Ep 31: CEO and Co-Founder of Mistral Arthur Mensch on the Next Frontiers for LLMs, Why Open Source Will Prevail and AI Safety

Mistral AI is often seen as the startup challenging OpenAI and incumbents developing LLMs. On this week’s episode of Unsupervised Learning, we sat down with CEO and Co-Founder at Mistral AI, Arthur Mensch. Arthur shared with us his view on why open-source will prevail, how Mistral gets LLMs into the hands of enterprises, the build vs. partnership decisions, the competitive landscape and future of LLMs, and how he’d regulate AI safety.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Jacob Effron?

  This is unsupervised learning, Red Points AI Podcast. I'm Jacob Effron, and I'm joined by my co-host Jordan Siegel. And today, we had a celebrity AI guest in Arthur Mensch, the CEO and co-founder of Mistral. Mistral is at the epicenter of the AI zeitgeist right now, building one of the leading open source LLMs, being used in all sorts of interesting ways. And we had a fascinating conversation. We hit on the future of AI policy, how Arthur thinks about the different products Mistral builds, and where they're going next. And we also talked about the future areas of development for large language models. I think folks are really going to enjoy this conversation. Definitely stick around for Jordan and my debrief at the end. Thanks so much for coming on, Arthur.

• 0:36Speaker 01

  Thank you for inviting me.

• 0:37Jacob Effron?

  I know, it's a busy week, so we appreciate you taking the time.

• 0:41Speaker 01

  Yeah, it's early in the morning, but yeah, a lot of things to do after that.

• 0:44Jacob Effron?

  Got the espresso there, so hopefully good to go. You know, I think to start, it's funny, I've listened to all your podcasts before. I'm not sure anyone has asked you this yet. I want to ask about like the name Mistral. As I understand it, it refers to this like specific wind in Provence. Why do you name the company that?

• 0:59Speaker 01

  So there's two stories, there's the official one and then the true one. I can give you the official one. So in Mistral, you have like two voyeurs, which are I and A. And in French, when you say artificial intelligence, you say intelligence artificielle. And so you have this Mistral AI, and so there's IA AI. And that's the official story. The true story is that we struggled a bit for finding a name. And then we thought about like using like primary forest names, like mountain names. And eventually, we came up with Mistral, and it sounded good. And I think that's for the official reason. And so we decided that it would be good enough. It's French, it's a French wind, blowing in the south of France, usually a cold wind, but it's a wind of change.

• 1:39Jacob Effron?

  Wind of change, I like that. And then I have to ask, I mean, I feel like you've taken the, you guys are just so good at kind of capturing the hearts and minds of developers. And part of that is, you know, I feel like you made this conscious decision to just roll out your models via torrent. And I feel like that was, people loved that, like, how did you decide to do that? Was that something you thought about? Was it spur of the moment?

• 1:58Speaker 01

  So I think we collectively decided that it was a joke because it actually occurred before with Lama. Lama was the first version of it was under a gate. And I guess some random guy just decided to put it on a torrent. And so we figured out it would be a good reminiscence of that and it worked out pretty well.

• 2:20Speaker 02

  Yeah. And then how about the font? I feel like your logo is a lot of fun.

• 2:24Speaker 01

  Yes, the logo. Yeah, so we have two logos. The first logo, I guess the backstory is that there was a random guy that stole our account on what was called Twitter at the time.

• 2:35Speaker 02

  Really?

• 2:36Speaker 01

  Yeah.

• 2:36Speaker 02

  Were they trying to sell it back to you?

• 2:37Speaker 01

  They took Mr. Hali and they pretended that they were the official account. And so we had to create another account to complain to Twitter that we were stealing our account. So we had to find a logo and I was at a wedding. I think Guillaume was also, it was during the weekend, so he was a bit busy. And so we figured out the world art would be fine. So it happened.

• 2:57Speaker 02

  I love that it stuck. Well, I think it really, you know, has defined the brand and I think everything happens for a reason, right? It worked out well.

• 3:03Speaker 01

  Exactly, yeah.

• 3:05Speaker 02

  So when you sort of look at the landscape right now, I think it's sort of solidified between the closed source offerings. You know, you have OpenAI, Anthropic, Google, and then you have sort of the open source side. You have you, Meta, you know, there's also Grok now, I guess. But how do you sort of see that evolving over the next few years? Do you think that's solidified for good or do you think it's going to change? And there will be other players, maybe?

• 3:26Speaker 01

  Well, we think that this is going to solidify with the open source prevailing. That's the, we really think that this is an infrastructure technology and that at the end of the day, it should be modifiable. It should be owned by customers. And so that's what we want to propose. So today we have two offerings. We have an open source one and commercial one. This is something that might evolve, but generally speaking, it's to find out the business model to sustain the open source development. And we really think that this is how it should be done. But currently, effectively, you have companies that are using closed source offerings and companies that are using open source offerings. There was a slight gap in performance that we are actively working to close.

