Anjney Midha on why he backed Mistral's $200M Series A, Sesame AI's companion hardware bet, and the post-training era

to be like I'm profitable. You remember this? He's a master of media. He's always on TV. Like how'd you do that? Anyway, uh our next guest is runs in the family. Uh welcome to the stream. How are you doing? Great. Looking fantastic. What's happening? What's happening? Um uh my name is pronounced Anjene.

Friends all call me An. You should feel free to as well. An fantastic. Uh great to meet you. Um I'd love to kick it off with a little bit of review of what you're investing in today.

Uh what you're excited to presenting at LP day and then a bunch of questions on open source AI and the specific trends that are going on in the in the in the market right now. Sure. I have the fun job of spending the $40 billion that Jen raises for us every couple years. Fantastic.

So I I'm a general partner on our AI infrastructure practice, which is basically all the computing systems that people need and rely on to build great products and services on top.

And I basically spend most of my time um as a glorified customer support um and compute intern for scientists building foundation model companies, training models, all the stuff that happens, you know, at the intersection of of of research and then actually trying to figure out how to get that research out into the real world.

Okay. So I I started at the firm two years ago. Oh, cool. Almost to the day. Um, and I came to the firm after a couple years as a founder. Um, and then I I I founded a company called Ubiquity 6, which we then sold to Discord.

And um, I ran the platform there for a few years during the era where we went from basically being a chat app for gamers to which was about a million monthly active users to about 200 million monthly active basically in 6 months because COVID happened. Yeah.

And um the the homework assignment was an go build a developer platform business for us.

And it was around that time that I got a call from some friends who were running research at OpenAI who said um hey an we're we're we've just trained this model called GPD3 and we think we're on to this thing called scaling laws where if we can combine compute data and algorithms in the right combination, we just might have a shot at creating God.

Classic. Um, okay. How much do you need to get started? They're like, sorry, our our audience, the live audience here got got a little excited. A little excited about God in the box.

Um, well, you know, as a having just gone through the journey as a founder, uh, I was like, okay, how much you think that's going to cost us? I mean, go God should be pretty pretty cheap in these times where where religion seems to have left the room. Um, and you're like, yeah, it we think we can get by with five.

And I said, okay, five million shouldn't be a problem. we can probably scrap that together next week. And uh this was Dario who is a longtime friend and now founder uh and CEO of Anthropic said, "No, Andre, I don't think you realize I meant 500 million. " So that's going to take a little bit longer.

So So I I was one of the early angels into Enthropic at the time. That gave me sort of my crash course to scaling laws for neural language models, right?

The idea being that you could get these these language models to to to predict the next word in a sequence and if you kept scaling up um both the compute the amount of compute you train these things on as well as the data set then you'd have no you'd be able to predictably improve the performance of these models.

I'd call that the pre-training era.

what I spend a lot of my time and so you know for the better part of the last three four years I've basically helped teams like anthropic I usually get a call from them before there's a company when they're scientists who who've who who get who get have this sort of moment where they realize they're they're they're about they've unlocked something um repeatable where where you can very predictably improve the performance of these models on some axes um and they usually give me a call and they say we've we've had the the enlightenment moment we have no idea how to commercialize that and turn that into a business.

Um, and so I've worked with teams like Mistral which is working um on open source models.

Uh, last year I led a our the first round into a company called Black Force Labs which is the who are the creators of stable diffusion which is the open source image model family that I think sort of introduced the world to the idea that open source language and image models can be pretty powerful.

And so that's that's what I spend most of my days doing helping folks you know get their clusters set up to train models to then build products on top. And I think at this LP day the there's sort of two big things I'm getting asked about.

The first is okay an you know two years ago it was all about pre-trained models like the pre-training era AI scaling laws. What's going on now? Why is everybody talking about post training right? Um so that's that's one big question.

Um and then the second is hey what what is going on with with sovereigns like why are countries announcing a hundred billion dollars into data centers and why are they all talking about wanting their own AI you know champion company whether that's mistrol in Europe or um whether it's Alam in in Saudi and so I would say sovereign AI which is really the idea that you want to control your own AI you want to control what it can and can't do um and the idea that that progress in AI is not coming necess necessarily just from pre-training, it's coming from from post training are the two big themes for for this year for me.

Yeah.

Um, on the on the open-source uh question, are you are you thinking about Mistral and Black Forest Labs as kind of like the Red Hat Linux model where they will be almost like implementers in the enterprise or is there a real world where you're underwriting against total and complete victory of the open- source uh of the open source paradigm at the foundation model layer?

