Les technologies profondes sont devenues une nouvelle catégorie d’entreprises en démarrage qui s’efforcent de relever des défis techniques et scientifiques très complexes. Manish Kothari, associé directeur de First Spark Ventures, cherche des occasions dans le secteur. Il discute des tendances qu’il observe avec Kim Parlee.
Print Transcript
[AUDIO LOGO]
* Artificial intelligence has taken investors and the world by storm in the past six months, but it's only one of many so-called technologies called deep tech that are potentially world changing. And our next guest is in the business of finding these deep tech opportunities. Manish Kothari is a Founder and Managing Partner at First Spark Ventures, a venture capital fund investing in deep tech who also focuses on high tech innovation in science and engineering. Manish, It's a pleasure having you with us. Thanks so much for joining us.
* It's a real pleasure, Kim. Pleasure.
* So let's start off with maybe tell us about First Spark ventures. You yourself and your colleagues who are involved with First Spark have a long pedigree in innovation technology. Maybe tell us a bit of how it came to be.
* Yeah, so First Spark is really a combination of three different groups. It's a brand new $400 million deep tech venture fund. We invest in companies where the scientific breakthrough has occurred, and there's massive IP advantages there.
* So it's founded by really three core groups of people-- myself, Stanford Research Institute, or SRI International, who is a partner with us, and Eric Schmidt, who has helped-- former CEO and chairman of Alphabet. We invest in deep tech. SRI has been at the foundation of a number of great investments, great creation of companies in the past, everywhere from Siri for your iPhone, to Nuance, to Intuitive Surgical. And we're sort of bringing that sort of thinking to the forefront here.
* Tell me what that sort of thinking that you're bringing to the forefront means. Because one of the things I was looking at, a previous conversation you had with somebody, you talk about the ability to shift the timeline. So with companies and innovation, I'm assuming that means making these things happen a little faster.
* Correct. Correct. So when you look at tech, which is really applications versus deep tech-- deep tech is the things that cause sustainable, long-term, fundamental changes. Think about Nvidia, think about Tesla, think about SpaceX-- these companies really transformed the landscape in a very deep and profound way.
* However, they often take very long to get to the scale of return that's great from a venture perspective. The good news today is what you've seen is that the fundamental component technologies have developed so much that we're now ready to accelerate much faster. And our fund really focuses on providing these startups the tools and tricks, if you will, to be able to go and hit the same sort of outcomes as you see in Nvidia, Tesla, and SpaceX but in time frames that are far speeded up. This would not be possible 5 to 10 years ago, but it is definitely possible today.
* It's fascinating. I mean, again, technology begets technology, I guess, is what you're saying in terms of things moving faster and faster. If people look at the fund itself, you talk about being involved in advanced computing, digital biology, and biotech, and cyber physical systems. I want to speak about each of those so you can educate us as to what that means and what you're seeing. But if we think about advanced computing and AI, just to start with, has ChatGPT been as big a deal as, perhaps, maybe the mainstream thinks it is in terms of advancing AI and bringing it into the world?
* The short answer is, "yes." ChatGPT, or generative AI in general, has really expanded things out. Let's talk about it. There's a few things that it's fundamentally changed, right? Let's start with this concept of data synthesis. So before computers even existed, we used to use books. Well, it was oral, and then books. Then you had encyclopedias.
* And then you had computer systems, and then you had the internet. Each of these things, what they fundamentally did was increase the speed at which you could synthesize information. So one of the things ChatGPT does, or any of these large language models do, is that they really inform our ability to rapidly synthesize information.
* Now, when you rapidly synthesize information, think about it like rapidly synthesizing food. From 1900 to now, you grew much taller, you grew much stronger, you grew healthier-- all of this happened. But then food became easier and easier to synthesize-- think fast food, and then you have other things happening, which is heart disease, cancer, other issues.
* So what we've got now is the next evolutionary step in data synthesis. We're able to synthesize things much faster. If you talk to a high school student today, a lot of them are using ChatGPT, and not in ways just to make things easy and cheap, but really to understand and synthesize better.
* The second thing ChatGPT has allowed, and this is pretty profound, is it's allowed us to talk to computers in ways like normal human beings. We've always talked like human beings to each other. Only in the last 20, 25 years, we've started using this sort of weird Boolean search speak. And it's even permeated in our day-to-day lives the way we talk about things.
* We're actually the aberrant generation, if you will-- two generations. The generations before the '70s spoke in normal English. The generations after the 2020s and 2030s onwards will interact with everything again in normal language. And ChatGPT is a fundamental enabler of that, which really does change the way one interacts with any computer.
