Erik Bernhardsson on Creating Tools That Make AI Feel Effortless

0:06Eric Berhardsen, founder and CEO of Modal. Modal developed a serverless cloud platform tailored for AI, machine learning, and data applications. And before that, Eric worked at Better.com and Spotify, where he led Spotify's machine learning efforts and built the Recommender system. Well, Eric, thanks so much for joining me today on NoPriors. Yeah, thanks. It's great to be here. So if I remember correctly, you worked at Spotify and helped build out their ML team and Recommender system, and then we're also at Better.com. What inspired you to start Modal, and what problem were you hoping to solve?

0:33Yeah, I started at Spotify a long time ago, 2008, and I spent seven years there. And yeah, I built a music recommendation system And back then there was like nothing really in terms of data infrastructure. Hadoop was like the most modern thing. And so I spent a lot of time building a lot of infrastructure. In particular, I built a workflow schedule called Luigi that basically no one uses today. I built a vector database called Illinois that, you know, for a brief period people used, but no one really uses today. So I spent a lot of time building a lot of that stuff. And then later at Better, I was a CTO and thinking a lot about like developer productivity and stuff.

1:07And then during the pandemic, I took some time off and started hacking on stuff. And I realized I always wanted to build basically a better infrastructure for these types of things, like data, AI, machine learning. So pretty quickly I realized, like, this is what I wanted to do. And that was sort of the genesis of modal. That's cool. How did that approach evolve? Or what are the main areas of the company focuses on today? So I started looking into, first of all, just like, what are the challenges with data, AI, machine learning infrastructure? and started thinking about from a developer productivity point of view, what's a tool I want to have?

1:38And I realized a big sort of challenge is working with the cloud is arguably kind of annoying. As much as I love the cloud for the power that it gives me, and I've used the cloud since way back 2009 or so, it's actually pretty frustrating to work with. And so in my head, I had this idea of what if you make cloud development feel almost as good as local development, like how does this fast feedback loops? And so I started thinking about how do we build that I didn't realize pretty quickly, like, well, actually, we can't really use Docker and Kubernetes. So we're going to have to throw that out and probably going to have to build our own file system, which we did pretty early.

2:12And I built our own scheduler and built our own container runtime. So that was like basically the two first years of Modo. It's just like laying all that like foundational infrastructure layer in place. Yeah. And then in terms of the things that you offer today for your customers, what are the main services or products? server? Yeah, so we're infrastructure as a service, which means like on one side, we run a very big compute pool, like thousands of GPUs and CPUs, and we make it very easy to get, you know, if you need 100 GPUs, we can typically get you that within seconds. So sort of one big multi-tenant pool, which means like capacity planning is something we kind of take, you know, it's something we solve for customers.

2:51They don't really need to think about reservations. We always provide a lot of on-demand GPUs. On the other side, there's a Python SDK that makes it very easy to build applications. So the idea is you write code, basically functions in Python, and then we take those functions, turn them into serverless functions in the cloud. We handle all the containerization and all the infrastructure stuff, so you don't have to think about all the Kubernetes and Docker and stuff. And the real killer app, as it turns out, we started this company pre-gen AI, but as it turns out, the main thing that really started driving all the traction was when Stable Diffusion came out.

3:22And a bunch of people came to us and were like, hey, actually, this looks kind of cool. Like you have GPU access. It's very easy to, you know, you have to think about, you know, spinning up machines and provisioning them. So that was like our first sort of killer app was like just doing Gen.AI in a serverless way with the focus of diffusion models. Like now we actually, we have a lot more of different modalities. Like a lot of usage is still like text to image, but we also see a lot of audio and music. So one example of a customer, I think is super cool, building really amazing stuff is Suno. which does AI-generated music.

3:53So they run all their inference on modal, very large scale. There's a lot of customers like that sort of dealing with like, you know, building cool Gen AI models. In particular, I would say in the model is like audio, video, image, and music, stuff like that. That's cool. And I think Suna's using all like Transformer Backbone now for stuff, right? Versus a diffusion model-based thing. I think it's a combination of both. I'm not sure. Yeah, I think they talk about it publicly. That's the only reason I mention it. You wrote in October this post, I think it was called The Future of AI Needs More Flexible GPU Capacity.

