer conference al organized outside
of the US
um here in Paris in the beautiful Paris
which is u my home city um the home city
of our company KOB um and a beautiful
place that uh if you're visiting I hope
you'll enjoy this conference is
particularly unique to us at COB Um we
it's the first time we're organizing an
event of this scale and some of our team
members spend the last three months
making this event happen for you. So I
hope you'll enjoy it.
Over the next 24 hours engineers, CTO,
engineering managers and a variety of
attendees coming from all over the world
will join us. you all are as far as I
know 70% coming outside of France. So we
are pretty excited to have you all here
tonight and to have this amazing crowd
to join us for this first event outside
of the US.
Today and tomorrow you will see an
incredible lineup of over 30 speakers
talking about how they build
foundational models, deploy MCP at
scale,
their experience by co coding and more.
We put a special attent attention in
inviting on stage several leaders of AI
startups founded and operating in Europe
including Leelio who will be speaking
tonight from Miswell. He's head of
engineering at Miswell. I hope you'll
enjoy his talk. I'm the co-founder and
CEO of Cray um where we provide high
performance serless infrastructure for
AI applications. We're not an event
company.
Um, so we're accelerating application
deployment with a seamless way to deploy
agents and inference services and models
across CPUs, GPUs, and accelerator. But
we love fostering AI communities, and we
love connecting both sides of the
Atlantic.
Um, tonight I want to thank our team,
uh, especially Aliser, Jen, and Edwan
who have been working relentlessly
during the last three months to make
this event a success. Um, the entire COB
team who you will probably meet during
the the course of this event and uh,
tonight and tomorrow. Um and of course
Rahul our MC who will be animating for
this uh 24 coming hours and the AI
engineering team um Swixs and Ben who
trusted us in making this event happen.
Um now uh I want to welcome Mwan on
stage. Uh so Mwan is head of startups at
station F. Station F has been our own
for the last four years and I want to
give him an opportunity to say award and
welcome you with me tonight.
[Applause]
Thanks. This one is working. Good
evening. Uh super happy to see all of
you. Super cool to see such a big crowd
here. Um, so my name is Marwan and the
head of startup programs and
partnerships as station and you know I'm
part of the funding staff of of this
beautiful place and you know when I see
you it's exactly what we are here for to
gather talented people and to create
connections between uh maybe some
funders here in the room. So where are
you exactly? You are at station F. So
station for those of you who are here
for the first time or people who are
watching online it's a massive startup
place. We launched eight years ago. We
have 1,000 startups here in the same
spot. They are participating to at least
30 programs. Some are in cyber security,
some are in consum computing, some are
in AI obviously. And a big congrats to
COB because Coy is not just one startup
among the 10,000 startups we have. It's
one of the top ones. You know, every
year we unveil a list of the best 40
companies of Station F. The best 40
across the 10,00 is part of them. So big
congrats and super happy that they have
this ability to gather such a big crowd.
And to conclude, you know, station F is
a big AI place, one of the biggest in
Europe. 70% of our startups are in AI or
they have some AI components in their
key offer. You know, the biggest alumni
we have is hugging face. Hugging face.
They were born here. So they were here
in 2017 and 18. So it's one of the
biggest alumni we were lucky to have.
And well, I just have one next step or
two next steps. Well, enjoy enjoy your
stay session. Enjoy the party just
after. Enjoy the sessions, of course.
And of course, if you one of you here in
the room thinks about building an AI
company, well, stay tuned. Working to
announce some big new offers at SF in
the coming weeks. Thank you.
>> Thank you, Mar.
Now that Marwan welcomed you and gave
you some insight about this UNX place,
I'm excited to give the floor to uh
someone quite important in this journey,
Swix uh the other co-ounder of AI
engineers. So please give please give a
round of applause to Swix.
Thank you.
[Applause]
[Music]
[Applause]
[Music]
[Applause]
[Music]
That was very dramatic. Hi everyone.