• 4:06Speaker 01

  And there was a slight gap in usability as well with the closed source offerings that had also better software surrounding on APIs and fine tuning APIs, et cetera. And this is also something that we need to close and that we're working actively on it.

• 4:21Speaker 02

  And you recently launched the closed source offerings with small, large and embed. How do you internally sort of decide what goes in the closed source and what goes in the open source?

• 4:30Speaker 01

  We need to be relevant on the two fronts. And so for the open source offering, we went to Mixtale, we released it. It's actually still the best model out there if you count in efficiency. And so this is going to be a very moving target, I think. We released Mistral Large as a portable solution, so it's available on our cloud. But when we ship it to customers, we actually give them access to the weight. So I guess the usability is similar to what you can have with an open source model. Right now, the way we think about it is that our very best model is shipped with commercial licenses. And the ones that are just after are actually open source. But we're not committing to anything because it's very moving.

• 5:15Speaker 02

  Of course.

• 5:16Speaker 01

  There's commercial pressure, there's competitive pressure on different sites. And so we are being very tactical in what we do. But our mission is to be the most relevant platform for developers. So this is definitely setting the high bar for the open source offering. And we'll keep producing some interesting things.

• 5:36Speaker 02

  On the commercial side, how have you thought about giving the weights to enterprises versus maybe, obviously, the other ones charge via API access?

• 5:43Speaker 01

  Well, there's some enterprises that know how to fine tune models, that know how to deploy models. And when you license the technology instead of licensing the service, you allow them first to deploy where they want to deploy it. So they typically deploy it where their data is. And so that solves a bunch of issues related to data governance. And then it also enables them to do the specialization they need and to potentially connect it to, well, to make things that are, well, to make applications that are more involved than just using an API.

• 6:21Speaker 02

  Yeah.

• 6:22Speaker 01

  And so we've seen a few customers do these modifications and make them.

• 6:28Speaker 01

  Well, benefit from the access to the weights.

• 6:30Jacob Effron?

  Yeah, no, it's super cool. I mean, you talked about kind of this enterprise platform and allowing folks to fine tune their models. And then obviously, you know, folks, you'll have a hosted solution. And I'm curious how you think about like, obviously you as Mistral can build all of this stuff, the entire developer facing experience. We've seen kind of an explosion of infrastructure companies that are trying to do parts of this as well. How do you think about like the parts that you want to definitely build and own versus, you know, maybe partnering with other folks?

• 6:54Speaker 01

  So the one thing that we do best is training models, that's for sure. And this is our core offering. The other thing we do well is specializing models. So that's why we're moving into offering some customization abilities. The thing that we need to do is to run inference, because that's actually at the end of the day, having like access to a model file is not, not, not cutting it. So you need to have something else. And so we have been leveraging some partnerships. We, we have been building our own inference pipeline, actually. But this is currently, this is not the best thing we do. But this is something that we believe will, will be, it's also within the realm of what we should be doing.

• 7:39Jacob Effron?

  Yeah, no, it's, it's super interesting. And I guess, you know, with regard to those partnerships, you've obviously announced some really interesting ones recently. I feel like Microsoft, Snowflake, and I don't know who's who of partners you could have. Maybe first, like, how have you thought about, you know, your partnership strategy? Is this, should we expect a lot more announcements down the line? How does it kind of fit in?

• 7:57Speaker 01

  So we, we announced, yeah, Microsoft, Snowflake, Databricks, Nvidia. The way we thought about it is to basically look at what enterprises would need, where they were operating, where the developers were operating, and figure out the channels that would facilitate adoption and spread. So optimizing for distribution, making sure that if you have a developer that is on Azure and he wants to use a model and he knows about Mistral, because he has been using the open source model on his laptop. Well, just look for, well, Mistral on the, on Azure AI Studio and he finds it. And that's, that's the very, well, that's the strategy we've been following to have, to be a multi-platform solution and to replicate the solution on the different platforms.

• 8:41Speaker 01

  Obviously, the hyperscalers come to mind. Those are very natural partners because enterprises are usually customers of them. The other thing that comes to mind are data cloud providers. So that ends the Snowflake and Databricks partnerships. It's a Enterprises have moved their data to these providers. And as it turns out, AI becomes very useful when you connect it to data and you want to do business intelligence with it. So that's a very natural channel to use. And there's great things to be done from within this platform using the AI technology we're providing.

• 9:14Jacob Effron?

  I was really interested in thinking about even with OpenAI, I feel like they obviously sell directly, then they have, you know, folks can access it via Azure. You know, it seems like similarly, folks can go direct to you guys or go through some of these other platforms. What are you seeing from enterprises or what do you think the future is of folks that kind of go direct to you versus, you know, via some of these other places?

