- 16 hours ago
As governments and enterprises race to build AI, they need the ability to inspect, adapt, specialize, and operate AI for local languages, cultures, regulations, and strategic priorities. Open models are becoming essential infrastructure, enabling greater customization, transparency, auditability, and faster innovation through shared tools, data, and collaboration.
Category
🤖
TechTranscript
00:01Thank you everybody and welcome to our panel.
00:05I'd like to thank my fellow panelists and
00:09you all for coming to spend time with us today.
00:13We're going to talk a little bit about transparent open models,
00:17data, how this enables us to build the ecosystem that's going to bring
00:22solutions democratically, right, to everybody.
00:26And so I really want to talk about that and
00:29I want to ask a round of questions.
00:31Everyone will introduce themselves in that first round and
00:34then we'll kind of open our discussion up a little more free flowing.
00:39So I'm going to kick off because we've got a lot of great content and
00:42a very limited time window.
00:44So Michelle Marie, can you tell us a little bit about how Linagora
00:49leverages open platforms, open data, open models to build even more
00:55trustworthy and transparent models on top of?
00:57Okay.
00:59So firstly, I would like just to thanks NVIDIA and especially John and
01:04the rest of the team to invite us to this panel.
01:07So I'm a general manager of an open source provider called Linagora.
01:10And so with Pierre Karl, but in some few hours, we had this sexy, crazy idea in 2023 to train
01:24and build digital command and generative AI and try to train really and truly open source models,
01:31not only open wide model, but models with open source data set to train these models.
01:39So it was crazy in 2023.
01:42So, but now because this is a bet also for Chinese planners, players,
01:49and also for NVIDIA.
01:52And so, and quickly we face, I think we faced three challenges.
01:56The first is to get to build this open source data set, but I will let Pierre Karl because he's
02:03the
02:03specialist and the expert on this topic.
02:06And I will let him to explain this part of the job.
02:10The second part is to access to the compute.
02:13And especially with the support of NVIDIA and the wonderful team of CNRS and GenC.
02:20So we have access to this platform and the compute needed to train this model.
02:26And so I know that we will deal it quite later in our panel.
02:32And the third part, it's related to talent and expertise.
02:37And so we started to work closely with the NVIDIA team in 2024 when we tried to use
02:46this supercomputer based in France called GenZ by starting to train this first open source,
02:57truly open source model.
02:58And so now we are part of the NemoTron Co-Aliation.
03:03And we are very proud to be there and to continue to work with all of the NVIDIA ecosystem.
03:12So the goal of our community and the work we are doing, it's not to build giant models.
03:20So the bet we are doing, it's to build a smaller model, what we call small language model.
03:29Very dedicated for specific use cases, train and align with our values and with our language.
03:38And model that you can run in the agentic world, very close to application, even in air-gapped system.
03:48Because for example, I can take one example.
03:53For example, one of our customers is the Ministry of Interior, the French police in France.
03:58So we deploy this kind of model combining with right system, agentic system.
04:03And so we are able to treat and to handle specific use cases for critical and the most sensitive system
04:12in France.
04:12And just to conclude, I think that this strategy based on very small language model are one level or one
04:27part of the strategy that Europe and France can develop its own autonomy and sovereignty on the general AI field.
04:35And just with this beautiful story last week about 5.5 from Anthropic, we know that restricted access to technology
04:49or model are no longer science fiction.
04:53So it's our real life and we have to deal it.
04:55And I think that the bet we did with NVIDIA and all the ecosystem around this community called OpenLM France,
05:04it's cool.
05:06And I think a good way to develop this server in Europe.
05:10So, Gautier, H&Co is known for building agents that work like people, right?
05:18They click a mouse, they look at a screen, meaning that we can integrate them into workflows.
05:23These virtual AI teammates, we don't have to change the workflow.
05:28We don't have to cut ourselves out of the workflow to integrate them.
05:32Can you talk a little bit about how that work opens up opportunities for other partners and in the Nematron
05:39coalition?
