The Infra Pod - How will the world run on GPUs? (w/ Stephen Balaban from Lambda Labs)

Episode Date: July 17, 2023

Ian and I sat down with Stephen Balaban (CEO / founder from Lambda labs), to talk about how Stephen started Lambda Labs inspired by Japanese Anime, into a Computer Vision API, and finally a GPU hosted cloud. We also introduced a new section called the "Spicy Future", and we got spicy :) Come listen to how we noodled on how we will get closer to AGI and what will happen when it does.

Thank you. And Ian and I, in this podcast, will be talking about amazing speakers like Steven, about their companies, what they've done in the past, but we're going to spend a lot of time also talking about the future. Spicy hot takes, spicy futures, this is what we're all here about. So let's go for this. I'm Tim from Essence VC, and I'll let Ian introduce himself. And I'm Ian, running some engineering stuff at Snyk and super excited to be joined today by Stephen at Lambda Labs. Stephen, welcome to our little shindig, our little podcast. How are you doing today? I'm doing really well. Thank you so much for having me. We couldn't be more excited. Stephen, give us a little introduction to yourself. How did Lambda Labs get started?

And also congratulations on recently raising your Series B. I saw that news come through as well. So I'd love to hear about your journey. Yeah, sure. Happy to give an intro here. So before Lambda, I was the first employee at a company called Perceptio, and we were training neural networks that ran locally on the iPhone. The iPhone 4S was one of the first iPhones with a GPU specifically, an imagination GPU. And what we were doing was running neural networks locally on the iPhone. And so doing face and image recognition. In 2014, that company got acquired by Apple and integrated into iOS. And so if you are familiar with some of this stuff, whether it's swipe up on an image on your photo roll, there's image and face recognition that's integrated into that. And that's some of

the stuff that got integrated into iOS. The founders of that company, one of them is an investor in Lambda Labs now. And my background broadly is just that of a deep learning engineer. You spent this time at this company that got acquired by Apple. What made you start Lambda Lab? What was the insight that you were like, I've got to go start a company, right? We all have different journeys and how we get there. I'm sure Tim and I could share our own, but I'm curious, what got you started? So one of the interesting things is that you always ask this question to founders, and there's two different answers. There's the PR answer, which is, let's make this story comprehensible and understandable for everyone.

And then there's the real nuts and bolts story. I'll try and give something that's comprehensible and understandable, but to be as truthful as possible with all the twists and turns. Fundamentally, the company actually, interestingly enough, started out as a company that was doing face recognition powered contact book. I had this future vision where you'd have a pair of sort of like AR glasses. I'm a big anime fan, and there's a great anime called Denocoil, A Circle of Children. And it talks about sort of the first generation of children growing up with augmented reality glasses and really what i wanted to build was sort of if you've ever played world of warcraft in world of warcraft

there's the characters have their character names and then the guild underneath their name and i wanted to build that in real life that's sort of like how i got down this path of building a face recognition contact book, because that's like maybe a more realistic version before AR glasses are a thing. I mean, that was probably about 20 years too early, let's say. But the really sort of fortuitous turn of luck was that I started Lambda, it was really getting into face recognition, right as, you know, Alex Krzyzewski and Ilya Sutskobar and Jeff Hinton published the original ImageNet paper, where they just blew away the entire state of the art with two GPUs underneath their desks.

A lot of people don't really recognize the true sort of shot heard around the world level importance of that paper, but it really totally transformed my thinking. And at that point, I literally immediately became a deep learning convert. A lot of people may not remember this, but there used to be sort of the University of Montreal deep learning tutorials where it was all Theano tutorials and they had Boltzmann machines and denoising auto encoders. And you could look into the weights that were learned there and you can kind of see some of the features that were learned. And I just remember looking into that and thinking, wow, this is so much better than

Gabor filters and so much better than local binary patterns and so much better than Fisher faces, which was the sort of the current state of the art at the time for traditional computer vision. And that's really how we got started as a company. And so we actually launched our first product was this face recognition contact book, Facebook acquired face.com. If you remember face.com, which was a face recognition API. And we thought, hey, there's an opportunity for us to take our homegrown face recognition software and make an API for it. And so that's actually one of Lambda's first successful products. Heads up, that face recognition contact book, it went nowhere, basically.