Thank you Yan for the uh kind intro and
um it's so nice to finally be in Station
F. uh we've heard so much about it from
the US and uh to see this live in person
is um it's really an update for me on
like the sort of the Paris AI scene and
uh the the tech scene in Europe in
general. So very excited to be here. Um
I was given basically a general task of
to just just give some kind of like
state of affairs. I've we've been
thinking a lot about agents this year
for AI engineer and we started the year
uh you know basically focusing on the
quoteunquote year of agents. So I just
wanted to give an update and uh give an
like a broad overview of where I'm
thinking that uh AI engineering is going
this year. Okay. So uh this clicker is
not working.
Okay. All right. AGI is not here yet. So
we have to uh solve that. Um so so I
gave this talk um at the first AI
engineer summit of this year, right? And
so basically we're just updating this
talk um on like what has happened since
then, right? So this was in this in the
early uh early part of the year and uh a
lot of back then a lot of people were
saying 2025 is the year of agents and uh
you know if you say it by Satia if you
say it by from Roman from openi if you
say from Greg Brockman from openi and
Sam Alman also from open eye maybe you
say it enough you you'll sort of it will
actually finally come true. uh I think
more broadly like a lot of people have
to discuss like what is the definition
of agents and I think this is something
that you you'll come up you come across
in your discussions over the next uh few
days and um I think like the the the the
evolving definition has has is is mostly
reflected here in Simon's post uh which
is something like last week which is
agent equals the an LLM with tools you
put in a a loop uh and you give it a
goal uh that is directed um I actually
did a bit more work uh on like all the
definitions of of agents. Um so if you
want to go over the agent engineering
talk, we have intent, memory, planning,
authority, control flow and tools. Um
and so I I definitely if you are you you
know in the agent discussion, I
definitely want to push you towards
discussing the harder parts and not be
too simplistic about forgetting that uh
really good memory, really good planning
and really good trust and authority
actually helps you build uh better
agents that people will not hate.
Um, I think the other thing that's
really driving us uh this year, I I can
see that uh this is this is out of date
slide. Um, is that we're starting to see
like one of the most epic infrastructure
buildouts of all time. Um, I think that
this is uh one of those things where um
the numbers kind of start to haze and
like not make sense. Um, when when
Stargate was announced at the start of
this year, most people were kind of
doubtful that there was actually money
to back into it. Uh, but now we're
actually able to see that. um we
actually have hundreds of billions of
dollars available to invest and I think
this is only the start. Um I think and I
think that is that bolds very very well
for the rest of us downstream uh of this
infer build. Um I'm going to refresh the
slides briefly just because I know that
this is out of date. Let's see. Okay.
Yeah, it looks like it looks like I can
play this. Um having done this enough I
know exactly what went wrong. Okay. So u
I think the other thing that's uh that
when I whenever I talk about this kind
of infra buildout with people a lot of
people who are skeptics are saying like
oh like you know that the AI usage is
not really there. Um and I think that's
not really true. Um Chad GBT is going to
hit 1 billion users in two months. Um
and like it is it is pretty much a
guaranteed that it will take over uh the
planet in in so far as like Google has
taken over the planet. Um and I think
these are if you look at the projections
and the the sort of historical estimates
of all the compute that's being used and
and probably will be used. Um it it's
pretty evident that this is actually
just the beginning of a of a giant infra
build up. So really I think the the
point I wanted to make is the same point
that Andre Karpathy made uh at at YC
startup school uh around the middle of
this year which is that it's actually
not 2025 is the year of agents. is
actually like the next 10 years we're
going to be building uh a lot of agents
and there's the most epic historical
tailwind you've ever seen in your tech
career. Um which is fantastic like you
know where the future is going to go in
roughly and you can sort of point your
your career your startups your your
businesses in that direction. Um the
second thing we're seeing this year is
increasingly agentic models. Um this is
a very hard slide to put together, but
basically this is every single um major
uh model launch that um you know that
sort of hit my radar. Um you by the way
uh you don't have to take a photo of
this. You can see it on my website um
and I can tweet it out later as well. Um
I bolded the ones that I think you
should will probably stand the test of
time. There's uh shout out to Mistral