And then related to the the international topic you mentioned like is Mistral uniquely um are you uniquely bullish on Mistral because it's an international company as opposed to an American company because it has this this specific advantage of being in Europe and and that driving extra value for the company.

Oh, I no I I think so, of course, you know, Europe is 400 million consumers. So, just massive economy that's that's decided to gear up um to build the basically the single largest infrastructure build out I think we've seen in the continent in like 60 years.

They just passed an $800 billion defense bill they're calling Rearm Europe and like a huge portion of that is flowing to uh AI and computing teams for sure. So, we can talk about that. But that wasn't when I led the the series A into Mistral two years ago, $200 million round.

Um like Europe was we I did not expect that we'd be in the middle of like this massive sovereign AI buildout. I mean the bet was very simple is that if you look at the history of computing infrastructure, there's basically two frontiers.

You've got the capabilities frontier and then you've got the efficiency frontier, right? And the capabilities frontier is usually dominated by closed source.

If you look at, you know, databases and storage and networking and so on, you usually have a company like Cisco and so on that that pioneers some new capability, right? Um and then or or actually in the case of Linux, like you said, right? Um you had Microsoft show up and build Windows closed source, right?

And that usually opens up the aperture for consumers because consumers are often the first to flock to new use cases. The enterprise cares about something slightly different. They care about cheaper, faster, more control. Mhm. And that's usually dominated by open source.

So, you know, two years ago when I was running the the platform or at Discord, we got early access to GPT4 and the OpenAI guys said, "Hey, we've got this new model. It's going to come out in 6 months. Can you guys just figure out what it's good for? " And we ran an experiment.

You know, we we we ran a bunch of uh tasks through it and it was extraordinary on a couple of them and we said, "Okay, great. Let's go to production. " Well, it turns out when you want when you're working with sensitive data, right? In our case, this was a social platform.

So GDPR, CCP, all of the compliance stuff, really critical that our data doesn't leave our servers, right? You want control over what where the weights are running and what the weights can and can't do. And so we looked around for a for an open source alternative to to GPT4 and we just couldn't find one. Mhm.

And that's when I realized, okay, for every dollar that that you're seeing in like enterprise or large company prototyping or proof of concept revenue, there's like 10 more waiting for the open- source alternative, right? And really, there was no alternative until later that summer when Llama showed up. Mhm. Right.

Llama was the first time there was a comparable open-source alternative for to to a closed source models and and well, the creators of Llama left and started Mistral, so that that made the investment easy. That's great. Uh can you take me through the deal for uh Sesame AI?

Obviously they went viral but the founder had had a previous Andre backed company I think in Oculus. Um how did that come together and uh how is the company? Yeah.

And I and even your first interaction with the product that that blew you away because it in in many ways that that that their initial launch and and the website and and I think you guys had already you know completed the investment well before that but so many people that moment was truly eye opening for them where it was like okay I could imagine this feels like talking to you know a friend you know it felt it felt really real.

Yeah. Right. Uh yeah, Sesame is a fun story. Um because we started that company two years ago at a time when like everybody in this space was like look the the this this idea that that we're going to have a new computing interface um was completely seen as like a crazy a crazy idea.

There's there's two parts of of this ask me. one is the the hardware and we're building AI glasses over there.

M um and the second part is the actual companion which is the voice interface you guys have probably tried out the the voice AI right and when you put those together the idea is that it's a it's a system that has so much context about everything you're doing about your life that it becomes the primary interface to to to computing and I mean you know two years ago it it it I would say a lot of people had seen and watched like her the movie right yeah But the the the when I'd sit down and I'd describe to people that hey what you really need is is this beautiful marriage of hardware and software.

Um and that's what's going to come that's what's going to be the primary interface after smartphones. People would just look at you like you're crazy. Um and so I was like okay the the there's only two people I know who are crazy enough to believe this.

One was Brendan who had been an angel investor in my last company Ubiquity S. Brendan was the CEO of Oculus. Yeah. You know, one of the few people on earth who who' both built a hardware startup and sold it for billions of dollars. I think Oculus sold in total for north of two billion.

And Ankit who was co-founder and CTO of Ubiquiti 6, which was my last company. And Ankit was running the voice uh part part of the voice AI uh SDK and infra at Discord. You know, about 60% of all discord voice minutes uh daily minutes are spent in voice channels. Wow. Wow.