But the API was the thing that first started giving us some user attraction. And there's a bunch of pivots between then and us selling our first workstation for people doing deep learning. But maybe we can dive into that later in the podcast. But long story short is we've always been building infrastructure for ourselves. We got really good at running and managing very large GPU clusters and eventually pivoted into providing both workstation servers as well as now a cloud service where you can spin up a GPU for, you know, A100 for $1.10 an hour, which is the cheapest in the world. We are literally the cheapest in the world,

which is pretty amazing given that our competitors are the biggest companies in the world. Yeah, that's totally incredible. And I have so many questions to dive into like how you achieved that outcome, but I just want to scroll back to a second, like the world of Warcraft idea of the heads up display to know that like, Oh, that's Tim over there. And they're level three. I totally love it played way too much while myself. So you had this long journey. Along the way, I think I heard you say, we're really focused on basically

building the best training infrastructure for our own pursuit around your training infrastructure and the work that you do with your cloud product. What's the driving like insight there? Is there some driving North Star that you focus on to say, hey, this is the thing that we design for when we build products, the things that we think about that need to be improved? Because I'm really curious to understand, is there some key insight there? And then also how that broader,

we get into talking about some hot spicy takes, as Tim talked about, but curious to learn more about how you think about empowering people to build this deep learning technology with the things you have and what you've learned about workflow and cost, performance, and all those aspects. I see what you're saying in terms of

is there a guiding principle behind all the sort of things that we've done? And so here's the long and short of that. There's a huge amount of qualitative decision-making that you make when you're designing a product. And when you're building for yourself, that feedback loop is very tight because you look at it and you say, this is trash.

Or you look at it and say, oh, finally, it solves my problem. But I think that maybe you're asking more like, well, is there a quantitative thing that's some sort of insight? And I do think it's really important to have sort of first principles thinking when you're running a company. The first principles are this early forms of AGI are here.

I think if you were to ask anyone from the 1980s, whether this, whether chat GPT was considered AGI, they would say, yeah, absolutely. That's AGI. And early forms of AGI are here. It's clear that there is a direct relationship between sort of the amount of compute used to train them and how good these models are. And access to more and more compute results in better and better models. From that perspective, if you think about now from a first principles way of thinking is that, well, if to make this AGI better, we need to make compute more available. Well, to make compute more available, you want to start saying, well, it's flops per dollar, basically, because availability is both a combination of how many flops are you going to be able to provide, but also how many incremental flops you're going to be able to provide per dollar input. So I believe that flops per dollar is,

if there's any one metric that we care about at Lambda, it's that. But noting also that the dollar of that is the really truly qualitative aspect there, which is to say that obviously there's amortized capital expenditure. Basically, if you're depreciating that capital expenditure over time, you know, there's that total capital expenditure cost. There's cost per kilowatt hour of electricity. But there's also some other costs like developer time, ease of use, simplicity. That's where you start getting into the qualitative aspects about a platform. We're really still going from zero to one in many ways on this most recent explosion in new architectures like the transformer and you've got deep neural nets. Like we're still just seeing the early beginning.

So I'm curious, like when you think about product decisions that move the needle through your journey, is it mostly indexed and focusing on like, how do I just make more compute available? Or is it also heavily indexing on how do I make that compute available and easy to access? Is it 50-50? Is it 30-70? I'm kind of curious how you think about the strategy behind what you're doing. So I think it's really important to just ideally be working towards something where your first principles aren't changing that much, right? And so one of my favorite quotes, and I guess I can pull up my Twitter here, is this quote from Henry Ford. And he says, basically, and success in manufacture is based solely upon an ability to serve that customer or consumer to his liking. He may be served by quality or he may be served by price. He is best