• 9:32Speaker 01

  So the small companies, digital natives are oftentimes willing to go to our platform directly. And the good thing is that they get like very direct support and we set up like Slack channels to support them. Enterprises that we've been discussing with, and in particular European enterprises, that were the first people we discussed with, typically want to use our technology, but they don't want to go through the procurement process. So that's where being able to say, okay, we're on Azure. So just use your Azure credits for buying the technology and we'll support you has been pretty instrumental in, well, for adoption.

• 10:12Jacob Effron?

  You've been involved in a lot of advancements in AI from Palm and Chinchilla, DeepMind to now Mistral. How do you sort of view the future and like what are the next frontiers of LLMs in your mind?

• 10:22Speaker 01

  Well, there's still inefficiency frontier to be pushed. The Mistral 7B was a pretty compressed model. I think we can still do better than that. I don't think we are at the end of scaling laws, so you can make models that are actually better than what we did. And I don't think we are at all, we have solved at all the question of making the model controllable.

• 10:48Speaker 01

  I think we've we've done a good job at that. But beyond the compute that you can spend on pre-training, you can make the model much stronger by figuring out how to slightly tweak it so that it follows what you tell it to do. So there's still a lot of research to be done in that direction. On architectures, I'm pretty sure that we can make better than we can make things more efficient than just plain transformer that spends the same amount of compute on every token. But obviously, this is kind of research directions. And yeah, that's the we're going to scale more, we're going to be more efficient, we're going to deploy models on smaller devices, we're going to improve latency, make models think faster.

• 11:31Speaker 01

  And when you make model think faster, you're you're opening up a lot of applications that involve calling an L L M just as a basic brick. And then you can figure out how to do planning explorations and all of these things. So as we improve efficiency, we're opening up areas of research that are immense, I think.

• 11:48Speaker 02

  Yeah, speaking of compute, Meta announced, you know, 600,000 GPUs and, you know, how do you sort of think about that relative to Mistral, especially on the open source side of things?

• 11:57Speaker 01

  So Meta has more GPUs than we do. We do a good job with it. I think that's also the the strategy of of staying lean, of having enough GPUs per person. I think we have a very good concentration of GPU per person. I think this is the way to be as efficient as possible to come up with creative ways of training your models. And so far, we've done a very good job with with just 1.5K H hundred. So that's and we're going to increase. This is increasing like in the coming months. So we expect to be able to ship better models with better compute. Going to free like for a startup, well, startup can't just buy 350K H hundred. That's that's kind of expensive.

• 12:41Speaker 01

  So there's also a question of unit economics, like how do you make sure that one dollar that you spend on compute, on training compute, eventually accrues to more than one dollar in in revenue? And I think that's going to take time, obviously, and that's why that's why VCs are there are there. But. That like, yeah, being efficient with the training compute is actually quite key in having a valid business model at the end of the day.

• 13:04Jacob Effron?

  You know, obviously, some of these other folks will just spend more on compute, you know. You've done an amazing job keeping up with the with state of the art, you know, throughout like, do you think as more and more compute goes to, you know, G P T five, six, seven, like. What do you think the timeline ends up being for, you know, the Mistral model to catch up?

• 13:21Speaker 01

  Well, I guess the the challenge for us is to to have enough compute to stay relevant, so we're. But we have a very good understanding of what we need to be like the best model provider.

• 13:31Jacob Effron?

  Yeah.

• 13:32Speaker 01

  And that might involve using less compute than our competitors. It's it's, I mean, I'm hesitant to answer that questions because I'm, I mean, there's still many things that we don't know. When does it saturate? Does it saturate? What do you need to do to avoid saturation because you're running out of data? So there's many unknowns, which means that we're doing research to try out.

• 13:52Jacob Effron?

  Well, if you don't know, I feel like everyone points to your chinchilla works.

• 13:55Speaker 01

  Well, Chinchilla showed that you could divide by eight the the size of the model and still get the same performance. And then what we showed with Mistral seven B is that you could get to like, you could get a factor of six again. So that's the, and I don't think we are at the end of it yet. There's obviously at some point you can't pour the entire human knowledge in like five gigabytes. So there's some trade offs and some some asymptotes that that you can see that you see appear. But I'm pretty confident that we can push that further.

• 14:24Jacob Effron?

  What about like alternative architectures to transformers? I feel like every so often, you know, we'll we'll get a pitch that's like, you know, ah, transformers are suboptimal on this. Here's like an entirely new way of doing things. Any chance or, you know, as you think about the next decade, how confident are you 10 years from now transformers will be what we're using?

