How AI Has Generated Trillions in Value Long Before ChatGPT and the Future of AI, with Konstantine Buhler, Partner at Sequoia Capital

Welcome to the Next Great Thing. I'm your host Andrew Greenstein, CEO at SFOPworks. It seems like AI comes up in every tech conversation we have these days. As if building a startup wasn't hard enough, now you have to contend with a new layer of complexity, but it also unlocks a new world of opportunities. If you're unsure of where to start, SFAPPworks can help. We work with startup founders and business leaders to design, develop and implement websites, apps, chatbots and more. Check us out@sfappworks.com Today's guest is Konstantin Buhler, a partner at Sequoia Capital and a pioneer in the world of AI. From his early days at Stanford, where his fascination with AI took root, to his dynamic roles at startups and as an investor, Konstantin's journey highlights a deep commitment to both innovation and investment and he brings a wealth of knowledge and a unique perspective on the AI revolution. During our conversation, Konstantin delves into the evolution of AI, highlighting key milestones like the deep learning revolution and the impact of algorithms such as Alexnet. He explains how AI has already created trillions of dollars in value by predicting users behaviors and optimizing business processes long.

Before the rise of ChatGPT.

Konstantin also shares his insights into the significant differences between startup life and venture capital. In startups, decision making is fast paced with numerous small, consequential decisions made daily. In contrast, venture capital involves making a few very long term decisions each year requiring a different approach and mindset. We discussed the role of AI in business today, from customer experience and co generation to AI enhanced security. Konstantin emphasizes the importance of focusing on real customer needs and using AI as a tool to solve significant problems. He provides valuable advice for entrepreneurs on how to build sustainable businesses by leveraging AI's capabilities without getting lost in the hype. One fascinating point Konstantin makes is about the future of AI. He envisions a world where multiple AI models collaborate and compete, each excelling in different areas leading to better outcomes for consumers and businesses. This approach highlights the importance of innovation and competition in driving AI development forward. We also touch on Sequoia's pre seed program Ark, which encapsulates 50 years of Sequoia's knowledge into a five week program designed to help startups with market approach, hiring, scaling, culture building and storytelling. Konstantin's belief that AI can make us more human is particularly inspiring. By taking over mundane tasks, AI allows us to focus on what truly matters, enhancing our creativity and connections. This episode is packed with insights for anyone interested in the AI landscape and its future potential. So without Further ado, here's Konstantin Buehler.

Konstantin, welcome to the show.

Thanks, Andrew.

Before we get into it, could you tell us a little bit about yourself and your background? And I'm really looking for how did you get into AI and what was the path to becoming a partner at Sequoia?

How far back you want to go?

Why don't we just start with when you got into technology?

Okay. So I got into technology in college. I grew up in Chicago, so I'm a Midwest kid. Chicagoland, Northern Illinois.

Me too.

Very cool. Yeah. And I came out here for engineering. So I came out to Stanford for engineering. I didn't know much about computer science. I'd programmed my first website when I was in fifth grade. But northern Illinois was not a place for computer science or startups or certainly not venture capital. Had never even heard those words in my life. But I came, I was lucky enough to come to Stanford and all of a sudden I got plugged into the innovation ecosystem here. And when I was deciding what type of engineering I would do, I kind of got pulled into the technology types of engineering and that eventually got me pulled into the company building of Silicon Valley and that eventually got me pulled into venture capital. So really it was an iterative path of finding really smart people, learning from them and listening really closely. And that's kind of how I got pulled into tech.

Have you taken a crack at the startup side of technology?

Yeah, I've worked at a couple startups and on a couple startups, including, you know, one of my co founders was co founder of a company I now have had that chance to work with for the past eight years. Awesome. And a company called Ethos and the Sequoia portfolio. One of the connections to Sequoia was actually a guy I worked on a company with. Very cool.

I mean, how do you compare startup life to venture capital life? Not many people see both sides of the coin.

Yeah. So they're dramatically different. I love learning and I feel lucky because I did angel investing, I did seed investing, venture, I did growth investing, worked at a family office, and then I even did consulting at private equity firms. And I got to work at startups over the course of maybe a seven or eight year period. So I got to see the gamut of what these things were similar or different. What was appealing to me. And not a startup versus investing is very different because a startup, you have to make really fast paced decisions. So there's this concept, Daniel Kahneman, thinking fast and thinking slow, where startups are a lot of really fast Paced decisions, thinking quickly, making decisions day by day, and investing, you have to be able to do that, but you also have to really slow things down because you're making very long duration decisions very infrequently. So to start up, you might make 10 small decisions a day that are consequential. And by the way, you cannot slow down because your company has to move really fast.

Yeah.

Whereas in the investing world, the biggest difference is that you have to make, in our case, two or three investments a year and they are really long duration decisions and you want to be well positioned to think about that well.

And you're doing it while the landscape around you is evolving so quickly. And I think that's one of the reasons why we're having this conversation today. You know, a lot has happened since the launch of ChatGPT, which I think to the layperson, that is sort of the start of the AI boom. As a partner at Sequoia, you have a front row seat to all this change.

Yeah.

And I just want to ask from your point of view, maybe you could describe from your perspective what has happened over the past few years in AI and where are we now?

Just at a high level.