Mattress Straw on there. Uh but also the
the Chinese models Quinn 3 coder GM 4.5
as well as the Frontier Labs. Um I I
think the the way that I can most
concisely describe what agentic models
look like is they are thinking with
tools. Uh we wrote this piece with uh it
was part of the sort of GT5 developer
preview and we wrote this piece about
how if you sort of zoom in over here ah
god I can't I can't zoom in on this
screen. uh you know like uh it oh god
sorry
um the the kind of the kind of thinking
that you can get with tools um is is
increasingly of this format where uh you
you can basically start to instrument
your thinking and you start to and the
the the thinking process starts to use
tool calls um and I think you see more
and more of this over time I'm sorry the
contrast is not very high but uh I mean
that's just more incentive to go read
latent space uh and see the actual blog
post itself. Um the other major uh side
effect or obvious observable impact is
you start to see increasing autonomy,
right? Um I think if you see like for
example like the replet uh agent launch
uh they'll talk about like now they have
like 200 minutes of of autonomy and I
think like we'll start to see hours and
days of autonomy and I think like that
that is uh the models are specifically
being post trained to do that which is
uh really fascinating for people who are
building agents and I think the last
part that is really driving this is like
now we're realizing how to do RL and LMS
uh starts to the point where we're
starting to allocate uh the same amount
of uh compute on post- training as as we
did on pre-training and this never used
to be the case prior to this year but
here's XAI probably saying that they are
doing it for Gro 4 they also said
recently that they uh did some more
similar stuff for Grock 4 fast and I
think like that's that's like obviously
where these things are going okay so um
that's the agent that's the model side
um on the engineering side which is what
we're all here to do AI engineering we
don't really uh you know most of us will
not really control the models but what
we can do is what we what we build
around the models um and that's agent
products agent protocol Agent Labs. So,
um here are also like some of the
notable agents. Uh there's absolutely no
way I can ever uh uh list all the good
agents. So, if if if I missed your
agent, um sorry, just uh find me outside
and I'll add you to the list. Um uh but
but again, I've bolded a bolded a number
of the the the important agents and I
think it's important to see how far
we've come along in the last nine
months. Uh we've only, you know, we're
here in September of 2025 and things
like opening ID research seems really
old now. Codeex CLI launched in April
but only recently became popular. Um I I
think it's it's a really interesting um
look back at like how fast the agent
field evolves. Um we in AI engineering
like really care about agent protocols.
Uh we made a particular bet on MCP in
the last AI engineer summit um including
a workshop that was very very well
viewed. Um and I think like u that that
is obviously has taken over. I I kind of
don't really need to to explain more but
I think like what else is after the MCP
uh or what else is being built on top of
MCP or with the MCP spec I think is is
still an open question. Uh so Google has
introduced A2A but I also actually would
want to shout out Zed for ACP uh which
is agent client protocol which actually
um sort of starts to be the interop
layer for all the terminal agents like
cloud code and uh open codeex. Uh I
think that's that's very interesting for
uh all the tooling that is starting to
to emerge, right? Like we're no longer
connecting to these things on the model
layer. Like we're not just changing the
model string, we're actually changing
the entire agent and swapping out the
user interfaces uh for different agents
as well. So u I highly recommend
checking out Zed's ACP if you uh haven't
looked at it. Um the other thing that I
think is an emerging consensus is
actually uh something that came out of u
Mistral's work on the agents API. Uh so
Mistra actually kind of solidified this
trend where every company every model
lab has an agent API that has a bundle
of tools that is this the standard
library of of what an agent's um sort of
platform should have. So you should have
a code exe code execution sandbox you
should have web search you should have
document library maybe you have imageen
uh I don't know I feel like the black
forest labs people will have a strong
opinion on that one uh and and you you
probably have MCP sort of connecting to
the the universe. So this is my update
of um onjarpathy's 2023 thesis on the LM
OS. We've really pretty much mapped out
what we definitely definitely need for
search for code execution for document
library uh and for multimodal input and
multimodal output. What is still missing
also we have MCP what is still missing
is good memory and really good
orchestration. There's some emerging
candidates but nobody has really won
here. I think there's a lot of
opportunity to build in the LM OS. Okay.