05:41Yeah, for sure.
05:42So, basically, two years ago, we made that bet, as you said, of computer use.
05:47So, that generation of AI that can connect directly to screens, to mouse, to keyboard, basically.
05:55Do you know how much time humanity is spending clicking and scrolling and typing stuff without even thinking?
06:02It's around like 60% of the time.
06:05So, if you're a nurse, if you're a salesperson, if you're an engineer,
06:08roughly 60% of your time is lost by trying to interact with technology.
06:12So, we are here to keep the intent, the intent being like what you want to do,
06:16but then remove all these like clicks, copy, paste, back and forth, alt-tab,
06:21all these things that you are doing with your windows.
06:26And it was a niche at the beginning, two years ago, when we started to focus on it.
06:33And it's getting more and more critical now to deploy in big companies because we've seen a couple of waves.
06:38The first wave was LLMs, so, you know, ChatGPT, Lusha, like deploying them in companies.
06:44Usually, it's a lot of money spent, but not a lot of ROI measures.
06:49The second wave were agents, so you plug these models to systems, so like software,
06:55and you get them to execute things.
06:58But in most companies, you don't have APIs everywhere.
07:01You don't have ways to interact with the existing tools.
07:04And the more important they are, often the more, the older they are.
07:08So, that's third wave.
07:11It's called computer use.
07:12That's what we do.
07:13And I think the reason why we started to work with NVIDIA very closely,
07:19NVIDIA and Jensen himself is very close to startups.
07:23They follow what's happening.
07:25They were one of the first ones to realize that computer use was going to become critical to deploy AI,
07:31not just as assistant or saving time or writing an email, but really in the core of the operations of
07:37a company.
07:39And Jensen once told me, you know, computer use is really hard and you guys are doing it.
07:45And we are very happy to be, you know, you have these like independent benchmarks.
07:49One is called OS World Verified.
07:51So, you send your model and a lot of tasks and it gives you a score.
07:55And we have the best score.
07:57Like, we are both anthropic, open AI, the Chinese models.
08:01We are the best in the world in terms of model to do computer use.
08:06And that's the reason why I think we joined NemoTron Coalition,
08:10which is a coalition of AI's company representing a particular sector.
08:15So, you have, you know, cursor for code, perplexity for search, edge company for computer use.
08:21And I think, you know, it's not just a new technology within the AI world.
08:26It's also a way to make sure that AI is really executing.
08:32So, you know, with ROI.
08:33So, ROI can be financial.
08:35I save money.
08:36I create new businesses.
08:38But it can also be non-financial.
08:40We have a hospital where the ROI is to decrease by two,
08:45the amount of time you spend waiting at the emergency room.
08:49And I think it's very important if we zoom out because we are in our democracies.
08:55People are aging.
08:56We have a demographic problem.
08:58And we need people more and more to be like doing the things that you are not supposed to do
09:02and less and less being in front of a computer doing repetitive tasks
09:06and spending too much time doing that.
09:09So, that's what we do.
09:11I'm very happy to work with NVIDIA.
09:13We started to work on the research side.
09:15Now, we are doing go-to-market together.
09:17We are also starting to deploy our own models at NVIDIA to help NVIDIA with their co-operations.
09:24And yeah, great to be in the team.
09:29So, Neil, when we spoke preparing for this event,
09:32you came up with a phrase, you said something that really resonated with me,
09:36which is that open models and platforms mean that you can build not local solutions,
09:45but locally built world-class solutions.
09:48Can you elaborate on that?
09:50Can you unpack that for us?
09:51Yeah, so maybe, you know, open source and open science, which also mean scientific publication,
09:59is the story of Gradium as a company.
10:02So, all of this started from research that my co-founders and I did at Meta and DeepMind.
10:08And the first thing we observed is all the AI ecosystem today has been built around open source
10:14and open science.
10:15It's because people have shared their discoveries that this field has been able to develop so fast.