There's two different pieces to this. What has happened over the last few decades, I'm going to say rather than years. And then the second piece is what's going to happen with Gen AI and this current boom that we're in.

I like that.

So for the first piece, AI is a long time coming. I have been an AI engineer in some way for the past 14 years. I was first introduced to neural Networks back in 2010 and the concept of a neural network was super powerful. It was introduced to me right over there at Lake Laganita on a run, where a friend of mine said, hey, you can design these algorithms that are based on these really simple perceptrons, like an individual neuron making a decision. And if you do tons of those in parallel in sequence, you can make really big decisions. And that ended up causing me to read about it all winter break and learn a bunch about neural networks. And I was very lucky because I was actually in school at a time when there was this revolution in neural networks for the first time. And it actually had good outputs. So there was an algorithm called AlexNet that basically took a big data set called ImageNet where there was this challenge around, hey, can you actually label images correctly? And for the first time, a big deep neural network way outperformed all the other, I'd say classical machine learning approaches and that was kind of the beginning of the AI revolution as we know it today. The deep learning revolution.

One of the first big breakthroughs.

Yeah, I'd say one of the first big breakthroughs for deep learning. But if you go before that, there were all sorts of classical machine learning algorithms that were doing really important things, by the way, creating trillions of dollars of value even back then. Yeah. I think one of the biggest things people overlook is that AI has already created trillions of dollars of value. Well before GPT3 or ChatGPT.

Before it became obvious.

Yeah, way before. So I'll give an example. So the biggest example of AI historically was taking a bunch of data and making predictions, which is what GPT does. It takes a bunch of data and makes predictions about. Next word. These were predictions about, I don't know, user behavior as an example. And if you basically look, I'll give a very tangible example. If you look at your Amazon feed or your Facebook feed, or frankly your Google search algorithm, all of those are based on algorithms that take machine data and make predictions. That's a type of AI. And a very tangible example recently was, do you remember when Meta stock was down like a ton, maybe 60, 70%, almost like some fraction of a trillion dollars?

Looked like they were placing the wrong bets at the.

Yeah, I think that's what a lot of people were saying. But the other thing was they weren't able to make as much on ad revenue because Apple put on this new privacy initiative which made it really hard to track people across apps. And you couldn't just carry those cookies simplistically, you couldn't carry those cookies with you across apps. And so you couldn't place ads as effectively. And that caused hundreds of billions of dollars of market cap destruction. Now everybody was kind of panicked because they said, wait a second, what are they going to do? And AI came to the rescue. We're not talking like a chatbot, we're talking predictive AI, which came to the rescue. And they said, hey, we actually have so much data that we don't need these cookies explicitly. We can actually make predictions about who you are and what you want to buy just based on these massive troves of data. And that's actually responsible for the majority of the hundreds of billions of dollars in market cap gain, not the new bets in Llama and the like. The majority of that back to the fundamental business got better in placing these ads. And I'd say today trillions of market cap are still responsible because of that type of artificial intelligence.

So maybe Another way to look at it is that the ChatGPT change was that it put the controls of AI into the consumer's hands. It became more of a thing that we could, we could create things with, we could do things with, as opposed to something that bigger organizations were already creating tons of value with.

That's perfectly said. Like, I sometimes this is a pretty tortured analogy, so bear with me. But I often compare it to the personal computer and I like the personal computer because it's a type of work mode, work mechanism. Just like AI is a type of work mechanism. And if, if it was before Windows 95 or before you had like a computer that looks the way it looks today, computers were terminals. I mean, you might have seen a command line before.

And I hate to say this, but I was going to say, yeah, my dad used to tell me about those.

Exactly.

Exact size computers in college.

And that's exactly right. Like, to me, ChatGPT was the graphical user interface is what it's called, like the, the first time where you actually could look at a computer screen and like point and click. And that was awesome.

Anyone could now, anyone could do it. This technology.

Yes, but the technology was there but for that interface. So to me, ChatGPT is a new interface and that interface is language. And, and everyone can speak language, just like everyone can look at a screen point and click.

And so then the market of people who can utilize technology is tenfold from what it used to be. In the same way.

Yeah, 100 fold, maybe. Yeah, that's exactly right. Like historically it was, you know, only people who could write these AI algorithms. And now it's everybody. I remember my sophomore year of college, we had this project where we wrote an AI algorithm. Everybody did, wrote an AI algorithm to predict a heart defect. And we downloaded what at the time felt like a ton of data from thousands of patients and we were able to predict heart defects at the rate of a physician. That was pretty magical. So think about the impact of being able to predict a heart concern. As a sophomore in college, AI was there, there was this amazing impact, but it wasn't accessible to the world. It was just a few people who could code and understood kind of how these algorithms work. And now everybody has access to it. And that's when the really exciting, interesting stuff is going to begin.

How would you compare it to the smartphone, Dawn? Let's say the era of the smartphone. Because I actually like your analogy better in the sense that it opens up the user interface. But I had been thinking about it in terms of a Smartphone in the sense that all of a sudden, now this new, powerful technology was in your pocket, in your hands, you could just use it all the time. Do you see an analogy there?

Or.

Or is it because you've also worked as an investor through the shift to smartphones?

Yeah.

And probably seen the impact it's had. And you studied during the AI growth, early growth. Do you see any comparisons?