So um I then I think the last part is
like I guess more pertinent to me and
like the stuff that I'm focused on. I
think code agent labs this year have
really exploded. Uh I recently blogs
about cognition and why I think uh
something like the the the sort of pool
code is code AGI will be achieved in 20%
of the time full AGI but capture 80% of
the value. I think like um not all
agents are created equal is basically
what I'm saying there. And this is why
uh over on over in the US for our code
summit, we're actually announcing our
first ever uh uh you know AI engineer
summit entirely focus on coding agents
and and so we're heading back to New
York in November for that one. So check
that out if you're if you're interested
in specifically coding agents. Uh I
wanted to close with some open debates
um where I think like this time if we
meet again next year uh we will probably
have some solutions. I actually just
want to encourage you guys to come up
with answers. Uh the first one is do we
need evals? Um, I accidentally, um, I I
got really pissed off at like some
people sort of overhyping EVELs. And I
think here's the, uh, built here's
Boris, the the, uh, the sort of chief
architect of cloud code, saying that we
tried really hard to build evals, but
yeah, man, in the end, it's all vibes.
Um, so the the single most successful
coding agents uh, so far this year, 500
million in revenue, no evals, just
vibes. And I think like, you know, like
that's cool. I think a lot of people who
are sort of have your professional
identity tied in with with evals I I
think like that that makes sense for a
small domain that you need to hill climb
on. But if you have a very general
field, it's actually not obvious that
you should do evals first or actually
have evals as a blocking restriction on
your product development. Um and I think
that's a bit controversial. That's why I
just put it as an open question because
I do not have the answers. I just I just
ask them. I just ask the questions. Uh
the second one is how to do context
engineering very well. uh we have
cognition and anthropic uh sides side by
side saying build a build multi- aents
don't build multi- aents but really uh I
think the multi- aent debate is also
part and parcel of the context
engineering debate um AI engineer worlds
fair in in June talked about this uh so
I highly recommend checking out that
article as well I think we'll develop
this more over time there's actually no
current um standard way of viewing
context engineering apart from if a
couple blog posts and I think probably
there's a couple startups in here that
are uh that are worth building that will
emerge over
Um the third one is actually more less
consensus but um it's is definitely I'm
I'm hearing it in the valley a lot and
I'm sharing with you guys is fast
agents. Uh what do I mean by fast
agents? Cereba's code is one example but
there's other other examples like
Samanova has come up with some stuff.
Grog has come out some stuff. Um but
here's every model provider providing
your you know your normal tokens tokens
per second at like let's say 100 200
tokens per second and there's Cerebra's
code all the way over the other side at
2,000 tokens per second. Um and I think
every 10x you get in speed, you get you
unlock different kinds of behavior in
just both users and your sort of product
uh possibilities. So um people are
definitely exploring this. Um so this is
this last part is my catch all. I think
uh over the next year we're basically
going to see a lot more development in
all these domains. Um email clients,
browser uh agents, voice calling, vibe
coding, low code and education agents. I
think the last part education is
something that u my co-founder Ben is
super excited about. We should be doing
we should be announcing the AI engineer
education summit pretty soon. So that's
all I wanted to that you know that's as
concise as I can make it for the state
of agents. If you want to come talk to
me about any of these I can obviously
talk your ear off about it but I just
wanted to set the context for all you
guys and thank you so much. I'm looking
forward to chatting with you. Thank you.