4:24And in general, what I've heard in the industry is that a lot of ways that people use GPU is reasonably wasteful. And so I'm a little bit curious about your view on flexibility around GPU use, how much is actually used versus wasted, how much optimization is left, you know, even just with existing types of GPUs that people are using today. Yeah, GPUs are expensive, right? And I think it's sort of kind of as like a paradox. It's like means that cloud, you know, a lot of the cloud capacity is like, you know, the only way to get it is to sign long term commitments, which I think for a lot of startups is really not the right model for how things should be.

4:58Like, I think the amazing thing about the cloud was always to me that like you have on demand access to like whatever many CPUs you need. But for GPUs, like the main way to get access has been over the last few years due to the scarcity has been to sign long term contracts. And I think fundamentally, that's just not how startups should do it. And that kind of get it has been sort of supply demand issues. But just looking at the CPU market, the fact that you have instant access to thousands of CPUs if you need it, my vision has always been this should be the same thing for GPUs. And that means, especially as we shift more to inference, I think for training, it's been sort of less of an issue because you can sort of just make use of the training resources you need.

5:39but for inference especially you don't even know how much you need right like in advance like it's very volatile uh and it's a big challenge that that we solve for a lot of customers is we we're fully usage based so when you run things on modal we charge you only for the time the container is actually running uh and uh and that's that's a massive hassle for for customer traditions like doing the capacity planning and thinking how many gpus and then and then having the issue like either you're over provision and you're paying for a lot of idle capacity or you're under-provisioned. And then you have, you know, when you run into the capacity shortage, you have degradation in service.

6:12And so whereas with modal, we can handle these very bursty, very unpredictable workloads really well because we basically take all these user workloads and just run a big pool of thousands of GPUs across many different customers. Yeah. One of the things that always struck me about training is to your point, you kind of spin up a giant cluster. You run a huge supercomputer, right? And then you run it for months in some cases. and then your output is a file. And that's literally what you've generated. It's kind of insane if you think about it. And that file in some sense is a representation of the entire internet or some corpus of human knowledge or whatever.

6:47And then to your point with inference, you need a bit more flexibility in terms of spinning things up and down or alternatively, if you're doing shorter training runs or certain aspects of post-training, you may need more flexible capacity to deal with. Totally. And that's something we're really interested in right now. Like traditionally, most of modal has always been inference. Like that's been our main use case, but we're really interested also in training. So in particular, like probably focus more on these like shorter, like very bursty sort of experimental training runs, not the like very big training runs, because I think that's a very different market.

7:13So that's like a very interesting thing. How do you think about meeting people's end to end needs? So I know that there's a lot of other things that people do. There's, you know, a lot of people are using RAG to basically augment what they're doing. or there's a variety of different things that people are now doing at time of inference in terms of using compute to take different approaches there. I'm a little bit curious how you think about the end-to-end stack of things that could be provided as infrastructure and where modal focuses or wants to focus. Yeah, totally. I mean, our goal has always been to build a platform and cover the end-to-end use case.

7:50It just turned out that inference was, we were well-positioned to focus on that as our first killer app. But my end goal has always been to make engineers more productive and focus on what I think was like the high code side of ML. I think our target audience tends to be more like sort of traditional ML engineers, like people building their own models. But there's many different aspects of that. There's like the data pre-processing, then there's the training, and then there's the inference. And that is actually probably like even more things, right? Like, you know, having feedback loops where you gather data and like, you know, online ranking models and all these things.

8:21And so my goal for Modal has always been to cover all of that stuff. And so it's interesting. You see a lot of customers now, we don't have a training product, but a lot of customers use modal for batch preprocessing. So they use modal to, maybe they're training a video model. So maybe they have petabytes of video. So then they use modal actually, maybe with GPUs even, to do feature extraction. And then they train it elsewhere. And then they come back to modal for the inference. So for us to do training makes a lot of sense. And in general, I think it makes a lot of sense to sort of build a platform where you can handle the entire sort of machine learning lifecycle end-to-end and many other things related to that.