So I could see it as like a parallel in the place that you do work has shifted. And by the way, audio is going to be a powerful form for accessing artificial intelligence. It already is. By the way, the hyperscalers knew that a decade ago when they were launching Siri and Alexa. And all of these Cordova, I think, was Microsoft's at the time, they knew this would eventually happen. So credit to them. So I could see that part of it be different, as in the places in which you can do work have changed. But the mobile phone was kind of an extension of the personal computer. It was a personal computer in your pocket connected by the Internet wirelessly, just smaller. Yes. And so, yeah, I could see it being an extension, but the grandfather of artificial intelligence is the personal computer, and maybe that turns into the smartphone and maybe that turns into, you know, artificial intelligence on the go through audio at some point. Audio and maybe video and these interfaces as opposed to. I don't view it as a communication revolution necessarily. I view it as a work revolution.

Yeah, Well, I guess maybe one of the contexts for that comparison is when talking about opportunity for entrepreneurs. I feel like there is a shift that happens every 10 or 15 years when it might be when the smartphone drops. And all of a sudden there's like a whole new playground. The rules are changed. We have new technology in our hands. So the things that used to be imposs are now possible. You can now have GPS with you, so you can now hail a cab.

Yes.

And when I'm looking at that, I actually went back and looked at when the smartphone launched in 2007, wasn't even an app store on the first iPhone. That came a year later. And then it was another year, I think, until Uber launched. And then. Yeah, you're right, another year or two until DoorDash.

Yeah.

And snap, I was thinking. But it was 10 years until TikTok launched. So I was thinking there's actually more time than you think. Although I think things are accelerating now. So I'm trying to understand, where are we today in terms of, like, the first couple of things that came out, it was like the flashlight app.

Yes.

To me, that's the same as like the first generate your profile.

Absolutely.

And then all of a sudden you had like things really started to disrupt.

Yeah, yeah, I think that's a great parallel. I think that's true with a lot of new technologies. So my partner Sean points this out pretty frequently, which is a lot of the companies we're seeing now might be like the early App Store days where they're super fast to adopt because it's a very constrained environment and everyone wants to try this out and so revenue grows quickly. There was the flashlight app. There was this beer drinking app where it uses the gyroscope to have the liquid flow down. Those were killer apps of the initial App Store era.

And that's sustainable.

Yeah, they were not sustainable. They were all trivial. Right. And eventually some of those were built in as utilities, just like some things will be built into utilities to the open AIs of the world. And then some things were distinct enough that they became big businesses like Ubers and doordashes and snaps of the world. That type of progress will play out here. And I think that happens in almost every technology revolution. To me, that's an example of we have these new capabilities, let's look at them as that and find great potential businesses from there. I mean, the Uber and doordash capabilities were directly from the gps. The Snap and Instagram capabilities came directly from a camera in everyone's pocket. Now we have intelligence as a service available. What capabilities come from there?

It's a good time to talk about. A few months ago, forbes launched the sixth annual AI50 list. And I'd be curious to see if you've seen any trends in that list over the last few years. This year that kind of line up with maybe the sorts of jumps we saw in past technological breakthroughs. What are you seeing trend wise?

Yeah, this is the sixth year we've done the list. When we started the list, it was very computer vision heavy because that was the big new capability. So if you look at the history of AI algorithms, early AI algorithms were handwriting detection. These were algorithms that basically could take handwriting and say what text it is. It sounds very boring now, but that was 98, 99. These algorithms were on the up and up in the 2000 and tens. Those were computer vision. And those launched a lot of really interesting companies, actually. Companies like Waymo, which you can now take on the street in San Francisco, or verkada or viz. AI. First FDA approved medical device. Medical AI device. And 2019, 2018, when we did the first Forbes list, a lot of these were computer vision oriented and those were in all different spaces. A lot of them were self driving. You've seen that hype, frankly. And these things come in kind of hype and troughs has died out. It's consolidated to a few great players, including the ones that I just mentioned. And now you have way more in the generative AI space. And I predict that in two or three years that will die out and there will be new capabilities of AI. There will be new algorithms that are on the list.

Then what is your advice to an entrepreneur who is trying to build a startup? Is it to build in the moment in generative AI or is it to try and look ahead to what will happen after this phase dies out?

I would say almost neither. So I think the things to do are take all of the hype out of the fact that this is AI and say, okay, I have this new capability. In 2017, that capability was the ability to see that a computer can see. Can see. Okay, a computer can see and identify what's in an image. Today it's a computer can write, a computer can generate art. A computer has some basic knowledge as a service. And I would take the words AI out of it. Oftentimes when we get a memo, it's Akoya and it's an AI company. Four, just take my finger, I put it over the AI and just remove that and say a company 4x. So step one would be remove AI from it completely. Step two is look at the capabilities, try them out yourself. Step three is what does this make possible? And then step four is, is there a problem that's going to be around in 10 years no matter what, that every human really wants and that I can move back from and say, hey yeah, this is a problem that's going to need solving in 10 years and we have a new capability to go solve it.

When we were talking to companies on the show, we would describe them as AI companies for a couple months and then we realize it's like electricity or the Internet. You know, you're all AI companies at this point.