10:21So, we decided to create Qtai as a non-profit lab focused on open science for two reasons.
10:28One, we wanted to be able to share our discoveries with everyone to help developing new technologies.
10:33We wanted to be able to train students. If you want to train PhD students, which are the next generation
10:38of researchers,
10:39you need to allow them to publish, otherwise they cannot get their PhD.
10:42And third, it's a great employer brand. Basically, if you allow people to publish their results,
10:48researchers know that they are a brand, that they need to be known,
10:52and there is no better visibility than being associated to a very strong open source model or a very
10:58strong research paper. So, we did the first conversational models with Moshi, the first three
11:03times speech-to-speech translation models and so on, really focused around voice. And what we observed
11:07after that is the impact of our models, open models and open science was significant. The models from
11:16Meta were based on it, the models of Alibaba are based on it, the model of Mistral, the model of
11:21some we
11:21cannot tell, but most of the audio stacks in the world are based on the things we shared. And at
11:27the same time,
11:27we were reached out to help these companies, big or small, to develop the next generation of
11:33algorithms so you can even get defensibility when you do open source. So, doing open source doesn't
11:38mean losing also a competitive edge on other aspects. So, we saw the potential to both be still
11:44committed to open science while being able to make a successful company. And that's what we are doing.
11:49We are doing competitive products. We still publish in the top conferences. We still develop models.
11:56We still share our knowledge because we feel that necessary for the field. And in a way,
12:01it's also a very good situation for us because since everybody is building their technology around
12:06our models, we have a very deep understanding of all of these works, which also, in a way,
12:10is a competitive edge that we can keep. And now, to relate that to the other topic,
12:14which is the European AI ecosystem. What is interesting is 2022 was a terrible year for
12:22open source, right? That's when all the culture of big companies started to close for competitive
12:29reasons. And then, there were a few labs that sticked to that. Mistral was one of them and others. And
12:36that
12:36created so much pressure on the big tech that Google started open sourcing again with Gemma. Even
12:41open AI is open sourcing with GPT. And so this, you know, there is this pressure where if there is
12:48so
12:48much brands that you acquire by doing open source, that when you do that, everybody is kind of, you know,
12:53asked to stick with this strategy. And for European, which is an emergent actor, it's fantastic. If you are
13:00building applications, you can build on very strong open source models. The open source models of today,
13:05whether NemoTron from NVIDIA or Gemma from Google are really, really strong. If you now want to start
13:10training your own models, when NemoTron open source is not only the model, but also the training recipes,
13:16you learn from something that is output of significant compute use and significant talent. So,
13:23I think that's why open source and open models and open science is fundamental to the emergence of a
13:30stronger European system. Because when you're catching up as an ecosystem, you basically just
13:36get a gigantic boost from the open source. And also, again, the employer brand of doing open source
13:42is extremely strong. And the strongest talent in AI is extremely, you know, concerned about this. And
13:50when you say not only you can work on competitive products, but you will also be able to publish the
13:55output of your research. It's a great deal for AI researchers. Fantastic. Pierre, can you talk a
14:05little bit about the role that data, open data especially, plays, you know, transparent data sets,
14:13etc., like personas, in the development of these sort of open, trustworthy models? Yes, actually,
14:22actually, it's kind of a bet we've made for almost two years to create players as a place to do
14:27data
14:27as research, particularly. So, putting research on data with the ideas that the data is actually the
14:32place where you can learn new things, you can gain new capabilities inside of the models. And how we
14:36can do that in the open. So, we have initial reflections about how we can create a series of assets
14:41of
14:41reusable data, which we call CoenCorpus, which is right now the largest fully reusable corpus because
14:48everything under free license. So, it's an accumulation of everything that's been released
14:53on the web for the last 20 years. It's Wikipedia. It's got public domain. It's open data, open
14:58government, open science, other kinds of things in one place. And data has actually a large influence
15:03because many labs have been trained on it, including people around these tables and also on Tropic,
15:08also Chinese labs. So, many people actually have been using this resource because it's the benefits of,
15:12okay, we can use data, we can track the provenance. So, I think the first initial idea that we've made
15:18and we train models on it, but now we have come into a new moment, I would say, where it's
15:23not just
15:23about training one model, but there we can build the open infrastructure that are going to be allowing
15:28people to train their own models. And we can disseminate all the kind of techniques that have
15:32been appearing at the frontier so that people can actually take sovereignty in the very strong sense of
15:37the world, not just about, okay, they can own the model, but they can continuously develop it,