• 14:39Speaker 01

  I'm pretty confident this is not the optimal architecture. On the other hand, transformers have been out there for seven years now. And so that means that everything has co adapted to it. So that includes, well, the way you train it, all of the training optimization algorithms, the way we think about it, the way we debug it. This has and also the hardware. This has been very well co adapted to transformers. And so it's hard to restart and consider a new architecture and then figure out every sliders, every lever that you should activate to get to the same performance of a transformer. So the ladder, the ladder climbing that we've done with transformer. Kind of set the bar very high and that means that changing the architecture in a

• 15:23Speaker 01

  non incremental way is actually quite challenging and that's why I believe that everybody is still using that that architecture. What we have done and proposed our improvement on the attention side, so how do you do sparse attention to improve memory efficiency. There's a bunch of things that can be done in that direction, in that direction, but generally speaking, I wouldn't be too bullish on a startup pitch saying that they have invented a new architecture because the path is pretty long to get to the baseline performance.

• 15:58Speaker 02

  It's good to know for us. I know you gave a keynote this week at TDC, you know, speaking of the hardware side of things, what, any thoughts on Nvidia and their latest chips?

• 16:07Speaker 01

  Well, generally speaking, it's improving. The important thing for us is total cost of ownership, so if you have more flops and the flops at the end of the day are are less expensive, that's great because you can train larger models. Nvidia is moving toward very big chips, which is quite interesting because it's basically focusing on the training side because like the GB200 is very useful for training. It's expensive for sure, but I think there's great things to be done with it. But yeah, I think we are, the good thing about Nvidia is that improvements in terms of dollar per flops are pretty much predictable. And so I think this new generation brings a further 30% improvement, basically, and

• 16:52Speaker 01

  that's, that's great to have.

• 16:53Speaker 02

  Yeah, great for you guys. You know, there's been a lot of activity in the last week regarding the EU AI Act. I know you've commented on the EU policy before. How do you sort of see that evolving over time and what do you think regulation should be like?

• 17:06Speaker 01

  So we have, yeah, we, we stated our position on that aspect that basically AI safety was something to be addressed through a product safety perspective, just the same way we have addressed software safety in general. So really focusing on what is the product, what is expected from the product, how do you evaluate that it works? And that's initially, that's what the AI Act was proposing. As it turns out, there was a lot of lobbying against the technology, against, I know, L L M's becoming a Bersac and and and getting out of control if you were to cross a certain fluff threshold. And so at the end, this lobby seemed to have worked because they introduced

• 17:47Speaker 01

  some like technology regulation, forcing, evaluation, re teaming, all these things. I think it's okay. I mean, we're doing it anyway, but we're obviously re teaming our models, we're evaluating them. But it doesn't solve the the product safety problem because basically models, L L M's are like, just like a coding language, it's program machine language, and so you can do whatever you want with it. And so you can try to evaluate it on the scope of of things that you know the model can do and shouldn't be doing. But at the end of the day, it's even if you have access to these evaluations, this is not going to certify that the products that are made with it are actually safe. So for

• 18:25Speaker 01

  us, it's just a, it's a, it's a ill directed burden. And it's not going to improve things too much, but it's very much manageable because we are doing the evaluation and documentation part anyway. On the A I Act, there's obviously some discussions around transparency of training data sets that that we'd love to enable with the caveat that you need to keep some trading secrets, because it's a very competitive landscape. So we've been engaging in discussions around that topic. We know that in the U S, it's evolving in like actually similar directions. It's going to be like a double burden because wherever you operate,

• 19:04Speaker 01

  you need to be compliant. But it's very much manageable. I think the issue of all of these regulations that, yeah, again, they don't solve the actual problem on how do you make an A I product safe? And the reason why they don't solve it is that it's a hard problem. You to make an A I product safe, you need to be Able to evaluate exactly how it performs and those are stochastic models. So it's it's a different way than we were. It differs from the way we were evaluating software before. So we need to rethink evaluation. We need to rethink continuous integration, verification that everything is happening as as it should be. And this is, I think this is more of a technological problem, a product problem.

• 19:46Speaker 01

  How do you, how do we as a company give the tools to developers to make sure that application makers verify that their application is working as expected? It's more of a product problem than than a regulation problem. But due to lobbying, we have seen regulation come up pretty early in the process. It's very much manageable, but it's not solving anything at that point.

• 20:07Jacob Effron?

  We see a ton of AI safety startups, eval startups. Do you think that should be handled primarily from the model layer and the companies like yourself or third party providers?

• 20:16Speaker 01

  It's great to see this ecosystem of startups doing what we could call middleware, like observability startups. It's great to see because they they have good ideas and and it's a bit like the user experience, the developer experience you need to have for this kind of product, evaluation product is a bit unclear. This is not something that we have been doing, but I think eventually that might consolidate into like a platform, an AI platform, an LLM platform should provide such kind of services at some point. And that could go through partnership or that could go through building it ourselves.