Well, AI, the term AI is also super flawed. Honestly, I have a problem with the term AI, even though my master's was in artificial intelligence. The term is really bad because it kind of means whatever is next. Like no one would consider handwriting detection AI today. But that was if you asked us 30 years ago if you had an algorithm that could drive you from San Francisco to New York within an hour, prediction with every turn, you'd consider that artificial intelligence. Today, that's Rote gps. That's simple. The problem with artificial effect, I read.

A little bit about this.

Yes.

You want to tell us about it? I mean, I guess you are already, but yeah, sorry to jump ahead.

And the problem with the term AI is that it uses the word intelligence and humans kind of measure themselves as intelligence. It's a very human oriented term. And as soon as a computer does something, we kind of remove it from being intelligent. And I'll just give two examples. Example number one is Blase Pascal, famous mathematician. He was talking about an intelligent machine 400 years ago. And his intelligent machine could add, subtract, multiply and divide. That's a calculator. And by the way, 400 years ago, that sounds like a really intelligent machine. Adding, subtracting, multiplying and dividing is non trivial. Yeah, no one would consider that artificial intelligence today. And then I gave the other example, which is, let's say it's 1980 and your parents are planning a road trip and they've got a map and they have to map it all out. And someone said, hey, I have a machine that can tell you exactly where to start and where to finish. And it'll tell you every turn and it'll speak to you on every turn and it'll know exactly when you get there. They would consider that either magic or artificial intelligence. And today that's Google Maps. As soon as you get a capability in the mainstream, it's no longer considered artificial intelligence. I'm already seeing this with computer vision from five years ago. No one really calls it AI anymore. It's cv.

A computer can detect what's in an image.

Exactly. It's computer vision. It's like, okay, great. Or autonomous vehicles. Like those are probably the most advanced AI in the world right now. It's amazing what they can do. They're just AVs now. They're autonomous vehicles. Same thing's gonna happen with generative AI. They're gonna be chatbots, probably some new term, chatbots and like picture generators. And the next thing will be AI.

Is there a term you would use instead of AI, or is it just that we, we think of AI as something that we can't quite achieve and as soon as we achieve it, it's not AI anymore. That's the next thing we're trying to achieve.

The term AI is kind of like saying the term high tech. High tech is, it's definitionally the frontier.

Right.

And then as soon as it becomes understandable, we kind of move on. So artificial intelligence, I'm fine with it. I'M fine with using the term high tech, but it's just going to constantly move on and it's a treadmill to the next thing.

So it's not a very good way to describe your company, as you pointed out.

I think it's not a long term good way to describe your company. The long term good way to describe your company is what do you do for your customers? What's the long term problem you solve?

You believe that as more AIs are developed, they'll begin to work together as networks of AIs?

Yeah.

Could you explain that?

Yeah. So a lot of this is also taken from the way AI has evolved over the past 20 years, where AI algorithms tend to start on their own. So you have individual algorithms that do some function on their own and then over time you stitch a bunch of algorithms together to do a task. When we talked about heart earlier, hard classification, you might have one algorithm that actually extracts where the heart is in an image, another algorithm that does volumetric analysis, another algorithm that does classification of what type of heart problem might this be, and then another algorithm that now with Genai can spit out an analysis of what's going on that will be stitched together and over time, each of those parts, the first three parts are now pretty standardized, pretty understood. The fourth part is the Wild West. All of that will be stitched together. That's one way where these things are going to be working as networks. The other way is I don't think there's going to be a world where there's one foundation model to rule them all. I think you're going to have multiple and I think that's a great thing because competition is wonderful for consumers and for builders. And you're going to have a bunch of these algorithms working together and against each other because they have different areas where they really spike. So I use one of our. Our portfolio company is called Dust. Anytime I search, it's Dust tt. Anytime I search or I want to use a chatbot or chat AI, I use GPT4 on it. So I call GPT4, but I also call Claude 3 and I also call. They have Llama and they have Gemini and they have Mistral and I can call all of those assistants and see which answers what in a good way. And I can also have them build on each other and say like anthropic, go and add citations to the GPT4 example. And I find that they work better that way.

So your models are in a way collaborating with each other?

Yeah, yeah. Both competing. First I usually have them compete and I say like, oh, which is the good, the best initial answer? Say I'm doing a market sizing. Which model does a really good job of figuring out quantitative market sizing? And then I have them collaborate like, hey, can you build on this without tearing it apart?

Do they know they're competing? I mean, are you actually presenting it as, hey, here's the they don't know.

They'Re competing, but they know they're collaborating because they're called. Their API is called in parallel, so they don't know when they're first called. It's just an API call that returns. But when they collaborate, they do, because I say, you know, add citations to act.

GPT4 because I've heard wild things about how to get better answers out of AI, including that if you tip them, if you say you will pay them for their response, their accuracy rate, or their answers get better.

Oh, that's funny.

I was listening to it as a client podcast, three part series on AI and he had a researcher on there who was talking about some of these quirks and one of the ones that blew my mind is they did a study and found that AI answers weren't as good in December as they were in May. And they didn't exactly know why, but their best guess was that AI has somehow internalized that December is a slower month and it doesn't need to work as hard.

Ah, interesting.

So I was just wondering if putting in the prompt that it's competing with other AIs is actually helping the answer or whether it's more you synthesizing from different models and seeing different approaches.