15:41and we can also preserve their own unique advantage within companies, within organizations and this
15:46type of things. So, right now, the best we have been making over the last six months is really about
15:51static data, static pipelines, static pre-trainings, or we can create also pipelines and create data we can
15:56access, typically for privacy reasons. So, we have typically one model we're going to present tomorrow,
16:02which we made with the subway of Paris, which maybe some of you have used to come here, for the
16:07monitoring systems, and we have to simulate a lot of data that, okay, they can have to be trained on
16:12this. Same, we're working also with bank sector on the type of things. And so, we got aligned,
16:16actually, with Nvidia and NemoTron, because as part of the many assets they have been releasing,
16:21so it's not just about training recipes, so many data sets they have been pushing, there are some
16:25which are particularly seeds to create static data. And one of the seeds that we have been working on
16:30together was a set of personas, so like people, like you and me, but they have to be credible,
16:35they have to make sense, like you have to have someone, okay, it's an age, STEM education, STEM background.
16:40So, the thing we released together was for France first, so when I started first with Nvidia and
16:45at the US, then we collaborated to make France, and tomorrow we're going to release a set for
16:50Belgium, which is going to be the 10th of their personal release. So, I think that's really a lot of
16:55things kind of lining up together, but I think it's so important also to recall also the context with
16:59these faibles, these type of things, about the anxiety right now that comes with these models,
17:04and suddenly we realize, okay, they can submit to export control, you don't really own these things,
17:08you don't know exactly where you start to get this, and right now I see much more interest of how
17:14most important models, it's not just something that is limited, but you need some expertise,
17:19you need to build on positions and collaborations, but it's something that's going to be making AI much
17:24more powerful the way it has been until now. All right, thank you for that, and so I want to
17:32move
17:32on now to a part, maybe we'll flow a little more freely and grab questions, chime in, feel free to
17:39disagree with each other. What are your views on compute access in Europe, especially around sort of
17:47supercomputing centers that have been trying to become more than just a traditional HPC center, but are now
17:53trying to help incubate AI startups in the AI space? Who wants to? I can start on this question, so
18:04to train open models, so we need to have public and freely access to compute because it's for us
18:13digital commons, and so, and for example, just to have amounts in some figures, so for 2023-25,
18:21the training of our models takes around 1 million of GPU hours, it's huge, and so we fought the support
18:30of
18:31the French government, the GenC, the CNRS, and because we publish all our research as open source
18:41documents and publication, so we have a direct and free access to this infrastructure,
18:47infrastructure, and it's huge, and now, because we started with GenC in 2023 on H1 GPUs infrastructure,
18:59and for example, I would like to raise this point because it's very important, we'll be the first,
19:07one of the first team in Europe, but even in the world, to use this new supercomputer based in France
19:16called Dahlia on the next generation of GPUs on the GPU 200 infrastructure, and so we will
19:27train to train a new model on FP4 infrastructure and using NemoTron framework for the training and the
19:40reinforced model we use, and so at the end of the day, because all of this research and all our
19:49experiments will be published on open source way, on the public and international conference,
19:57and I think it's the most important when Neil talk about open source, open source is a way also to,
20:04you know, to make innovation incrementally, and each part of our work on data, on engines, on voice,
20:12we can mix it and speed up the innovation of all the ecosystem, and I think it's a great way
20:19for
20:20Europe to accelerate and speed up its sovereignty. And where do you, oh go ahead. Yeah, I wanted to add
20:28something also about compute in Europe. I think, again, if we see Europe as an emergent ecosystem,
20:35right, a challenger to more mature ecosystems, there is an ability from very large companies to
20:42just drain the compute market, you know, by deploying more capital very quickly and so on and so forth,
20:50which can become a challenge to develop domestic AI models in Europe, which is a bit sad because
20:57France, in particular, is a fantastic country to host data centers with our energy production,
21:02and so that's something that, you know, becomes more and more strategic. I think for us, Mishra was
21:08kind of early to understand that, but it becomes a very strong topic, and then there is not only the
21:13compute for training, but then there will be also the compute for inference, because now training a
21:17model is also done during the inference, right? Most of the, a lot of improvement from the model comes
21:21from very large-scale deployments that allows you to collect traces of, of, of, of your model and
21:27and user feedback and so on and so forth. So again, this, I think at some point now with reinforcement
21:33learning, there is such a direct way of converting compute into intelligence that having more compute in
21:40Europe means more intelligent models in Europe in a very straightforward fashion, I think.