[Applause]
Ladies and gentlemen, please join me in
welcoming to the stage head of
engineering at Mistral I Leo Lavo.
[Music]
[Applause]
[Music]
[Applause]
[Applause]
Hi everyone. Very happy to be here. Uh
thanks to the Coy team and to the
engineer team for welcoming me. Um it's
an honor to be in Paris where Mitch was
born um a little more than two years
ago. Just need another slide.
I'll just use my um arrows. It's going
to be fine. Um so what is Mistrol? Um
Mistrol has been founded as I said a
little more than two years ago by uh
scientists rooted in um I mean
engineering and science. uh the initial
authors of Gemini, Llama, Chinchilla
papers um and with the main goal and
objective uh to bring AI to the
enterprise world um and also obviously
to advance open source and open models
and this is why um after a few months uh
of work we released Mis 7B which was the
state-of-the-art small model um that
actually triggered um a spur of
experimentations fine-tuning and new uh
usages uh among the community and the
open source community. Um after a few
months we also released the first
mixture of expert model uh mixed roll 8*
7B um which was actually a you know
great model that also was very welcomed
by the community. Um and it was fun in
San Francisco at GTC um in spring 24 uh
when some people from one of the biggest
uh old company in the world came to me
and said we are building such amazing
usage on top of mutual 8* 7B and we have
fine tuned it and we're actually putting
this in production and getting amazing
returns but we had never heard about
them at all. They never reached out um
and we never built something with them.
That's where we really saw that the
mission of providing open source models
to actually uh bring the AI to the
masses uh was also compatible with
pushing enterprise beyond and building
some strong um focus. So what we build
at mistrol um sorry about this slide
it's very corporate I did not make it um
but basically we do have foundational
model this is what the science team is
building. We're also continuing
pre-training with specific customers to
bring more rare uh languages or specific
capabilities um to those customers. We
do provide um misi studio which is a
wrapper on top of that providing any
kind of tools and API you need to build
complex AI applications. Obviously this
starts with completion APIs. Uh but then
you have conversation and stateful APIs
and more than that. Um obviously we also
provide some products. So you have heard
about Lasha which is our assistant uh
Mistral code which is our agentic uh
code completion um tool basically um and
we do build custom solutions for
specific customers with a strong um team
of solution and applied engineers. So we
started with seven uh now it's more than
100 engineers 100 researchers and 100
applied engineers actually implementing
those solutions. But you have seen um
this uh numbers and you know that
actually PC's fail in production and
that enterprises actually struggle to
find some value in AI. And so I'm not
going to dive into uh uh the precise
numbers and everything but I'm going to
talk about uh how we do uh try to
overcome some of those obstacles and how
we try to make AI a reality for our
customers. So what do they face? they
face issue with data uh with the
observability layer um with the skills
gap internally and externally and
obviously sometimes it's our fault
sometimes the models are not great
so about data um data is massive and
it's unstructured uh enterprises have
accumulated data for the I mean for the
few years for a few decades for some of
them and they've put this data pretty
much everywhere um without u an AI
policy and that mean that makes sense
because AI did not exist
But without the strong AI governance,
this means that you have silos of data
everywhere in different providers um
that is not you know attached to the
proper metadata that is not meant for an
AI system to actually investigate and
dig in. And this means that those poor
management I mean data management
practices um actually result to really
strong challenges in bringing this data
together and bringing value. The fact
that we have so many providers is also a
challenge because each provider now
knows that this data is worth gold. And
so each provider is now inventing their
AI assistant and AI agent. And I mean if
you're using SAS you now know that you
have access to 15 different kind of SAS
uh assistants that basically all wrap up
you know OpenAI API cloud API or
sometimes Mishold's API and basically do
the same thing but over the data that
they propose but none of them work on
interoperability. Obviously, we do have
MCP servers and that's great, but that's
usually not enough. And that usually
means that the providers are in charge
of providing you with search for
instance, but search is not always good.