So that happens, but we shouldn't have to do that. Prompt engineering shouldn't be something that we as humans have to do. It should be something that's eventually abstracted away. It'll start that way, but eventually, like AIs should work together and against each other to do more of the work for us.

Yeah, that was exactly the follow up point to that, which is that we think prompting right now is really important and it's a skill. Everyone said prompt engineering is the new skill and the follow up was no, actually AI will just get better and better at deducing and figuring out, because at the end of the day, these models just want to try and figure out what you want.

Yeah, that's right.

What happens from a content creation perspective when AI models are collaborating and learning from each other? There's this horror scenario where everything is just an AI ripping off another AI off another AI and where does the human creator fit into this?

There will always be room for the human creator. I'm very high conviction on that statement. There's always room for the human creator as long as they're differentiated. The net new novelty of the human creator's ideas have to be significant. If you're doing something like rote or undifferentiated, that's scary in this world. Hand wave for a second. But for those who are listening, a lot of AI can be represented in a multi dimensional space. What is a multidimensional space? Think about a 2D chart, okay. And then we make a 3D, we added a new dimension and then 4D, it's moving through space over time. And then there's 5Ds and 60s and 70s. All we need to know about that is it's like adding new columns to a spreadsheet. Just like one column could be X, Y and z in a three dimensional interface. There's also alpha, beta, gamma, delta, etc. So like we have a multi dimensional space and AI is really good about learning the things that are in the clusters of space that it already has data about. So in this multi dimensional space, AI is really good at saying there's a ton of information, how to write a great podcasting script. I don't know, there's probably some blogs out there and information out there. There might be very little information out there about how to discuss prompt engineering. When prompt engineering started to become a thing, so 18 months ago, there would be almost no training data about that. And that's actually where the opportunity rises for people. Is the net new or the net different in this multidimensional space? Think about a three dimensional cloud with a bunch of clusters and if there's a net new cluster that the algorithm hasn't trained on, like that's where humans need to look at.

But a model doesn't know that those columns are populated by human content versus AI content.

It should in the future. Like it doesn't right now. You're right. Maybe, maybe some of the labs are doing that. I would imagine they have initiatives to do that.

I've heard of companies doing it. I haven't heard of great success there. Just reading New York Times article on restaurant reviews. Auto written restaurant reviews. And yeah, they trained it on Yelp and they, they then did touring tests where they asked people, could you tell the real reviews versus the false reviews? At first they could tell. So then they went back and they tweaked the LLM to add more typos and yeah, and Then it got to the point where it could tell us often than they.

Yeah, well the big way to do this is basically we know that pretty much the whole Internet was human made prior to November 2023 or 2022. Like so that's the biggest way to do it is all the. Many of these companies have huge databases of the entire Internet almost in plain text form or enriched data or what have you. And like November 2022 and prior, almost certainly human generated 90 some percent after. Still very human generated, but less and less probably over time. That's like the biggest way where you can actually differentiate in your training data. And if companies are really smart about this, they'll continue to say, I think at some point you need to have clarity about where you're getting the training data from. And hopefully they label things and say, oh, they won't. They probably won't be able to algorithmically determine if something's human created or not, but they probably will be able to say, hey, this came from Andrew's website, is a post that he created and is attributable to him.

And the big companies, I think they're very incentivized to want this to happen. I think they want to understand whether they're training on good data or bad data.

Good data or bad data, I think is where the problem is. There might be machine data that's better than human data, better in terms of producing an output that's better for the end consumer. I don't think that's the reason for the incentive is actually fairness data attribution slash. If retroactively things change on how we give credit to algorithms to the creators, those are the areas where it actually makes more sense. The problem with using machine data in it actually might work well and in some ways it does because when you have AI algorithms self infer a few times the quality of output goes up. So that actually might be good for your experience. The problem is not that. The problem is that those are all in the same clusters. Remember that multidimensional space I was kind of describing? Like, oh, you're just making one cluster denser and denser and denser and denser. You're not adding net new clusters for this multidimensional world.

I feel like I sound like I'm negative on the prospect of AI and what it'll do to the Internet, but I'm actually very positive and excited about it.

I thought you sounded positive.

Yeah, okay, that's good.

Realistic.

If I think about the way that like the Internet has gotten from A consumer perspective. When you're out there looking for stuff, it's not as great as it was. Like, it was fun in the beginning, and you could find, like, that expert who was creating a video on woodworking. And now it's just so hard to find anything. And there's so many ads and there's so much SEO. And I'm actually wondering if this is an opportunity for us to change the way the Internet works. Again, for the positive. But I'm curious, what is your take on the state of the Internet and what the impact that, you know, all these companies are going to have in light of some of the challenging things that the Internet has struggled with over the last 10 years?

I'm pretty positive, actually, on the state of the Internet. I can think back to. Yeah, the early days of the Internet were particularly fun, but that's also because it was new. And humans have a novelty effect. And we love novelty, especially the older we get. Actually, our brains consciously like novelty, the older we get, and younger actually like more redundancy. That's why, like, a toddler listens to baby shark again and again and again, and an adult that drives them crazy. It's just different development patterns in the brain. So I think the Internet was more novel earlier and now has become mundane. It's not just where you go to watch the end of Z World, which was a early Internet, a movie, YouTube that I remember watching as one of the funniest memes of, like, the early 2000s. It's where you do your online banking and it's where you do your online accounting and where you do your business and all of these mundane things. I think that's great.