served by the highest quality at the lowest price. And any man who can give the consumer the highest quality at the lowest price is bound to be a leader in business, whatever the kind of an article he makes. There's no getting away from this. That is a quote from 1922. So it's 100 years old, as they say online. It's about as lindy as you can get. And when you look at other entrepreneurs and how they approach business, this is not too different. When you look at early Jeff Bezos talking about what people care about, people just want the thing as fast as physically possible. And so I'd say that, you know, we spend our time worrying about cost, both from a actual like capital expenditure reduction perspective and optimization perspective and designing a system that is like perfectly in line with what the

customer needs. And then a lot of the cost reduction comes into like, how can you make it so that somebody instead of committing to a three-year contract, maybe can you right-size a compute job so it's like, hey, you're trading a large language model. So here's the exact amount of compute you need to train your large language model, for example. It's a mixture of all these things. So I guess maybe to ask additional question on top of this, your website, the about page, you have the timeline. It kind of talks about your evolution towards the facial recognition API, towards the workstation, to enterprise products. Definitely one thing that jumps out is a lot of enterprises tend to jump on, like we got adopted by Apple and a bunch of these companies.

Maybe talk about what was that inflection point like? What got their interest to use your products? Was it the timing? Was it there's a lot of need that you start your infrastructure first and make it When I first started Lambda, and I'm in my Chinatown apartment, and I read the AlexNet paper, at that point in time, you really had to go out and convince people, hey, look, neural networks are cool again. It wasn't just something that died during the AI winter of the 1980s that everyone had been taught about in school. And you had actually legitimately convinced people that deep learning was not just some sort of mirage. And you talk about customer adoption, you get a certain customer adopting your thing. Basically, it was not going to be possible for us to build this business until probably around 2015, 2016, when TensorFlow came out. Maybe a little bit before that, when Cafe came out,

if you guys remember, Cafe was a slightly more user-friendly sort of inferencing type of thing that Yangqing Jia came out with. And before that, it was like Theano, which was like a very complicated programming model that TensorFlow actually ended up adopting and then sort of realizing they made the same mistake that Theano had made with this sort of graph programming approach. And then there was Torch, like Lua Torch, Lua Torch. Those are the two frameworks that people used. AlexNet, the original one was actually just written in raw CUDA, if I remember. It was just a CUDA repository on Google Code, if that really dates it. Google Code was a thing still. And it was sort of like this. We started selling workstations. And then literally within the first 10 customers, it was like Fortune 500 companies. Funnily enough, we actually started selling the workstations on Amazon.com. Second or third sale was to an Apple engineer in Cupertino.

And so it was immediate. It was immediate, unbelievable customer that you land like that. But you couldn't have done that in 2012, basically, if that makes sense. Yeah, that's awesome. So we really want to get to the next section, which we call the spicy futures. Spicy futures. Spicy futures. What we love to get out of you is you live in the future, right? You started Lambda Labs and you built this infrastructure and you made it available. That was one of the first back then. Now we see the crazy amount of HEI, all the LEMs.

AI is everywhere now.. AI is everywhere now. GPU demand is everywhere now. But what is the future that you believe in that I don't think many, many people have believed in yet? Being able to predict the future is a super useful skill in business, obviously. But the frameworks that I use, it's just like sort of some crazy dream that you come up with. For people who are sort of really intuitive people, you just sort of like have ideas or whatever. But a few of the things I think are like good truisms. One, the future is already here. It's just not evenly distributed. William Gibson. That's definitely a really good one. Two, if you want to predict the future, invent it. That's an Alan Kay quote, right?

And sometimes you're really, really wrong. For example, that sort of future of augmented reality glasses from Denocoil and whatever. We're not there yet. I don't have my little AI augmented reality Pikachu that's sitting next to me right now, even though I want that. And so sometimes you're really wrong. And that's sitting next to me right now, even though I want that. And so sometimes you're really wrong and that's fine too. The other one is like, finally, I think it's also important to reference science fiction. I think science fiction is like just people have spent many hours thinking about the repercussions of this stuff.