• 20:50Jacob Effron?

  And if you were like, if you were with Adam and Mr. All and you know, put into government to kind of figure this stuff out, like what would you do to kind of make the, you know, it it sounds like, you know, which makes a ton of sense. You you'd rather see the application layer, you know, monitored. Obviously, maybe you go build some some eval tools that could do that. But what would you do if you were kind of in the policymaker's shoes?

• 21:09Speaker 01

  So I would, and that's what we suggested. I would put some pressure on the application makers to basically force them to verify that the task that they are solving is actually well solved. So for them to do some form of evaluation, the same way you do safety testing with Caus. And and that would put.

• 21:29Speaker 01

  That already puts like a second order pressure to the foundational makers, foundational model makers. Because if you want your model, if you want your application to be good enough at a certain task, you do need to verify that it solves that the model you're buying is solving.

• 21:47Speaker 01

  And that's the way we think about it. So since application makers needs to comply with rules, needs to comply with the existing regulatory framework, we should be providing them with tools to comply. So evaluation tools, ways of improving the model so that it actually, so that you verify that it solved the task it should be solving.

• 22:08Jacob Effron?

  Yeah, it's interesting. It sounds like, you know, if you end up regulating the application side, it will still flow back, you know, to you guys ultimately because, you know, they'll just come to you looking for some of the stuff.

• 22:17Speaker 01

  It flows back and it puts some competitive pressure because obviously the application makers are going to pick the model that they can best control.

• 22:26Jacob Effron?

  Right.

• 22:26Speaker 01

  And that puts some, I think, healthy pressure on the model makers.

• 22:31Jacob Effron?

  Totally. A lot of model they can best control seems well suited for what you guys do.

• 22:34Speaker 01

  Yeah. And I guess the alternative that they took is to put some pressure directly on the technology. But that's not a healthy competition because it also favors the big players. Because they can send an army of lawyers to capture the standard making. And that's the one thing we are afraid of, to be honest.

• 22:52Jacob Effron?

  I mean, I guess one trend we've been seeing that I'm curious for your thoughts on is it feels like there's, you know, in recent months, we've seen the foundation models for different countries pop up. Like, you know, we've seen one in India and one in Japan. And I'm wondering, like, one, do you think that makes sense? And two, you know, should we expect in the future there'll be like an LLM company in every major geography? Or how do you think this plays out?

• 23:14Speaker 01

  So I think what makes sense is to enable countries and to enable developers in general to deploy the technology where they want to deploy it. Portability is our approach to sovereignty.

• 23:25Jacob Effron?

  Yeah.

• 23:26Speaker 01

  And the other thing that makes sense is language. Because as it turns out, this is a technology that speaks languages. And The entire world wants to benefit from from generative AI. And as it turns out today, models are just much, much better in English than they are in other languages. And so this is something that we have tried to address, that we have started to address and that we'll continue addressing. In the sense that we want to make models that are great in every language. And we've started with French, because we do speak French. And our models are actually, I think, pretty much the best model out there in French. But this should be, this should be spread to to the entire world, and this is an important aspect.

• 24:06Speaker 01

  Our approach is to be a global company, but to give, well, to be portable and to be multilingual. So that the technology we're making is ubiquitous and can be used by by any companies in any part of the world.

• 24:20Jacob Effron?

  Yeah, it makes sense that that, you know, there'll be, you know, back to our earlier conversation of of kind of the six or seven major model providers. That that's kind of what will, you know, that's that's probably the folks that will still be around. And as the models get better and better, they probably work, you know, around the world.

• 24:33Speaker 01

  You can have a couple of companies that are specializing models on on specific languages. But as it as it turns out, like the language selection and getting good at language is something that is very much on the pre-training side. So it somehow belongs to companies like ours.

• 24:47Jacob Effron?

  Yeah.

• 24:47Speaker 01

  And so that means that, yeah, we do need to work on it pretty heavily and to, yeah, have a bunch of people that can manage different languages.

• 24:56Jacob Effron?

  Totally. To me, it feels like even more of a political question than a technological one. It feels like technologically, the optimal thing is to have a few folks like yourself globally. And then I think the question feels like politically, you know, are major countries going to want to have their own homegrown, you know?

• 25:08Speaker 01

  Well, if they can get access to the technology and be able to modify it as they want.

• 25:12Jacob Effron?

  Yeah.

• 25:13Speaker 01

  Which is the way we want to distribute it. I think that that that should be enough for them to be to be confident that they control the technology. And it's very, so as long as there are companies that are playing a platform play. And so deploying, basically shipping the models to various countries in a distributed in a distributed way, then then I think we're good. But if you if you only have a couple of companies that are offering SaaS services.

• 25:40Jacob Effron?

  Yeah.