And if you want to actually interact
with all this data, you need this kind
of layer, unification layer to actually
make sense of things that can be spread
out on Google Drive, on Microsoft Teams
or on Slack. Obviously all those
providers have a different vision of
access control of our back and it's
really hard to have like a unified model
of who has access to what. So this is
why we're um and obviously you're
thinking about rag right if you have
access to multiple MCP servers and
multiple knowledge bases you can use rag
everywhere and everything is going to be
sold. Well, that's not exactly true
because you know if you work at super uh
in a quoration making the best socks in
the world and you really want to know
what are the socks sales of last year if
you just provide this query well you
might find you know great answers in
annual report PDF which is from 2013 and
then maybe your rag is going to provide
you with 2014 PDF report but if you do
not have the proper context and this
context is very very you know enterprise
and customer specific
then the model might not even be able to
actually understand what it is because
you will have five chunks with annual
report PDF that are in different folders
but this is not necessarily provided to
the model. So building a strong context
engine that is able to build some
uristics and that is able to actually
understand the link between entities
that might be residing in different data
silos among different providers to
actually make sense of this data that is
spread out everywhere and map it to some
kind of ontology understanding the
entities um and basically providing an
AI first data organization is the key to
building um strong databased systems.
So another issue obviously is the black
box phenomenon. Um so everybody wants to
observe and everybody wants to explain a
technology that is inherently
nondeterministic. Obviously this is a
challenge. So instead of that you want
to actually provide all the tools that
you need to achieve the goals of your
customers which are trust. How can I be
sure that the model is right? Well, you
cannot be sure, but what you can be sure
of is that I will provide you sources
that you can check or I will classify
that based on the input, the output that
you got actually makes sense and that
the model did not completely diverge.
And so this is providing the set of
tools um that will get you the kind of
trust that you need and safety. safety
being very relative because obviously
the safety if you're building something
to moderate Reddit commands is very
different than if you're trying to power
uh kids chatbot. So this is all context
dependent but this is also extremely
important especially in a regulatory
environment that is moving very fast
where you also have to prove to many
different actors that you're doing
things right. So once you actually
collect all this data, you want to watch
people do what they do. And in
enterprise um context, it's really
important to use this to understand how
people are actually doing their job. How
are they, you know, processing their
days? What kind of workflows are they
putting in place? And observing is a key
to that because AI is not just um what
people have been using which is
basically JGPT like assistants um which
are providing a lot of individual value
but do not convert into profits and
losses optimizations for companies. So
you actually need to understand how
people are using it. If you give them a
task, how do they do it and what are the
different queries that they actually
perform and how do you automate that?
And for this you need to actually
observe all these data and then you know
extract some insights out of them to
help the companies change and transform.
And then once you've collected all this
data obviously you can improve your own
models. What does it mean to improve
your models? Well that's very easy. You
have seen how the you know chatbot
assistants have improved over the years
uh the past few years. This is because
they collected a massive amount of data
and were able to optimize the models.
Well, you need to do the same thing for
enterprise and large customers
situations. You are collecting this
data. You need to make something out of
it. You can make models that are even I
mean sometimes better, sometimes smaller
that are optimized to work with certain
set of tools, certain sets of
connectors. Um, all of this can be done
if you're collecting this data and
really try to use it the same way that
the big players used it to improve their
chat bots.
So how do we get to maturity after all
that? Well, obviously there is a model
performance part. Um I will take my
responsibility here. Um you need very
strongly aligned models because you're
getting into more and more complex
processes with like longer system
prompts and you need uh very I mean you
need the model to strictly adhere to it.