21:46Yeah. Yeah, maybe I can add also on this, because the model we're going to release tomorrow is RTP,
21:51is the first example of a collaboration between Gen-C, when we train the models, and ScaleWave, where it's
21:57going to be deployed. So I think this is the kind of interaction that we're going to need, because
22:00right now, HPC is really, really precious, not only because it's a place where you have GPU and compute,
22:05but also where you have expertise, actually, where people actually have been using distributed compute at
22:09scale for many years now, which is kind of lacking the private sectors in Europe right now. So we need
22:14this kind of
22:14interaction, but they get it to get read. And I think also, clusters are kind of underused right
22:20now, I would say, also as a place of knowledge and experience and places of where you exchange
22:23the data, the type of thing I don't see as an end-off right now, intervening in policy in Europe,
22:27and we definitely need this kind of expertise.
22:29Yeah. So, oh, go ahead.
22:31I mean, I think you've all covered very well the topic. I'll reinforce what you just said, Neil.
22:38For a lot of us, training was about, like, putting data in models for a very long time,
22:42and I think that's still something we need to do. But to give you an example,
22:47we shifted from doing that for computer use to environments. We have, like, tens of thousands
22:54of environments where you have these virtual humanoid robots, if you want, that are trying
22:58to do things, that are breaking things, that are trying to click, to scroll, to do things.
23:03That's what we call reinforcement learning, and that needs another kind of compute that is more, like,
23:09the same compute that you use for running the models or inference, rather than training.
23:14And it's the same kind of compute inference that you need at our customers.
23:19So we also start to see a phenomenon where, you know, our customers are using it during office hours,
23:24and then during lunchtime, during the night, you can use your inference to do training of your models.
23:30You can basically put more virtual humanoids, like, playing with things, and then they become quiet during the day.
23:36So for a company like us, which started by training models, it becomes more and more interesting to have a
23:42full stack.
23:42So you can train, you can inference, and you can inference for training.
23:47So on that topic of the full stack, right, how valuable is it to you to have a consistent stack
23:54that means you can train on one supercomputer and migrate some of that to another,
23:59and infer at a computing center, in a hyperscaler, or on-prem?
24:08What I can say is, if I go to my research team and I say,
24:13hey guys, we have to switch our compute provider, it's not going to, you know, it's so much work.
24:20It's, you are storing your data, you have a lot of things that you learn to be comfortable with, and
24:25so on.
24:26So this continuity and providing the ability to get a consistent experience,
24:32expectable results in the performance of your training, in the reliability of the GPUs, and so on,
24:37I think has a lot of value.
24:39However, what is sure is the fact that most are building around CUDA.
24:44It's something that is very strong, and even for talent, you know,
24:48it's a, it's, that's also why it's a barrier for other kind of accelerator is quite hard,
24:53because now it's kind of a race on so many fronts. Honestly, the last thing you want to do is
24:58to have
24:59on top of that to reinvent the, you know, reconsider the very basic. So maybe, you know, at some point,
25:05there are some considerations that make it necessary, but most of the time, it's, it's quite a strong and
25:12reliable foundation, yeah.