8:57Also the data pipelines and nightly batch jobs and all these things. Yeah. I mean, what you describe is a pretty broad platform-based approach. I think there's a handful of companies who are sort of in your general space or market. How do you feel that modal differentiates from them? I think, first of all, we're cloud native. Like, we're just like cloud maximalists. like we went all in and said like basically we're going to build a multi-tenant platform that runs everyone's compute and the benefits of that are like very tremendous because like we could just do capacity management much better uh and that's one of the ways we can offer like instantaneous access to hundreds of gpus if you need to like you can do these like very bursty things so we just give you lots of gpus right i think the other benefit or the other sort of differentiation is very general purpose uh we focus on sort of what i think as i mentioned like high code like we run custom code in our containers, in our infrastructure, which is a harder problem.

9:49Like containerization and running user code in a safe way is a hard problem. And then dealing with container call start. And like I mentioned, we have to build our own scheduler. We have to build our own container runtime and our own file system to boot containers very quickly. And I think so, unlike many other vendors, like they're only focused on, say, inference or maybe only LMs. Our approach has always been to build a very general purpose platform. And sort of, you know, in the long run, And I hope to sort of that sort of manifestation will be more clear because I think there's many other products we can build on top of this now that we have the compute layer sort of kind of becoming more and more mature.

10:22When I talk to large enterprises about how they're thinking about adoption of AI, many of them already have their data on Azure or GCP or AWS. They're running their application on it. They've bought credits in the marketplace. They want to spend resident. They've already gone through security reviews. You know, they've kind of done a lot and they worry about things like latency or pings out to other third party services versus just running on their own existing cloud provider or their hyperscaler that they work with or set of hyperscalers. You know, many of them actually work across multiple. How do you think about that in the context of modal in terms of your own compute versus hyperscalers versus, you know, the ability to run anywhere?

10:59Yeah, totally. And of course, there's also a sort of security compliance aspect of this. I think it is a challenge. I look back at when the cloud came. And I remember back in 2008, 2009, and the cloud came. And my first reaction was like, how the hell, why would anyone put their computer in someone else's computer and run that? And I think, to me, that was just insane. Why would anyone do that? But over the next couple of years, I was like, actually, it kind of makes a lot of sense. And I think now, even among enterprise companies, there's a sort of recognition that, yeah, actually, probably our computer is more safe in the big hyperscalers.

11:33And in a similar vein, I remember talking to Snowflake back in, say, 2012 or something like that. And they had a sort of similar approach where they basically said, we're going to run databases in the cloud and it's not going to be in your environment, or maybe in your environment, but we're in infrastructure as a service. And I thought that was nuts. And then obviously, I think Snowflake now is a very large publicly traded company. I think they showed that infrastructure as a service makes a lot of sense. And so I think there is a little bit of resistance to adopting this multi-tenant model. But I think when you look at security and adoption of cloud, I think we have a lot of tailwinds blowing in our direction.

12:09I think security is moving away from sort of a network layer into an application layer. I think bandwidth costs are coming down. I think there's a lot of tricks you can do to minimize bandwidth transfer costs. You can store data in like R2, for instance, which has zero egress fees. It's something that I think is realistically going to mean we're going to have to push a lot. But I think there's so many benefits of this multi-tenant model in terms of capacity management that to me, it is very clearly like a big part of the future of AI is like running a big pool of compute and slicing it very dynamically.

12:40You mentioned earlier that one of the things that really caused early adoption of modal was stable diffusion and sort of these open source models around image gen. Are there any open source projects or models that you're seeing be very popular in recent days or in the last couple of months that have really started taking off? That's a good question. I think, if anything, it's actually been a little bit of a shift towards more like proprietary models. But like proprietary open source models, I guess. So like Flux, I think most recently has been, you know, a model that's getting a lot of attention.

13:14I'm personally very interested in like audio. I think Unreal is like very underexplored. I think there's a lot of opportunity for open source models in that space. But I don't think we've seen anything really cool yet. What else do you think is missing in the world today in terms of AI infrastructure or infrastructure as a service? So I'm very biased, but I think modal is missing. Like basically a way for engineers to take code and run it. And look, I'm very bullish on like, you know, code and like people wanting to write code and building stuff themselves. I think outside of sort of LM space, which is like a very kind of different world, in my opinion, I think there's always going to be a lot of applications where people want to train their own models.