So those are the frameworks. And then I think this is one of the futures. The future is 20 years from now. And I kind of look at where we are today as being a little bit like being in 1979 in the computer, personal computing revolution. And one of the quotes they used to pass around at Apple or one of the sort of missions actually they had at Apple was, one person, one computer. One person, one computer. one person, one computer.

And that seems like really obvious today, right? I mean, everyone has more than one computer even. But at the time, it was a very, very avant-garde statement. One person, one computer. There was very hashtag super smart people at IBM who thought there was like no market for personal computers. And so I kind of rephrased that a little bit as one person, one GPU instance. Because I really think that 20, 30 years from now, you're going to look at the world and everybody in the United States every day will be interacting with the neural network at

home, at work, while they're being entertained. And the sheer amount of compute that the world needs is like 100 or 1000 times greater than it is today. When I look at numbers like that, what's really clear is we're just one small part of that equation at Lambda, for example. It's obviously much bigger than us in the same way that Apple was only one small part of the one person, one computer equation. And so that's sort of like the high level sort of numbers on it. It's like we were one person, one GPU instance. But let's say some of the qualitative predictions for the day to day. I think that, you know, flash forward 30 years from now, you're probably going to see, for example, using this sort of the future is already here.

It's just not evenly distributed yet. Throughout an organization, there's going to be sort of this hierarchy of LLMs that have different levels of visibility in terms of the inner and outer chatter of a company. And with ever-expanding context windows, today we're at, let's say, 64,000, 128,000 characters. As those context windows get larger and larger and larger, I think that, 64,000, 128,000 characters. As those context windows get larger and larger and larger, I think that one, the size of the network and the expanding context window sort of is going to start to obviate the need for too much fine tuning. It's really just going to be more about shoving more contacts into the context window. But then you're going to start to see like

at the very top of the organization, there's going to be this sort of corporate overmind where it has all of the Slack or Microsoft Teams chatter in and out, all of the video chatter in and out, all of the emails in and out of a company. And the CEO is going to be able to have complete visibility. A low-level manager may not have visibility into the emails of the CEO or the executive staff, but the CEO might have full, complete visibility. And they're going to be able to do queries that are very powerful and would historically only have been able to be accomplished by many, many layers of management, summarizing sort of middle managers whims of like, hey,

how effective is this person at their work? And you're going to be able to do a query like, hey, who's the most knowledgeable person about InfiniBand at the company that I can talk to? Who are the top 10 highest engaged, most performing salespeople who are being sort of, if you will, under-recognized, for example, right? You can imagine these very complicated queries that are based off the entire streaming context window of all the information the company's producing. And who are the top 10 most hated people at the company? Fire them. It's that kind of interaction, I think, where you're going to be giving executives and managers and people within the company visibility into who they should talk to, who can help them on projects. And then eventually, I think, again, 20, 30 years out, this is going to be even more weird. We'll

be like, some of your colleagues will be virtual people. Some of your colleagues are going to be non-humans that are interacting with each other. And maybe there's going to be non-human middle managers. And there'll be some sort of weird symbiosis between human individuals who are employees at a company and then AGI agents who are employees at a company. I don't think it's that weird to think 20, 30 years out that there's going to be people who, like in the movie Her, are going to fall in love with their AGI agent. And the government's going to want to tax income for AGI agents. And there's going to be a AGI agents rights movement. And there's going to be

a set of people who are going to say, well, AGI agents, are they people? Do they deserve rights? And there's going to be another employees of a company will be AGIs. And I can actually think today of things I've been looking at recently, specifically in the developer landscape, where we're starting to see semi-autonomous feedback loops. It's human still in the loop, still semi-in control, but expediting the development loop. So you can see how quickly you'll end up with these AGI agents that are actually like these full-blown things. Humans or, well, it's human-like, right? They're replacements for that human function that used to be a person. It's now a robot effectively, like an AGI. Now, maybe a robot in physical form, but in conceptual form. Another great example of that would be like how McDonald's replacing the worker at the drive-thru, right? No longer takes the order. I'm curious, thinking about that future,

what do we need to get there? You brought up we need a massive amount of compute. What are the other barriers to achieving that outcome? Is there some fundamental thing we're missing? Or is the only thing we're missing compute power? Are we missing key architectural components and sort of the same with the transformer unlocked? Yeah, absolutely. I mean, you know, this is the thing.