* It's fascinating. You're right. I think about the term "prompt engineer," and people who learn how to speak to things. And now, we start to speak more like ourselves. As we do more of that and as that expands, how do you see generative AI changing how we do things? The food analogy is really interesting-- that was very productive in the beginning and then it kind of went sideways in other ways. Do you see that happening with AI at all?
* It could. It could. Look, we're about to enter into the very beneficial period of increased synthesis-- think of the early first half of the food synthesis and the ease of access of food sort of approach. We're about to enter a place where people are going to spend even less time trying to collate the data, organize the data, and more time trying to apply the data.
* Let me give some examples. I'll start with example, let's say pediatric medicine. Machine learning is very useful in many areas, but it always needs a lot of data. Here's the challenge in pediatrics-- any one hospital will not have enough data to run machine learning significantly on it.
* So now, what you need is to have groups in McGill correspond with groups in Stanford, correspond with groups at Harvard. The challenge with all of these things is that you don't necessarily have enough data in any one location. With generative AI, you can start creating synthetic data. So synthetic data allows you now to quickly leapfrog this weakness in data and suddenly be able to take advantage of modern machine learning techniques, thereby changing things, like in pediatrics, for example.
* Take another example, Waymo. Waymo has had such a big advantage in self-driving, because it is years ahead of everybody in data collection. Today, with generative AI, people are starting creating synthetic data sets to augment their data set and try and catch up faster. So you heard over the last five years, data is the next oil.
* What you're hearing now is that that itself is being challenged, and part of the reason is generative AI. That's just one example of something that generative AI is able to do for you now.
* Can I ask you-- I know you focus on, you say, advanced computing, digital biology, and biotech and cyber physical systems-- cyber physical systems, what is that?
* So cyber physical systems are physical systems that can scale with software. Typically, you see physical systems being cars or something else. And cyber physical systems are more systems that you buy the hardware once and scale.
* So think of a lot of the applications with space. Space is a perfect example of a cyber physical system.
* OK. And I know another one you mentioned about digital biology and biotech. You were talking a bit about pediatrics and how data can help in that area as well. But what other things are you looking at?
* Well, the ultimate Holy Grail for-- there's two Holy Grails for medicine. One is detect the disease before it really impacts you. And the second is, can I personalize the medication as a result of it?
* What you're seeing right now is the beginning, just the beginning, of a profound shift in medicine. And the real shift is we're now able to predict clinical trials much better, so less of them are going to fail. Since less of them are going to fail, we're going to be able to get more drugs to the market for the same dollar value, which means now we can have a much greater degree of personalization on one side.
* On the other side, we're really being able to address the differences between your genetics, your proteomics, your interactome, and let's go all the way to your regulum that regulates your whole body-- the combination of that combined with machine learning to be able to interpret these very vast data sets and changing data sets allows you to personalize. So on one hand, you're going to have a lot more drugs that are targeted.
* And on the other hand, you're going to have a lot more personalization, which is a great use. Because in 10 years from now, I think very few of us will be taking non-personalized drugs.
* It's fascinating. Sorry, just I want to, if I could, just circle back on that-- and I was going to ask you about the time frame. You're saying in 10 years you think that everyone will be having much more personalized drugs?
* I think within 10 years, it's going to start. And it's going to be something that we really think that 2035 to 2045 is the period of time where medicine will fundamentally shift. This is a very complex shift, because you can imagine most pharmaceutical companies are used to making very large batches of things, getting one thing regulatory approved.
* Even the way they make things, even their quality control systems will have to change. So the transition is well underway, but it will take 5 to 10 years for the entire infrastructure to adapt to this transition.
* It's fascinating. Bespoke pharmaceuticals for whoever needs them, and I guess made for your genetics. I know another thing that you are also looking at is with regards to location and location accuracy in GPS. Tell us a bit about that in terms of what you're seeing either with your portfolio companies or what you're watching right now.
* Yeah we're looking at a lot of companies in this area. GPS has been a gift that has been given. none of us can imagine a world today without using Google Maps, or Waze, or something equivalent. And without that, the economy would probably be significantly hurt.
* So the question is, if you've ever looked at your Google Maps when you're at an airport waiting for your Uber, or Lyft, or whatever, you watch that dot move back and forth. The trouble with GPS today is that it is not precise enough in terms of it is not fine enough and it is not reliable enough. If you're in a city, sometimes, it's a 20-meter error. And sometimes, it's a 3-meter error.
* And sorry, not to interrupt, Manish, but tell me why that matters. Because I know for someone like me who's walking around, it's like, oh, it's close enough. But tell me where that precision makes a difference.
* Precision makes a difference. So let's just start-- if you're waiting for your Uber, I think many times we've gone to a position where we're not sure exactly which side of the road Uber's going to come on. If you're DoorDash, and you want to deliver, and you want to deliver optimally, the driver going to the wrong house is problematic.