Socializing.

For socializing, I guess worse, I'd say, yeah, I agree with that. But for the better is, look, online banking is safe and secure. There are so many more protocols. I'm not going to say the Internet's safer because that's. There's nuance there. Loaded, but, yeah, totally loaded. But there's a lot more you can do on the Internet with way more advanced protocols. So I'm very enthusiastic about how much has been going on the Internet. Do I think that things are harder to find to your point? I don't know objectively. Like, if we were to measure it now versus then. I think companies respond to incentives and that's why you see so much SEO stuff. And frankly, I just want the people who are building an AI to be very conscious of that so that we have a better user experience going forward.

What sectors are you Most excited about being transformed by AI.

I think we will see the sectors that are most material in the years to come, but things that are clearly working. There's a big swath of things in customer experience and customer support and a lot of people have talked about that at ad nauseam where especially now that we have text to speech and we have the ability to understand large databases of customers or customer experiences and problems. That's impactful. And we have businesses in that category like Sierra or Cresta or the like, all working and doing well. There's also this huge swath called AI code generation and that's also a huge productivity booster. And we've seen that pretty broadly. And we're going to see that not just help with autocomplete, but we see things like Cody and Source Graph or factory doing full implementations of code. Those are two huge ones. And I think that there are also big areas like entire verticals. The legal industry is about writing and writing structured text. You can kind of understand a contract like a, like code. We have wonderful businesses in the contract space like Ironclad. We have businesses serving legal firms and doing legal work like Harvey.

I was going to ask if Harvey was one of your companies.

It is, absolutely. I think it's a great response to the need of having associates and analysts and frankly more younger people be more productive and not have to be billing 3,000 hours a year just to get by and get the promotion. And then there's all sorts of new exciting spaces. We're seeing more and more around AI for security. Not AI security, but AI for security. So actually leveraging artificial intelligence to be a better security engineer and boost productivity and maybe just detect vulnerabilities that you previously might not have been able to detect. We're seeing a ton happening in AI for go to market. So central source of truth data and then making predictions about who to upsell to when and how to better serve your customer and how to draft and send that, or being live in a conversation and having a deep technical understanding of what your customer's asking and what the response might be. But what you'll hear in everything I'm saying is it's all really pragmatic. We're all thinking about hey, what are real customer needs? What's a real response because of this new capability, not hey, it's AI, but rather this new capabilities here. How do we use it to serve customers?

And one of the great things about where we are now, it kind of speaks to your accessibility and interface. You can now basically talk to an expert in anything and figure something out, whether that's, you know, you're not a coder, but you're like, can you walk me through how to build a simple app? And you ask it the right questions, you can get there. I mean, I've talked to experienced engineers because I think you still need to be experienced to really utilize it. I don't believe a junior engineer is now a senior engineer just because they have AI. I think that, you know, you can learn new things. So I was talking to a senior engineer, 20 plus years of experience on, you know, he's building AI tools now, and I was asking him if he feels like he needs Someone with a PhD in machine learning to be able to figure this stuff out, if that would help him. And he was basically like, all the questions I have, I get answered. Yeah, it's just now going through the work a bit, you know, if you want to of course, learn a language, if you want to learn anything. I just feel like it opens this door of how I can myself go and figure any topic out. It really feels like a door opener for opportunity.

Yes, it is. So I'm going to use this concept again of like the multidimensional space and I'll use just like a three dimensional space in this case. So when I was programming, when I had a question like that, I would Google it and hopefully find an answer on Stack Overflow, which was kind of the de facto place where people would ask coding questions and get coding answers. Let's come up with an example. So let's say I wanted to implement a sorting algorithm. So just an algorithm that puts numbers in order and I wanted to do it in Python. And let's assume that there wasn't direct sorting algorithm for Python answer on Stack Overflow. Of course there is, but let's assume there wasn't. Historically you could go and you'd find maybe a JavaScript answer to QuickSort algorithm and then you'd read it and you'd say, oh well, for Python we don't actually use these types. We use these types and we actually have a different way of calling functions. And so I'm going to implement it slightly differently and you do that translation. So now in the AI world you have a cluster here called quicksort. So how to do it, and it's been trained on and you see this algorithm and then you have quite literally like a shift in direction Python to, or JavaScript to Python. And what the AI world is able to do is it takes this cluster of how this Sorting algorithm works and it shifts it over to hey, how a Python algorithm would actually be written here. And it actually gives that direct answer. So you take a bunch of clusters and you're able to basically combine and have infinite answers as opposed to the human connecting two different sets of answers. That's really powerful. And that does make a custom expert. Now, the only downside is nothing's deterministic in that world, as in nothing is verifiably. You run the code and it works. Whereas before you could actually say this is something that was exactly called and the comments say that this is exactly how it worked in this new future or in this new present. You get a better answer because it's between these two clusters, but it's not verifiable. You have to run it on your own and actually see. So it's a different set of skills and you might have to debug that on your own.

And that's an opportunity for the human element.