• 25:41Speaker 01

  Then then there's definitely a sovereignty problem. And I'm, and I think this is a sovereignty problem that has been identified by many countries already.

• 25:50Speaker 02

  You know, I would love to just like circle back to like the start of Mistral, where you know, you raised this really large seed round. There are some skeptics out there saying, well, there's already OpenAI, there's Anthropic. You know, what sort of gave you the confidence to go out with Timothy and Guillaume and go start this company in the first place at that point in time?

• 26:07Speaker 01

  We had done it before, so I think that was part of the confidence. We had secured a few people to join us from day one, and that also gave us the confidence that we could have a team sufficiently large to speed on.

• 26:22Speaker 01

  And then we saw some interest in Paris, because people realized that there was actually a big pool of talent there. ChatGPT helped, because it increased awareness in the VC world. And so we had an opportunity there, and we were confident we could ship very good models in a limited period of time. And so we just went for it. That's how it happened.

• 26:47Speaker 02

  Yeah, I don't think there's any skeptics anymore, so definitely not.

• 26:50Speaker 01

  There still are, but we'll show them.

• 26:53Speaker 02

  I feel like one thing you guys did recently that was fun is you obviously released LeChat. And I'm curious, like, does that, you know, do you think you'll do more kind of applications on top of models, or how do you think about about doing that?

• 27:06Speaker 01

  So our approach there is to, we realized that for many enterprises didn't know how to get started. And so the best way to get them started is to show them the value, and to show them that they can increase the productivity of their workers by giving them an assistant that is contextualized on the enterprise data. And so we started with this assistant as like just a V0 or something that we want to ship to enterprises, which is called LeChat Entreprises.

• 27:34Speaker 02

  I was scared to try and pronounce that.

• 27:36Speaker 01

  I'll do it for you. I think this is, this is quite important as an entry point. It's also quite important for us to solidify the APIs behind it. So for, for instance, we have on LeChat, we have a moderation tool that detects input that could be deemed like off railing the model, and those things we want to expose on the platform side. So it's really a tool. To be closer to the end user, to get feedback for the developer platform. And it's also a tool to, uh, well, get the enterprises into generative AI. Before they figure out exactly what they should be doing with their core business. Because that's really where the value would accrue. But getting them started, showing them what they can start building. Through something that users can use and not only developers.

• 28:21Speaker 01

  Uh, is the strategy we've been following so that we have these two products. But the most important one being the platform.

• 28:27Jacob Effron?

  And one thing I'm curious about, and obviously you're, you know, in the weeds, working with enterprises, helping them train their own, you know, uh, fine tuned versions of, of Mistral models. How like, to what extent do the big folks have, you know, I think there's this outstanding question of like, you know, if you're sitting on tons and tons of data and you throw that out at, at kind of, uh, adapting a model. Is that much better than you're sitting on, you know, you're a midsize player and you don't have that. I'm curious, like, it's probably an overly broad question, but like, to what extent do you feel like the folks that are, that are sitting on tons of data have a big advantage? And, uh, to what extent, uh, is having some examples enough?

• 29:01Speaker 01

  I think that's, that's a good question. So typically you don't, uh, when you have a ton of data, you shouldn't be fine tuning the moog right.

• 29:08Jacob Effron?

  Yeah, I guess you try.

• 29:08Speaker 01

  You should be using retrieval augmentation, uh, generation, retrieval augmented generation. And you should be thinking on how to use the context to actually empower the assistant with tools, uh, with, uh, data, with the ability to do requests to some databases. Uh, so that's the, that's the first thing that you should be doing. Then the data that you could use to make fine tuning work are more demonstration data. So traces of what, uh, users are doing so that you can imitate them and provide them with an assistant. And in that setting, I mean, we haven't seen a lot of enterprises have this data readily available. And that's something that is probably missing to make assistance that are very robust and reliable. Um, and I think the. So it's a brand new kind of data that you need to acquire.

• 29:53Speaker 01

  And in that respect, I think it's an even field. Companies can start acquiring it faster than others. The data they had before is a great enabler, but they should rethink their data strategy in the light of the co-pilots, the assistant they want to deploy.

• 30:12Jacob Effron?

  Yeah, super interesting. Well, we always like to end our interviews with a quick fire round where we get your quick takes on things.

• 30:17Speaker 01

  Okay.

• 30:18Jacob Effron?

  You know, to start, would love one thing that's overhyped and one thing that's underhyped in AI today.

• 30:23Speaker 01

  One thing that's overhyped, synthetic data. I keep hearing about synthetic data. What does that mean? It's ill-defined. Underhyped, I would say optimization techniques.

• 30:35Speaker 02

  What has been the biggest surprise in building Mr. All? And then one thing that you thought would work and didn't? And one thing that you thought didn't work, but actually did work?