That's when we realized people don't realize it, but for most for many many many consumers, voice is actually a much more frequent interaction modality than like looking at a screen. Yeah. And so the idea was if we can combine those two, we might have a shot at building whatever comes next after smartphones.

So that that that was the idea, but the the key like insight I guess or bet was that it has to feel realistic. You're not like Siri just doesn't work. like you just nope out completely when you try to talk to Siri and you're like, "Okay, I'm talking to a robot. " Yeah.

But instead, if we could get the companion the the to to cross the uncanny valley, if we could get you to think about talking to Sesame, and you know, we have two companions right now. One is Maya, the other is Miles.

If you could get get you to think about using and talking to Maya and Miles as companions, not as robots, then that's when you'd really start using it in your daily life as an interface constantly to all the apps and services use. Does that make sense? Yeah, totally.

How do you in the context of Sesame what's been the thinking because obviously sounds like you've been on the board since day one or or you know I don't know the exact mechanics but how how has the company thought and and how do you think broadly across the portfolio around this a balance balance between you know being heads down building and needing to capture the attention of of uh you know the tech community the broader you know potential user base because I feel like it's this interesting dynamic right now where there's so much to build like in some ways companies should just be heads down and almost be silent, right?

And and this is like the SSI approach is to basically say like we're not you know maybe going to release anything that that we we consider less than um you know a really significant evolution. Um and but Sesame kind of like popped its head up and said hey look at us look what we're doing.

But then now it seems like, you know, it's taking more of the approach of, you know, being willing to fly under the radar for, you know, call it the next who who knows how many months. Um I'm I'm happy to report that it it doesn't get any easier the more money you raise.

I found like that on day one um we we didn't we hadn't raised any money, right? So we we're just three guys sitting in a room and talking about ideas. And um this I I I would say the heristic for most people who come from the software era is like hey ship fast, ship early, ship something you're embarrassed about, right?

And then iterate and AI is a little bit different.

Frontier is a little bit different because you often need a critical you you you need to have a a capability threshold that is sufficiently transformative enough that people would will put up with all the tons and tons of like friction that there still is to use AI today, right?

So you guys may have noticed but when you go to the Sesame site and interact with the demo, right? It's quite fast today. it. One of the things we really the team like spent a ton of time honing was the latency of voice responses. We had to make it feel like it was sub 200 millconds.

So it felt like you're having a conversation, right? Two years ago it wasn't there. It was is excruciatingly slow and we knew that nobody it people would miss the underlying personality of the model if we hadn't solved the systems problem of latency.

And I I find it comes down to this really delicate bal the the the overarching problem of like hey when do we ship when do we actually put this out in the world versus um being heads down comes down to this constant tussle between taste which is who at the company has such a strong opinion that the product experience is good enough and the and the and the ruthless sort of machine learning approach of running evals right the idea that you you build a model you test it in in an environment you benchmark it and you see how it does on that eval.

And if it hasn't improved that evaluation score, you know, you keep going. Mhm. And it it it it basically you you got to have like I've lost count of the number of hours we've spent debating that that tension, right?

And it's extremely uncomfortable for traditional software teams to do the the eval driven approach because you the the list of features you can kind of deterministically write out and say check check these are the things we need to ship build. Once you've got the vo, the MVP, we ship. Mhm.

That's not how that's not how AI works, right? AI research and post- training is is eval is eval driven, which is you have an intuition for what you what you need to do to improve the model, but you don't actually know until you run it through the an empirical test.

And so the answer comes down to how good are your evals. Mhm.

And so if I was to boil down what's what works for the best teams, it's that they have great product taste, but they also have great taste in eval what what is the right evaluation metric to build um for your team and then you you basically stay heads down until until you've you've unlocked both of those.

Does that make sense? Yeah. Yeah. Fantastic. Well, thank you so much for joining. Uh good luck with the rest of LP. Come back on again soon. Yeah. Yeah. We'd love to have you back and talk. Anytime. You know where to find me. Fantastic. We'll talk soon. Bye. Cheers. Cheers.

Uh, next up we have Martin Casado, general partner. He's on the board of Cursor, World Labs, DBT Labs, Kong, Fiverr, Idoggram, Ambient AI. This guy has a lot of board seats. Brain Trust, Coactive, Netlifi. Um, uh, he leads the firm's $ 1. 25 billion infrastructure practice and has built quite the AI portfolio.

So, we're excited to have Martin on the show. Uh, discuss open source AI, his views there,