It's like, you know, predicting the exact thing. Well, if I could predict the exact thing, then I would be off going and doing that as fast as I possibly could. Right. So we could unlock the next thing. And then this is sort of like where the limits are for predictive capabilities however i would say that what what do we need it was very clear that we need a lot more compute in the world thankfully because of the success of chat gpt and chat gpt probably being the fastest growing product in the history of capitalism well that's gotten the attention of, let's say, a handful of companies in the world

that are looking to do large AI infrastructure deployments, whether it's Amazon, Google, Microsoft, Oracle, and Lambda. We're doing a lot to deploy a lot more AI infrastructure that the world very, very clearly needs. So we need more compute deployed today. I think that we're going to need plenty more advances as far as like, you know, all the next generation of architecture, et cetera. And having followed this industry for a little over a decade now, right? 11 years since 2012, when I started Lambda in April, 2012, it's like been this series of S curves. And it's like, the overall industry is going to be this one big S curve, you know, this sort of sigmoid type of thing where there's a exponential growth phase,

a linear phase, and then a sort of logarithmic phase at the top end of it, right? And the overall thing feels like that. And I'd say that right now with LLMs and the transformer model sort of getting brought out to scale, it feels like we've hit that first inflection point or well, well into the scary accelerating part of the initial part of this sort of ConvNet, like AlexNet came out in 2012. And I just remember thinking very distinctively 2015, 2016, like VGG, ResNet era type of thing. I was like, we've kind of hit a plateau here. Is this all AI is going to be is recognizing defects on a manufacturing production line or analyzing trains to see if there's stowaways on the train or, you know, dystopian camera systems that are like analyzing people's movement in China or in the United States or in the UK. And it's like, okay, well, great. Is that, is that the end here? And it definitely felt like we hit that sort of saturation curve on the

ComvNet side of things. I think we're still actually well into the deployment phase, but let's just say the acceleration really slowed down, it felt like. Maybe we got into the linear thing. But the LLM stuff is that next thing. And so we've seen a couple of different phases like that. I have, and I'm sure people who have been in the industry for longer have too. And so there's definitely going to need to be a couple more breakthroughs that keep us accelerating forward. But to answer your original question, what do we need? We need more compute.

I think that's definitely true. I have to think more about what other things are definitely true. We need more compute. One of the best parts of deep neural nets is you no longer need as much feature engineering. But we still need high-quality data to feed into the LLM. The analog-digital divide isn't as clear as it used to be. It's one that I often think about.

What other analog things do we need to capture? What other things do we need to record and need to have accessible and available to us to feed into AI models? I'm curious to pull on LLMs for a second. And you may not have an answer, but it's something I consistently think about in terms of like reaching and thinking about both the S-curve of the current LLM infrastructure and architecture, but also thinking about how we actually get these things into deployed to the real world. Have you thought a lot about the hallucination challenge of like, how do we get these models to be generating output that is factually true in the world. This is like literally already a solved problem. And I think it's super funny that, you know, people spend a lot of time talking about hallucination because the fact that they're talking about interacting with a raw LLM where you're just like, all right, it's generating the next token. That's what you're taking as like God's word. Obviously, that's a bad idea.