Um there is a lot of things being I mean
going on with structured outputs and
tool calls. You need them to be very
reliable. You need the model to
understand the intents. Um and obviously
speed. um switch was talking about you
know cerebras and and fast agents. I
think that's something that is obviously
I mean once you've tasted to a fast
model even though you might say it's not
the most important thing it's really
hard to go back and so optimizing the
models to make them smaller or to
optimize inference is also a key part of
what we're doing. Um the expertise I
mentioned it briefly um well people do
not really know what AI is and so the
lack of expertise really grows us grows
on us because it's hard to explain to
the customer what is AI and what is not
AI what to expect what not to expect and
we need to train the workers at the
customers we need to train our own
workers we need to train everyone to
understand how this models behave how we
can improve them um how we can make
everything better and how you know
system prompting works how why the model
is behaving quite differently in a
certain way. evaluating this and not
just vibe checking or you know vibing
your system prompts is something that
you really need to educate uh the
customers with otherwise they will just
send you a problem and say well it works
well it works but does it really work on
all cases um and so this is really an
education problem uh that we're facing
and finally all of this is just bringing
change and huge change to corporations
and organizations so AI can be just a
tool but then it's you know a finite
project but AI is really more of a
capacity and once you've integrated
this, you can actually change the way
you're actually doing business. It can
be a transformation enabler. I did not
mention agents a lot. Um, and this is
because I usually don't like talking
about things I don't know and I don't
know what is an agent. But an agent can
be, you know, just doing a micro action
and wrapping an MCP call. An agent might
be performing a very complex set of
tasks. An agent can be pretty much
anything. what you want to be. But what
matters is the actual business case or
you know workflow that you're trying to
solve. What problem are you solving if
agents are a part of a of the puzzle or
or the bigger I mean the the big puzzle
whatever what matters is that you're
trying to solve a problem from end to
end and agents can be a part of that or
not whatever that's uh important.
So how does uh AI transform
organization? Well, there is growth. So,
how you basically increase profit? Uh,
you create new use cases, new
applications. You're doing new stuff,
creating new products, processing new
data that you did not used to process
before. And this is really bringing new
revenue and creating new opportunities.
On the other hand, efficiency is also an
extremely important part of what
customers are asking. How do I make my I
mean, how do I reduce my cost? basically
how do I make my organization and my
workers more efficient by streamlining
their processes by saving them time by
you know crunch crushing all this data
and providing ready to go insights um
and those are really two sides um of the
same coin um that are crucial to
understand how we can use cases uh to
solve enterprise situations.
So to sum it up, um you need to build AI
your way. And this is really uh a matter
of training uh and making sure that you
actually set up the the flywheel of AI,
right? You you will you will build, you
will deploy, you will observe, then you
will improve and you will start over
again. Uh making sure that you can
actually customize things and that
you're not just thinking AI as an
off-the-shelf solution and you just need
to, you know, enter your credentials and
then boom, then it works. um you really
need to tailor um the AI to your use
cases. obviously leveraging the
community. I mentioned open source at
first but we learned so much by actually
publishing our models uh and the
community has done so much by
fine-tuning providing data sets
providing so many examples on why we
were failing and you know we uh were
great for instance on role playinging
well I discovered you know this amazing
community of role players that were
actually providing with uh extremely
long prompts and super weird scenarios
sometimes that also help to align the
model because it now can understand
instructions that are very complex. So,
open source uh has been the core of what
we've done and we're also now pushing uh
enterprise solutions. And finally,
breaking barriers uh is is essential.
Making sure that you can actually uh you
know, get the data where it is. And some
people are relying exclusively on MCP
servers and say, okay, well, you know,
you'll have MCP servers for everything.