25:14Maybe on that,
25:15the size of the models
25:18that we see are being used is decreasing, because, okay, it's easy to use a jet engine to do everything,
25:25but if you are on a small scooter, you can have a small engine, and at the beginning,
25:28we saw these companies like starting to deploy big models, of course, they could do everything,
25:33but they cost a lot of money, and they are very bad for the planets. What we,
25:36what we see right now are models that are becoming smaller and smaller,
25:41is that sometimes you don't even need compute, as in, you don't even need to contract with the
25:46hyperscaler. Nvidia is doing a DGS Spark, for instance, it's the size of a big red rose, like,
25:53it's a small square, and you can plug it on a computer, and we do, some of our pilots doing
25:58that,
25:59and it's, obviously, you can't, like, you know, create a super sophisticated 3D diagram of something,
26:06but a lot of the tasks that company needs are actually very simple tasks, and so you can execute
26:11it almost what we call on edge, meaning on the device. We also have collaborations with a big
26:17phone provider in Korea, where they want to put AI in their phones, so, so I think, obviously,
26:22data centers and GPU centers and, like, huge projects are useful, but I think it's interesting
26:29to see, also, that modems are getting smaller, while being good enough to execute tasks at a new
26:35thing. Yeah, I think you can also add on that, I think what's also interesting with the shift we
26:41have been seeing, also, is how we think in terms of continuous training, which also kind of push
26:45towards the edge, because, obviously, you don't want to retrain a super large monster,
26:49but, suddenly, if you have good approach to take data, and you know exactly what you want to do,
26:54and you want to really codify a very bonded agent, well, okay, you have, you want to put the
26:58infrastructure out of the company, work a lot, find always bonds, and they have their own GPUs,
27:02so, it can be an issue, obviously, on the side of infrastructure, but what we see, actually,
27:07that more and more, actually, models are integral part of the infrastructure, and the nice part is that
27:11once you have inference working, you have many things that can work in terms of the data processing,
27:17and this type of thing is only going to align the model itself. So, yeah, I agree, I think most
27:21or more, we are going to start to think into the model parameter size in terms of almost a search
27:26space, basically, how much information you want to put inside, how much you want it to be updatable,
27:31and how much you want to get performance out of it, and sadly, going to be the largest one,
27:35but you need to have this kind of integrated vision, so you can spend much more time on
27:40static data, which is both generated by the model itself, but also going to help training it,
27:45and everything is becoming kind of fuzzy in the end. So, we've touched a little bit on data again
27:52in this discussion. Compute is now more mobile, we can move compute around to the edge, we can
27:59move it into the cloud. How is data anchoring where you choose to run compute? Is that now the deciding
28:08factor, or is it still where you can get the compute? And maybe we'll go quickly on this round,
28:14so we get just one more question, and before the end. I can talk for voice, right? So, when you
28:19train an audio model, you train it on audio files, uncompressed, WAV, FLAC, very large files,
28:27so you are pre-sharing your file system pretty hard, and all of this needs to not bottleneck the
28:33training of the model in itself. So, your IO on the file system is very, very intense,
28:40and you need very efficient access to data. So, already voice is, in terms of capacity of storage,
28:47and ability to read and write very quickly is extremely important. Now, I think for a new,
28:52you know, also other modalities like video world models and so on, at Qtizer is a world models project,
28:58that means reading frames and videos in real time. It's even worse than voice by at least an order of
29:05magnitude. So, then if you don't do it right, and you don't have the right data system and abstractions
29:15and so on, you're going to lose all the upsides from your compute in terms of efficiency, because
29:20most of the time will just be done pre-processing the data. And so, going back and kind of capstoning
29:29some of this, what does it mean to you to be, you know, a French-based company,
29:37but playing on the global stage, right? Working with global partners, working with global customer
29:42bases, maybe just run around, and that'll kind of take us right through the end.