• 30:44Speaker 01

  Would work and didn't. I think so far, things have worked up pretty well. I'm thinking of, yeah, I mean, there's some challenges in hiring. We're like, we're opening a US office for science people and it's actually like getting the gravity to get the best profiles is actually quite hard, but we're starting to get there. One thing, something that surprised us and we didn't know would work is basically we got attention more quickly than we thought. And so that's a great problem to have, but then we, it also creates some prioritization problem because you get a lot of inbound.

• 31:27Jacob Effron?

  Good problem to have.

• 31:28Speaker 02

  Yeah. What do you think of the latest GROK model?

• 31:30Speaker 01

  Ah, it's a big one. It's a little too big. They can make it smaller.

• 31:37Speaker 02

  You don't think 340 billion parameters is where it's at?

• 31:41Speaker 01

  Well, I mean, you can have many parameters, but then you need to have the performance that goes with the parameters. What matters is the paret au front, in between the size of the model and the performance of the model. And they are not quite there yet. They will probably improve.

• 31:53Speaker 02

  So you can't say Mr. All for this, but which AI startup are you most excited about right now?

• 31:58Speaker 01

  There's a couple of Paris startups that are good. Dust is a very good startup, is focusing on knowledge management and they have a very sleek UI and it's great to see.

• 32:10Jacob Effron?

  Yeah, no, it's awesome. It's I feel like you're at the heart of this Paris ego system, but it's amazing to watch it just rocket to the top of the scene. I guess if you if we said you weren't allowed to build models and you couldn't work at Mistral and you had to go build an AI application, what application would you go build?

• 32:28Speaker 01

  This is a very good question because that's a path that we could have taken. I'm very excited about hard science, but this is very much the North Face. So how do you accelerate material science, for instance? It's something that we discussed before that could have been an alternative Mistral. But how do you like accelerate the synthesis of ammonium, for instance, which is a very carbon intensive process today and that could improve? And that requires a lot of exploration. You could use LM for that, maybe. So material science, I think, is still lacking a foundational model. And I would be happy to speak to anyone that is willing to come in.

• 33:05Speaker 02

  Awesome. Amazing. Well, we have wanted to have you on for a while and we're glad we can make this work in person. There's obviously a lot of exciting stuff going on, Mistral. Where can people go to learn more about you and the work you're doing?

• 33:16Speaker 01

  So we have we have two developer relations today. Sophia Young, in particular, who joined us from the US. And so we have been working heavily on solidifying the documentation, the guides, to simplify the usage of our APIs. And so our documentation is a good starting point to learn more about us. And this is going to grow as we as we create more content.

• 33:39Jacob Effron?

  Well, Arthur, thanks so much. This has been a fascinating conversation.

• 33:42Speaker 01

  Yeah, you're welcome. Thank you for inviting me.

• 33:44Jacob Effron?

  Yeah, that was a lot of fun.

• 33:46Speaker 02

  So what did you think?

• 33:48Jacob Effron?

  That was a ton of fun. I mean, God, I feel like Mistral is just at the center of like, you know, the AI zeitgeist right now and everything interesting. And, you know, even being at NERPS, I feel like they were just the talk of the town. And it's incredible just the different things they're doing. And Arthur's just such, he's a great guy.

• 34:05Speaker 02

  Yeah, no, I mean, the first time we met him, we were like, we have to work with this guy because he's so humble and, I mean, just so smart. But yeah, I mean, I feel like every practitioner we chat with, they're like, yeah, I'm using Mistral.

• 34:16Jacob Effron?

  There were just so many fun parts of that interview. I love the idea that he's like... At a wedding, you know, it's over the weekend and they're trying to figure out like, what do we, you know, what logo do we launch with? And it's like, I guess word art will be just fine. And it's become kind of this like iconic logo that they have. And, you know, it feels like, it was funny, but we asked him, you know, if anything had, you know, surprised him to the negative. And it's like, it just feels like all these, you know, these decisions they make quickly just resonate so well with, with the, with the broader community.

• 34:42Speaker 02

  Yeah, I also like the name part. That was pretty cool, too. The Winds of Change. I think that's a pretty good name.

• 34:46Jacob Effron?

  Yeah, exactly. I like the, you know, like every good name, there's a real story and, and then the actual, like the one, the one that you tell the, the, the press and whatnot.

• 34:54Speaker 02

  Exactly. What did you think of sort of the, the foreign sort of policy side of things?

• 35:00Jacob Effron?

  Yeah, I thought the policy side was, was super interesting. I mean, it's definitely a theme we've heard throughout, you know, throughout these conversations is the idea of regulating the applications and not the models themselves. But I thought his point, which makes a ton of sense is like, look, if you go and regulate the applications, the pressure is going to flow up to the models. Because the application makers will just then go back to their model providers and say, hey, we need X, Y, or Z. I think that actually puts, you know, if Mistral's main focus is around controllability, having the underlying weights of your models, puts them in a pretty good position to, you know, help folks adapt to different kind of regulatory regimes.