And people hallucinate stuff all the time. What I mean by this is literally a solved problem. It's like, if you just use the link chain vector DBQA class and put your facts into the vector database and go look up your facts in the vector database and have the LLM summarize the output of those facts and then provide citations into where from the vector database this is drawn from. Well, there you go. You've got all the real world information that it's making a decision based off of. And you say, hey, look, summarize this context. Okay. It's going to do a good job that and you're going to have fully cited response. So like literally to me, that's the hallucination thing is like a fake problem. Because why would you ever just take an LLM's raw generated output as like

what you're giving back to the customer? It's obviously a bad idea. To take that one step further, where we are today is like, again, future is already here. It's just not evenly distributed is that you have things like lane chain that allow LLMs to do tool use, right? And so they can do web queries. They can do database lookups, they can do all this other stuff. And I mean, you know, the prompt is not freaking hard, right? There's a

lane chain thing, it figures out what tools to use, it figures out what sort of data gets put into some sort of context, like the prompt will just be like, hey, summarize all the data you've put into this context, and give that as the response to the user. And so that just be like, hey, summarize all the data you've put into this context and give that as the response to the user. And so that could be like some data they W get off the internet or curl off the internet, some data they look up in a database, some data they have from CSV files that are from the company, right? And then summarize that output.

And then it's going to be able to point back to where all those, you know, it did a curl and it did this action that some of that context is from the curl some of that is from the csc this particular csc file and then it's like all right then just point to where your facts are coming from you know this hallucination problem is just using lms in the wrong way we're so early in this adoption phase of this technology because we saw the demo everybody saw how easy the chat thing is and so like i said that is god god has spoken so we'll just take the output of the 10 commandments and we just stick it everywhere oh okay this is not the gospel here unfortunately but i think definitely the next phase is we have to prepare our knowledge graph or knowledge base or the vector databases

with our own data, because that's supposed to be the source of truth. If we want to play even a little bit more further, I wondered if you have any sort of takes on the more immediate future. Okay, we're going to have more data into vector databases, as we have today. There's still tons of limitations, I feel like, you know, using lane chains, QA vectors, classes I feel like today there's so much questions about how do you actually implement something correctly when it comes to, let's say, building this virtual employee that can maybe answer HR questions or SRE questions

and do something on its own. Well, performing tasks is really difficult right now. You can call some APIs, but we don't even have any sure how do we trust the outputs completely. Even though you have source of truth, I still may not be able to trust a virtual employee completely doing

something. It feels like an employee being helped. But if they're fully replacing a human, you need to trust enough so that we can be able to debug and believe it. There's some level of trust that still needs to happen, even with Decker databases, because the similarity search doesn't guarantee

the right intentions and outputs for now. And I think there's actually quite a few things that we have to figure out. I just wonder, do you have any particular takes on your own about what might be the next thing that can help us trust even the next level of outputs to take actions for these? Yeah, I think you make a really good point here, which is like, this is like a new way of making software. And we're all as an industry figuring out right now, hey, how do you actually do this?

You know, like you said, what is the correct way of doing this? We're still figuring out all those details like, hey, what is the correct way to write software where a big part of the logic of your software is this large language model embedded inside of this and interacting with traditional software as well as itself being a part of this new form of software? And so the answer is that the industry is going to figure all this stuff out collectively together. But of course, I've got a perspective on what I would like to do. And Lambda, where right now we're just hyper, hyper focused on providing the underlying

compute infrastructure. But more broadly, of course, I think like a couple of years down the road, how can we play a bigger part in making things like virtual people and virtual agents be a reality. And I think that I look at it as something like there's going to be more standardized ways that you log data in and out of interactions with these neural networks and interactions with these LLMs. Standardization of how you're logging conversations, standardization of how you're collecting feedback so that it can be then put back into, you know, whether it's RLHF or some other human feedback powered retraining system. At what point can you give an LLM access to the

payment tool? At what point can you give an LLM access to the read the CEO's emails tool? At what point can you graduate the LLM from writing a draft to writing a draft and sharing it throughout the company? And I think that the answer there is like simulation. Simulation of environments and comprehensive examinations is one aspect of that. I think that it's very interesting that when you start phrasing this as virtual people, virtual people, that's kind of like helps me really think about it. And the reason why is that when you phrase it with virtual people, then it answers a lot of the questions for you. So for example, at what point do you give a new college graduate full wire transfer capabilities at the company, not until they become the CFO and only with the

signatures of the CFO and the controller and the CEO. At what point do you give that intern a Divi card? Well, maybe not as an intern, but maybe you give them their corporate card with a $500 limit, right? What is an exam or a test that you might give somebody before they're coming in? If you have a bachelor's degree in computer science and you can pass all the tests that were involved with completing a bachelor's in science and computer science, well, then maybe they can start to learn how to program. And of course, that's a joke, right? Obviously, you don't need to have a college degree. In fact, you don't even need a high school degree or any type of degree to do any work in the world. I mean, that's all fake stuff pretty clearly.