That's always the opportunity for human element. Yes. The human becomes more important in an AI world, like I've super high conviction, the human becomes more important.

The AI can translate anything and that opens up knowledge to everyone.

Exactly.

But you need the humans to verify.

Exactly. And then the nuances become more important. You used a great example, translation. Automatic machine translation was an example of artificial intelligence. I don't think a lot of people call it AI now, but that was a huge breakthrough in artificial intelligence and it hasn't really eliminated the need to speak other languages. There's still a lot of power in the nuance of speaking another language and the human connection of speaking another language to a person in their native tongue. I love speaking Greek to people because that's my cultural tongue and I want my children and future generations to understand that language. For example, Spanish for you.

I love the movie Arrival because it hits that point about there's concepts that don't exist in your language that exist in a different language.

Yes. And AI is not going to replace that. Humans still will benefit from understanding that. But if I need to speak Spanish to you or you need to speak Greek to me, it is nice to have this additional tool to do some of that translation. And then your skill of actually understanding this nuance will make it even better.

I want to talk a little bit about investment approach and I'm basically trying to understand what you look for in a company. What's interesting, you know, oftentimes we're all the time, actually we're talking to startup companies. This is the first time we're talking to a venture capitalist, so I feel honored.

I feel like more of a startup kind of person. So hopefully it's not too much of a transition.

No, it's good to get different perspectives. But I think the way I want to ask this is I'm going to ask you about three portfolio companies. You mentioned Dust. You can start with them if you want. I want to know what was the pitch? I want to know what you found most compelling about them.

Yes.

And I want to know what their biggest challenge as you see it is.

Yes. So Dust, the pitch. We met at the seed 16 months ago. We met the company and at the time it was two founders and an idea. And the pitch was frankly the two founders and a direction. The two founders were exceptional. One is Stanislaus Polu. He is a brilliant artificial intelligence Engineer. The past 10 plus years was the first OpenAI researcher in Europe. In an era when they only hired in person, they opened the Europe presence to have him on board. He was a Parisian AI researcher and we did the same grad program in artificial intelligence and I just knew reputationally how brilliant he was. And the second founder was the best product person by references in his Geo and of the smartest product people in general. He'd been running product at a big French product oriented company called Alain that was well respected for how thoughtfully they approached product and their differentiation was perfect for who they were. They said, hey, we're going to build a product that harnesses AI for where it is. We're not going to build a foundation lab. We're not going to try to do things that are outside of the realm of where the current capabilities are. We're going to try to build a product that's usable for productivity for people in the enterprise using these new primitives that AI allows like via API. That was the whole direction at the time. It wasn't more concrete than that. And over time it's become this really cool Slack like utility. Like I use it as frequently as I use Slack and I use it for communicating with and from AIs and asking questions to and from it. And now they've added multi actions and they're able to query all of my proprietary information. And when I, when I write a memo I usually write the first pass of it. With Dust I have a memo custom memo bot that does it. So that's them.

That's such a cool way to frame it with the Slack analogy because you're basically saying it's a communication tool for LLMs to communicate with LLMs.

That's right. That's right. And to do work that then communicates with my other partners. You asked their largest challenge. Their largest challenge is they're so smart and so capable and they've assembled such a wonderful team that the world's kind of their oyster in product possibilities. And they need to limit and focus the tangible actions that can be taken with this really fundamental platform. And they understand that. And they've done a really good job of saying no to a lot of different directions and saying yes to a few that matter. But that's the continued challenge.

What's the second portfolio company you want to talk about?

Kumo AI K U M O AI Their pitch was they had a new look on this technology called graph neural networks. And really the way to think about graph neural networks is they would do neural network training and inference over a graph of information. A graph is a computer science data type that's super connected, as in there's edges between a bunch of data types. It's not just a specific data type, it's not a string like a piece of text, but it's connections between the pieces of text. And they were frankly the best team in the world at this very specific type of technology. The professor who was kind of at the forefront is a guy named Yuri Leskovich. And I always found him super brilliant professor in the topic. Both brilliant and able to communicate and charismatic at the forefront of graph neural networks, this new technology. And he was working with two co founders, a guy named Vanya Jasakovsky and Hema Raghavan, both who were executives at big companies that were using graph neural networks to generate billions of dollars for the companies. So Vanya was the cto of Airbnb. Hemo was running all growth AI at LinkedIn. They were all using graph neural networks to use the topology, basically the interfaces between all of this different data to make predictions and boost revenues. And they said, hey, we actually think it's now at the point where you can put these on data warehouses, run inference and kind of bring the power of these big tech companies like Airbnb and Yuri was at Pinterest and hemo was at LinkedIn. To every enterprise and every consumer oriented business. That was the pitch. And we literally worked together for six months before even forming the company on direction and idea and having conversations. And that was another formation stage business. And today that's what they do. They helped companies make hundreds of millions of dollars more in revenues through their product.

Amazing.

And challenge for that is I'd say also constraint like this is such A foundational primitive that they have to find. Hey, what are the use cases that we want to focus on? Because you can do almost anything with the type of broad technology and they focused really on personalization. Next best action. Some KYC and fraud and are really good at those things.

It's interesting. I've heard of course that focus is one of the hardest things for a company to achieve.