• 35:31Speaker 02

  Totally. I mean, that, he mentioned the transparency and the training data, and that's gotten a lot of attention because it's very broad, you know, it's very open for interpretation. But for a company like Mistral or OpenAI, the training data is just so vast. And how do you make that transparent?

• 35:44Jacob Effron?

  Well, it's also the secret sauce.

• 35:46Speaker 02

  Right, exactly.

• 35:48Jacob Effron?

  Totally reveal all the, all the company secrets. And so, yeah, very curious to see how that gets implemented. And, you know, it all, it feels like it all comes back to eval always in these conversations. It's such an unsolved problem. And, you know, it's, I think it's points well taken that even if regulators wanted to, to do things on the application side, there's, you know, we don't even have the products yet or the tooling to, to really be able to, to look at that.

• 36:08Speaker 02

  Totally. You know, speaking of eval, I mean, we see eval startups all the time. What do you think of his answer? And, you know, we may partner with that. We may just incorporate into the platform. Not really sure yet.

• 36:18Jacob Effron?

  And I love just how transparent and open he is about like, we're still figuring out exactly what like, you know, there's, we basically reserve the right to be flexible here. And like, when you have as much momentum as they do, and you're into as many Places, I think it's a great perspective to take. And so as they think about building out this broader developer platform, I love the idea that, you know, certainly they'll be, they'll build some stuff that they're really good at. And obviously it seems like fine tuning and the ability to train your own models is, you know, real strength of the team. And then for other things, you know, they'll be the first place people look to. If you're, if you're using a Mistral model, you'll say, okay, what, how do I, what do I use, you know, to do eval or to host or whatever. And I love the idea that they can decide what of that they want to build, what they want to partner for. And, um, you know, I think we would love, you know, from our investing standpoint for him to just tell us,

• 37:01Jacob Effron?

  I've chosen a partner and it's this and we're like, great, we'll go do that. But I think, you know, the reality is everything's moving so fast that I love just kind of the flexibility they keep.

• 37:09Speaker 02

  Totally. No, I completely agree. And what do you think of, you know, we talked about like, is there going to be a model or a company for each GIO? He mentioned that, you know, we're doing many languages and we're hiring for different types of GIOs. How do you think about that? And like, do you agree with him that maybe Mistral will just do a ton of languages? And there won't be a need for, you know, a centralized company in each part?

• 37:31Jacob Effron?

  I kind of feel, I mean, I think I said this to him, like, it feels to me more a political question than a technological one. Like I do, I, you know, think from everything we're seeing these, as these models get more and more powerful and more and more generalizable, that you will have the top models be quite effective across different languages. I think there's a broader political question of, you know, do you want the main model that's powering all these economies? To be, you know, to, to just reside in countries that are, you know, sorry, in, in companies that are present in certain countries or as, you know, our major countries going to want to feel like, hey, we have our own provider. And I don't know how that plays out, uh, you know, maybe we'll, we'll have to get some more political folks on the, on the podcast.

• 38:10Speaker 02

  Sounds good.

• 38:11Jacob Effron?

  Yeah. What, I guess, anything else that stood out for you?

• 38:14Speaker 02

  No, I mean, look, I, I mean, we're obviously super excited about Arthur and the team over at Mistral. And, uh, you know, I think every time we chat with him, we just get more and more excited about the direction and how, uh, You know, they are both really strong researchers and just making these crazy advancements, but also reminding ourselves, it's like a 30-person team.

• 38:31Jacob Effron?

  Yeah, and then I thought like another really interesting part of, you know, what he was saying toward the end is just his kind of interest in foundation models for material sciences. And I think this is a trend we've been seeing over the past few months of, you know, obviously there's foundation models for biology, material sciences, robotics is one we've seen with the announcement of physical intelligence. And I think these domains are fascinating because they seem to exhibit similar scaling laws. I think the big question is, how do you generate enough data? You don't have the entire internet to go scrape. And so, you know, I think it'll be really interesting across all those areas. You know, I'm really glad we have Arthur's like brilliant mind focused on the LLM world, but I think in a different world, him going on material sciences would be pretty interesting.

• 39:11Speaker 02

  Yeah, that would be pretty cool. But no, I mean, I agree. I think we've been seeing that a lot on the investing side. I think we're going to see more and more this year where we have, you know, real domain expertise going after foundation models for these specialized industries.

• 39:22Jacob Effron?

  So that was our conversation with Arthur. If you enjoyed that conversation and you've been enjoying the show, please consider sharing this with someone you think that would also find it interesting. We're trying to get the word out about the podcast and we have an awesome lining coming up, including next week with Edo from Pinecone. Look forward to seeing you then.