But the point is that that is at least how society looks at credentialing people up. And so I think when you use that sort of virtual people as a conceptual framework, it answers a lot of the questions of what you need to build. Well, you need to build, what does college for LLMs look like?

What does a payment platform and system that's for LLMs look like? What does a simulation platform look like to give them exams to show that their performance is good? What does the performance review process look like for LLMs? And then importantly, this is this sort of like crazy future. It's like, you know, as a company, I've started already to start thinking about how do I market to LLMs? What I mean by that is like, I've already purchased something using the chat GPT feature. They've got a shop feature. I've purchased a few things from it using that feature. Well, this sort of like a new form of search engine optimization, right? It's like,

in the same way that you used to be like doing SEO to make sure that you're showing up at the search engines, like, what is your LLM marketing strategy? And I know it sounds a little bit far out, but like, I'm already talking to our head of sales, who the marketing organization at Lambda reports into, about how are we marketing LLMs? How can we start thinking about this stuff? I love that framework on the virtual person. It totally encapsulates a lot of the key questions. I'm curious, as one of our final questions, do you think we need new laws around AI?

Or where do you sit on that question? I have a personal take, which is I actually think current laws cover us. But I'm curious what your perspective is on what is currently the question du jour, at least in the media. In terms of new laws, you have to be really careful. Regulation has a lot of unforeseen consequences. You have to be

very careful when you let people who are very far removed from the reality of things make decisions about what the structure of regulation should look like. Of course, you have to worry about things like regulatory capture. You don't want one actor coming in and making regulation that favors them. So I think there's a lot of unforeseen consequences with that. I'm sure I can generate a spicy take for you right now, but I'm going to choose to just say that the world is a complicated place and you should be really thoughtful about the laws that you put onto books and that many of the people who framed the Constitution

and created a lot of the laws that has made the United States the longest running constitutional system in the world were very skeptical about introducing new laws. And many of them thought that you should get rid of all the laws every couple of years. And so I think you just got to be careful with it. Appreciate that diplomatic and also very thoughtful answer. Steven, this has been absolutely awesome. We thank you so much for joining us. How can people find you, your social media, your company?

How can they learn more about Lambda? And is there anything you want to have a parting thought before we go? You can look up Lambda Labs. It is Google Lambda Labs. LambdaLabs.com is where you can go and spin up a GPU instance. We've recently launched our generative AI demo platform called Lambda Demos, where you can spin up a stable diffusion web UI or an LLM chatbot and actually sort of run some of this stuff yourself for very inexpensive. And my socials are Stephen Balaban, S-T-E-P-H-E-N-B-A-L-A-B-A-N on Twitter.

And that's it. Tim, any last thoughts before we end this? No, I just hope that future comes sooner. And Lambda Labs definitely is going to provide a huge part of it. So we're looking forward to it for sure. Yeah, I think that if you really let yourself think on a 30-year time scale, that a lot of this stuff and more is going to happen. I think that when you look at an interview of, let's say, Steve Jobs in the early 80s and stuff, and you add 30 years into from 1984, which is the launch of the Macintosh, 30 years is 2014,

at which point we were very much well on our way to having one person, two computers, and let alone one person, one computer. And I really think that what's exciting is that everyone in this room is young enough that 30 years is something that we can look forward to seeing in our lifetime. And what is really cool about this industry, specifically deep learning, I'm saying, and AGI is that this is the most important technological shift that we're experiencing in our lifetime. And it's like, it's a really big honor to be a part of it and to be building a small, small part of the super future that we're going to all live in. Awesome. Thank you so much.