And then you're dependent on many
multiple providers doing their job of
implementing MCP servers, maintaining
them and providing a good quality
documentation, good qual good quality
open API. ML um sometimes it doesn't
work and this is why the community is
also pushing uh you know to get you know
custom browsers or browser add-ons and
stuff like this because you don't want
to reauthenticate to 15 different SAS
providers just to get the data. So
breaking the silos um is also extremely
important uh and making sure that you
you avoid vendor lock in so that you can
actually build this you know context
engine that brings everything in a
single representation um that helps the
model perform well. So thank you very
much for your attention and uh happy to
talk after after the talk.
>> Thank you Leo. What an insightful
presentation. I'd like to invite you to
join me for a couple of questions.
>> All right,
let me pull the questions on my phone.
Okay, so what I found particularly
fascinating and interesting is you
mentioning that enterprises actually
have these data silos where not all the
teams has access to all the data that is
generated by by by those teams, right?
Um so how does Miscell actually help
them in structuring that data?
>> That's a good question because uh I
mentioned that the data governance was
not there. Um and you might you might
think okay we need like all those humans
to actually set up a data policy and set
up all those processes and stuff like
this but it's actually a bit of a loop
right we do have AI to actually parse
the data silos that we have and classify
them and understand the ontology and put
things in the right order. So using AI
to get data to structure data to create
more AI projects.
>> Exactly. And use the humans to guide the
AI but not requiring the humans to now
say okay you're going to tidy this you
know 15 years old drive. Good luck to
you.
>> All right. Okay. Um I have a second
question for you. So uh Mistrol is known
for its open models and you mentioned in
your presentation uh the 7B model that
made you so famous because it was
state-of-the-art and uh open models when
uh when it was released. So I I was
wondering how you strike a balance
between building for enterprise
customers but also have this o focus.
>> Yeah, that's a very good question. At
first we only did open models but then
obviously we saw a lot of companies and
a lot of different actors capturing a
lot of value. um that we were actually
spending time and money uh to to train
models. So we we're still convinced that
pre-trained models are going to become a
commodity and are going to be mainly
open source. Uh but then the post-
training phase where you can actually
tweak a model or you know the continuous
pre-training where you just add some
more language or some more domain
specific knowledge is somewhere um that
requires a lot of science, a lot of
engineering work um and a lot of compute
power that many customers do not have.
And this is where we find the enterprise
sweet spot is by getting the base model
making sure that the community can make
something out of it but also trying to
uh make it more specialized in some
ways. Uh so continuous pre-training but
also in post training because um
typically distilling like bigger models
into smaller ones you can do that in
many different ways. if you don't do it
efficiently in FP8 or things like this
um you might need the expertise that we
have um and then you need all the
expertise to actually deploy them and
link that to actual real world use cases
um and so I think we try to balance that
as much as possible by pushing
everything we can in the open domain but
also keeping some edge uh so that we can
actually you know um make money I guess
>> so you guys basically have the compute
power and the resources and the knowhow
to train these these models. So, might
as well just make everybody make it
available to enterprise customers as
well to solve their problems.
>> Yeah, absolutely. And we really try to
accompani those customers so that
they're not alone with a you know base
model and say like good luck uh here is
the repo and just you know press enter
uh really try to accompanize them
because uh knowing what kind of data to
put uh is the model actually converging
in the right direction is the data
diverse enough or problems that are not
exactly solved um and we're really
trying to help there.
>> Awesome. That's all for me, Lilio. Thank
you so much.
>> Let's give it up to Leio.
All right. Oops.
This is Can you hear me? All right. This
second time it happens to me. Thank you
so much.
>> Thank you.
All right. We before we wrap this day
up, how how does everybody feel?
>> Yeah. Nice.
Okay. So, before we wrap it up, we have
one more thing for you. We have a
welcome party upstairs. So, I hope I'll
get to see everyone of you tonight.
Yeah. Um, please don't forget about your
ticket. Uh, that will give you a free
drink. I think I think it's important,
especially at this time of the night.
So, um, I'll see you all there. All
right. Thank you.
Oops.
[Music]
[Applause]
[Music]
[Music]
Heat. Heat.
[Music]