13:51They want to run their own models or at least like run other models, but have like very custom workflows. And I just don't think there's been a great way to do that. It's like pretty painful to do that. And so I think that's pretty exciting. I think on the storage side, there's some other really exciting stuff. We haven't really touched storage at Modal. We focus very much on compute. So I'm personally very interested in the vector database. How's that going to evolve? I don't think anyone really knows. I'm pretty interested in more efficient storage on training data. I'm also very interested in, I guess another thing I'm very fascinated by right now is training workloads.

14:26in order to train large models efficiently, you've had to really spend a lot of money and time setting up the networking. So one of the things I'm really excited about is what if you don't, you know, what if we can make training less bandwidth hungry? Because I think that would actually change a lot of the infrastructure around training where you can now like kind of tie together a lot of GPUs in different data centers and not have to, you know, have this like very large data centers with like, you know, Infiniband and stuff. So that's like another sort of infrastructure thing. I'm looking forward to see more development on.

14:59How important, so there's sometimes been a little bit of debate around vector DBs. And you mentioned that you actually built one when you were at Spotify. I think Spotify today hit$100 billion in market cap. I think it's one of the first European technology companies to get there, which is pretty cool. So a lot of folks I know may use one of the existing vector DBs or in some cases are just using Postgres with PG Vector, right? How do you think about the need for vector databases as sort of standalone pieces of infrastructure versus just, you know, adopting Postgres versus doing something else?

15:33Yeah, I feel like everyone's debating that. I don't know necessarily. Like, I think there's a lot of there's a case to be made that, you know, you can just stick everything into relational database and you're fine. To me, like the bigger question is like in the long run, like, you know, if you think about like what's like an AI native data storage solution? Like, I don't even know if it's like necessarily has the same form factors, the same interface as a database. So that's actually a bigger question that I'm more excited about. It's like, I think people look at like vector databases and like, you know, whether it's relational or not, they sort of shoehorn it into this like, you know, sort of old school model of like you put data, you get data back.

16:09But I don't know. I think there's like a lot of room to sort of rethink that in the age of AI and have very different like, you know, interaction models with that data. I know that sounds a little fluffy. Yeah, it's super interesting. Could you say more on that? I mean, like one thing I think a lot about is like maybe the database itself be like the embedding engine, right? Like instead of like you put a vector in, you know, you search by that vector. I think there's a lot of, you know, the more native, like AI native storage solution would be you put text in, you put, you know, video in, you put image in, and then you can search by that.

16:38Like to me, that would be like a more sort of native, AI native sort of storage solution. So that's like one line of thought that I've had is like, maybe we're just like so early to this that like, I think it's going to take five, 10 years for it to really, for it to shake out. Yeah, that's really cool. I guess one other thing that you mentioned was more people seem to be training their own models, at least in a lot of the areas that Modal works with. Do you think there's any heuristic that people should follow in terms of when to train their own model versus use something off the shelf? I think eventually, for any company where model quality really matters, unless you kind of train your own model in the end, I feel like it's going to be hard to sort of defend the fact that you have a better solution.

17:20Because otherwise, what's your moat? If you don't have your own model, you need to find a moat somewhere else in the stack. And that might be possible to find. It might be somewhere else for a lot of companies. But I think at least if you have your own model and that model clearly is better than anyone else, then that sort of inherently is a mode in itself. I think it's more clear outside of the LM space when people are building audio, video, image models. I think if that is your core focus, it's very clear to me, you kind of have to train your own models in that case. Yeah. If I remember correctly, you're an IOI gold medalist.

17:56Yeah, that's right. Obviously, you think a lot about code and coding. And how do you think that changes with AI over time? Or do you have any contrarian predictions on what happens there? I don't know if this is contrarian, but I actually think that this is just one out of many improvements in developer productivity. And you look back at whatever, like compilers was originally a tool that made developers more productive. And then higher level programming languages and databases and cloud and all these things. And so I actually don't know if AI is different than any of those changes in hindsight.