Yes. Third company. Yep. You know, I use one that's generative AI. I use one that's predictive AI. I'm going to talk about one that's Computer Vision and that's Verkada.

That's not AI.

Oh, it's AI. Absolutely. Computer vision's not. Oh yeah, it's funny. Yes. So Verkada which use computer vision to do physical security cameras. So they actually are primarily a hardware company. They make physical security cameras. Initially the initial pitch was we make physical security cameras that stream to the cloud instead of stream to like a box in your back office which is how a lot of security cameras worked back in the day. And their pitch was we're a wonderful hardware team that has great product sensibilities and software and we're going to build the next generation of physical security cameras. Now again the founders fit that. Philip Callison was a brilliant product founder who had known for 10 plus years really good at product always in software. His co founder Hans was a phenomenal hardware builder and go to market person as well as their co founders James and Ben who were great on the software side. And at that point we actually intersected with them. Later on they had already been growing so they had really good sales. People were buying the product and using the product as part of the pitch they were able to say hey, we're making this next generation physical camera. People are buying it. We're great at product and hardware. We're going to make a better solution for enterprises to have physical security cameras. That plus over time we're going to make the entire operating system for a physical building. It's going to be the camera, it's going to be access control. You're going to be able to use people's faces if they so choose. Privacy preserving to access a building, not just swiping a pass. We're going to be able to know if people are in a room to turn down air conditioning and help the environment or turn off lights at the end of the day and turn them on. All privacy preserving as well by adding AI inference on the edge. And they have been just amazingly fast growing.

I love company has like A base hit, maybe the video camera, computer vision. And then they have their home run like the operating system. Yeah, you have like, you know, me too.

Yes. Super brilliant of them. And they always knew that was the direction, but they kind of were building, building toward that. And challenge. You asked for challenge for them. I think challenge for them is, look, they're at the point where they're scaling very rapidly and managing thousands of people is hard. They're at the point where it's thousands of people or maybe crossing thousand plus. And yeah, that scale, 10 to the fourth scale people is difficult. And Philip has developed so much as the CEO, he's a friend who I get to watch. And that's one of my favorite parts of this job is you get to watch your partners who become friends develop in such a way that they become so great. So I think that's the, that's the current challenge there. Interestingly, all three companies I picked, I picked them intentionally. Ones at the series A stage, the B stage and the growth stage. All three companies I picked were actually second time founders. Now that I am running kind of data backwards. The first sold their company to Stripe. The second sold their company to Pinterest. The third sold their company to one to Chegg and one to Cisco. So they'd seen it happen before. And I think that's something else that I always appreciate in founders.

Yeah.

Quick bonus question.

Yes.

What was your pitch to them?

I'd love to ask them that. I'd love to ask them what they saw because I'm very grateful for them working with me. They're all, all of them are founders that any venture capitalist in the world would be lucky to get to work with. My pitch, I imagine is kind of the same. I'm a transparent person who honesty is like number one priority for me. Honesty, I believe is always the best policy. And my pitch is generally, look, you're going to get someone who's honest, hard working, dedicated, and who by the way, understands the product very well, the engineering aspects of product very well. That's kind of the baseline and then the capabilities that you get on top of that are me and our firm's ability to help with go to market, hiring, scaling and beyond. So that's the fundamental pitch. What I think really resonates and at least for the relationships that I mentioned there, what I think really resonates is, hey, is this a human that I can grow closer with for the next 10 years through good times and through bad. And I certainly feel that way about all those founders believer that at the.

End of the day, we're all humans and the human element is so important.

These are, were these series A companies seed. So initial investments. Seed. Seed. B.

Okay. You also have a pre seed program.

You want to talk a little about that? Absolutely. So we have a pre seed program called Sequoia Arc and it's really 50 years of Sequoia tribal knowledge compressed into a five week program. And what you get from that is certainly our understanding of, hey, how to approach a market, how to start hiring effectively, how to scale up, go to market. But also, you know, fuzzier but incredibly important things like how do you approach culture, how do you share a story, how do you have a base plus a home run story like you were just describing? Andrew, it's been a really fun program to get to be involved with. I've had several companies already go through it. They love it because of all that knowledge, but also because of the community.

And where should someone go to learn more about that?

Ark.sequoia cap.com ark. Sequoia cap.com Absolutely. Awesome.

I'm going to get you out of here on one last question we ask everybody.

Absolutely.

Why is AI the next great thing?

Artificial intelligence is the next great thing because it has the possibility to make humans more human. If we do it right, artificial intelligence is going to take all of the things you don't want to do and let you focus on things that you do want to do. If we do it wrong, it will be a net negative. That's why I'm so passionate about being in this industry is I really want us to do it right so that AI is a huge net great benefit to society and allows us to focus more on being human, make us more human.

That's an amazing answer and this was an amazing conversation. Thank you so much for your time to acknowledge your experience. I really enjoyed it.

So good to have you here at Sequoia. Thank you for coming.

The Next Great Thing is hosted by Andrew Greenstein, CEO of SF AppWorks, a technical agency that helps organizations and entrepreneurs design, develop, launch and maintain web apps, mobile apps and platform integrations. The podcast is produced by Kristen Sills with marketing by Leah Roos. For more episodes, search for the Next Great Thing wherever you listen to podcasts or check us out@sfappworks.com.