29:47So, for a company like Sinagora, we started in 25 years ago. So, the main market we address now,
29:54it's France and a little part of Europe, because we are very dedicated to critical and very sensitive
30:01system. And so, we have decided by strategy to make an operator to unveil this kind of deal.
30:11But, in terms of R&D, it's what we said before, we work with everyone, everywhere in the world.
30:22We take some part of DeepSeq, we take some part of QTI, and it's why open science and open source
30:30is
30:30very important, even for a company like us, because even if we are very dedicated to the French market
30:36or European market. So, you can speed up your own technologies by using, you know, the piece
30:44that comes from everywhere.
30:47Pierre?
30:49Yeah, I think what's important is the fact that entire research is a global conversation,
30:53particularly at this point. And it's really important to keep it up with this and to have
30:57this kind of interaction. We have a lot of interaction right now, even with Shining Labs,
31:00and in terms of, okay, we get some data out, and they get us access to some models, this type
31:05of things.
31:06Also, actually, even more so, I would say there is this kind of emerging new countries in India,
31:11Korea, and Japan, which does have the kind of consensus Europe in terms of sovereignty,
31:17in terms of entiling quality. So, there's a lot of communities in terms of research you can also
31:21benefit from, so I think that's one of the aspects. And the other one is, yeah, definitely
31:25in terms of opening markets, we have been, we definitely see this kind of shift, because
31:29usually when you're a small company, okay, you start building nationally, and maybe you expand
31:33very close. Here, we definitely get the kind of opportunities which can be very, very far
31:37from our original place, because, yeah, we are kind of known in the ecosystem, we see
31:41circulation for user data, and definitely opens new doors.
31:45Okay. I mean, sovereignty, we talk a lot about it, we saw orthopedic mythos, we saw these things.
31:50I think you only win if you are a global leader. If we, you know, decide to do something,
31:57and we stay in France and Europe, it's like, you know, trying to run, but not going to the Olympics,
32:02you're not going to fight with, like, the best players, you're not going to be in the most
32:05competitive markets, and then you're going to slowly become a second zone model. So, so to me,
32:14that's the most important. We need to be in Korea, we need to be in the US, we need to
32:18work worldwide.
32:19We are based in France, the capital and the people, most of the people are based in France,
32:25but when I'm in the US, I need to be a US company. When I'm in Asia, I need to
32:29be an Asian company,
32:30meaning I need to understand what people expect from us, what the market needs, and that's very
32:35important. And I think in France, we had a lot of innovations in the past, if you look, but a
32:40lot of them
32:42failed because they were only in France. We invented the mini-tail,
32:47it stayed in France and never really become global. Maybe if it had been become global,
32:52the story would have been different of what happened next. And I think it's very important.
32:57Sovereignty is great. Sovereignty, we need to be sovereign, but if it's an excuse to choose within
33:04a smaller league, I think it's not serving us well. So that's what I would say.
33:12Yeah, I agree with what Gautier just said. So, I mean, it's quite easy for me to be very bullish
33:18on
33:19European AI because, you know, I saw from the inside how Lama was developed in Paris.
33:23The GMA models from Google that I mentioned, they are developed in Paris, actually.
33:27Most of the audio technology of Google was developed by my team in Paris. So the talent is very strong,
33:34and AI is an industry of exceptional talent. And talent in Paris is exceptional, and Paris is
33:40very attractive to other European countries, Asia. So we have a very strong pool of talent,
33:46but the goal is absolutely not to restrict ourselves to the domestic market, but rather to
33:52become, you know, an international leader and deploy our go-to-market internationally. So for us,
33:58we see international expansion mostly on the go-to-market side. And France and Europe has a very strong
34:05anchor in talent. Fantastic. Thank you all so much. It's been a pleasure. I'm sad that we've run out of
34:14time. And thank you all for joining us and staying with us through the whole time. Thank you.
34:21Thank you. Thanks.
Comments