18:32And by the way, every time that's happened, it turns out there's so much latent demand for software that actually the number of software engineers goes up. So I feel like you look back at the last 40 years of software development, every decade, engineers get 10 times more productive due to better frameworks or better tooling or whatever. And it turns out, actually, that just unlocks more latent demand for software engineers. So I'm very bullish on software engineers. I think it would take a lot to sort of destroy that demand. I think people look at a lot of like AI as like it kind of fixed something.

19:02But in my opinion, it's like, no, it's just going to unlock more latent demand for more things. So I'm very bullish on software engineering. And then I guess the other field that you touched a long time ago was, I think you won a Swedish physics competition in high school. And I'm curious if you followed any of the physics-based AI models or some of the simulation. Like that's an area that strikes me as very interesting. and the way you think about the models for it are different and yeah i i i did win uh the spritish high school physics competition i was a total math lead nerd when i was uh you know my teenagers okay yeah i think it's a really fascinating area right now like it's one of those areas that seems um like there's some real reinvention needed and not as many people working on it so it's one of the areas i'm kind of excited about just in terms of there's there's lots and lots of different applications that you could start to come up with relative to it yeah i think i mean like physics in It's like, you know, it looked like the golden era of physics, like the 20s and 30s and 40s.

19:59I kind of feel like it's like it hasn't really evolved much to the field. So I don't know. I would love for you to be right that there's like a resurgence of, you know, new physics based models. Yeah, I don't know if it would necessarily help in the short run with basic research. I think it just helps with simulation. It kind of feels like physics as a field really kind of doubled down on sort of the Ed Witten path of physics and maybe got a little bit lost there or something. I'm not sure. Are you talking about more like material, like doing more like compute-based methods? It's kind of like Ancest or other companies where you simulate an airplane wing, you simulate load-bearing in a...

20:30Oh, I see. So like HPC. I mean, it's always existed, right? Especially in like oil and gas and stuff like that. But it's a lot of kind of small, bespoke, kind of fine-tuned or hand-tuned models for specific things versus... Yeah. I mean, meteorology is like something I actually think like deep learning should like change, right? Like it sort of makes a lot of sense. Like, you know, deep learning should be very good at like, you know, predicting, you know, turbulence and things like that. Like, because turbulence is actually very hard to solve with traditional physics models. Right. And so deep learning should, in theory, I kind of feel like makes a lot of sense.

21:03Yeah, I think there's been a couple of papers on that out of NVIDIA. And then I think Google has a team that's worked on it. And so there's a couple different sort of weather simulation teams that have started to publish some pretty interesting stuff, it seems. Yeah. Yeah. I mean, I would also point to like an adjacent area. like biotech, I think it's like been, you know, computational methods have been enormously successful, right? Like if you look at protein folding in particular, but also other things like sequence alignment and things like that. And that's actually a field where we start to see a lot more usage as modal as well.

21:32There's a lot of, I feel like there's like a kind of a resurgence of computational biology. It's really exciting. That's really neat. Yeah. Are there specific use cases that you see people engage with most across your customer base relative to the sciences or? There's a lot. I'm not a buyer person. So this is kind of superficial, just kind of looking at our customers. But like one thing I've seen a lot is actually medical imaging. Because my understanding is like with modern methods, like you can do like very automated, like, you know, get like millions of, you know, experiments and do, you know, automated electron microscope imaging of that.

22:07And so we've actually seen quite a lot of customers like use modal for like then processing and doing computer vision on those images. which is kind of cool. It's really cool. Is there any area that you're most excited about from a human impact perspective for some of these models? You know, with my background at Spotify, like I think Zuno is like, to me, very exciting thing. I think it's still like very early sort of AI generating music. You can still hear that it's like not right. It's sort of uncanny value a little bit. But like Zuno is like every generation of their model is like getting better and better.

22:37And first of all, like music in itself tends to be like sort of always like one of the first areas where you see real impact of new technologies, whether Spotify or iTunes or Piracy or all these things, or gramophones going back, right? So I always think music is an exciting area in that sense. It always shows the opportunity of new technologies. And I also think Suna is fundamentally something you couldn't have done before Gen AI. So that, to me, is really exciting. It's really pushing the frontier, enabling a completely new product that there's no way Suna could have existed five years ago. So that's cool.