* If you wanted self-driving cars to be able to stay in a lane more precisely, 3 meters is definitely not good enough, because that's the size of the lane itself. So you have many, many use cases where this is really important.
* And what is the technology that will enable that greater precision?
* So one type of technology for this is a new set of satellites, right? So when you launch a whole new set of satellites, you can now get to low-Earth orbit. So satellites traditionally for these things have been at middle Earth or high orbit geosynchronous orbits.
* If you can come to low-Earth orbit, you can now switch to very cheap satellites. Think about what Starlink and SpaceX did for communications in Ukraine. You could do the same thing for GPS.
* And these are techniques that are now coming online that should allow GPS even down to 1, to 2, to 5 centimeters. And we can talk about these very meaty applications, but there are also simpler applications like AR and VR on your phone that became enabled in much greater degree of accuracy once you get to the GPS of that level.
* I could talk to you for hours. I'm not allowed. I have 2 more minutes, Manish. So I'm going to try and ask you-- I know a couple of other ones were federated learning and planetary health, both important to you. So maybe you could tell me which one you want to talk about.
* I'm just going to talk for a minute about federated learning. Planetary health is really crucial, but it's a tough one to do in 1 to 2 minutes. So federated learning-- look, we all use browsers. I think many of us actually appreciate the targeted ads we get.
* We don't appreciate the fact that our data is always being taken and sold, and we don't really want to pay for the internet. So we want to find alternative mechanisms. Federated learning is a way which allows all the data to stay in remote locations and never be taken to a central site, yet you're able to do the analysis at a level that allows you to get the benefit of getting some of the ads, for example.
* So in a browser, you would have privacy protection, but you'd still receive ads. If you're a complex, large company-- let's just say, a Walmart, for example-- you could now interact with all of your suppliers in a way that you could do much better and efficient inventory management without necessarily having to get data from all of them. So the pediatric example I gave in the past is another great example.
* What if Harvard, Stanford, McGill, others could share all their data for analysis purposes but never have to actually share the data itself? So this is a concept called federated learning or distributed computing plus federated learning. It is a relatively new concept. And it's something we're very excited, about because it brings back the concept of privacy back into the game.
[AUDIO LOGO]
[MUSIC PLAYING]
* Artificial intelligence has taken investors and the world by storm in the past six months, but it's only one of many so-called technologies called deep tech that are potentially world changing. And our next guest is in the business of finding these deep tech opportunities. Manish Kothari is a Founder and Managing Partner at First Spark Ventures, a venture capital fund investing in deep tech who also focuses on high tech innovation in science and engineering. Manish, It's a pleasure having you with us. Thanks so much for joining us.
* It's a real pleasure, Kim. Pleasure.
* So let's start off with maybe tell us about First Spark ventures. You yourself and your colleagues who are involved with First Spark have a long pedigree in innovation technology. Maybe tell us a bit of how it came to be.
* Yeah, so First Spark is really a combination of three different groups. It's a brand new $400 million deep tech venture fund. We invest in companies where the scientific breakthrough has occurred, and there's massive IP advantages there.
* So it's founded by really three core groups of people-- myself, Stanford Research Institute, or SRI International, who is a partner with us, and Eric Schmidt, who has helped-- former CEO and chairman of Alphabet. We invest in deep tech. SRI has been at the foundation of a number of great investments, great creation of companies in the past, everywhere from Siri for your iPhone, to Nuance, to Intuitive Surgical. And we're sort of bringing that sort of thinking to the forefront here.
* Tell me what that sort of thinking that you're bringing to the forefront means. Because one of the things I was looking at, a previous conversation you had with somebody, you talk about the ability to shift the timeline. So with companies and innovation, I'm assuming that means making these things happen a little faster.
* Correct. Correct. So when you look at tech, which is really applications versus deep tech-- deep tech is the things that cause sustainable, long-term, fundamental changes. Think about Nvidia, think about Tesla, think about SpaceX-- these companies really transformed the landscape in a very deep and profound way.
* However, they often take very long to get to the scale of return that's great from a venture perspective. The good news today is what you've seen is that the fundamental component technologies have developed so much that we're now ready to accelerate much faster. And our fund really focuses on providing these startups the tools and tricks, if you will, to be able to go and hit the same sort of outcomes as you see in Nvidia, Tesla, and SpaceX but in time frames that are far speeded up. This would not be possible 5 to 10 years ago, but it is definitely possible today.
* It's fascinating. I mean, again, technology begets technology, I guess, is what you're saying in terms of things moving faster and faster. If people look at the fund itself, you talk about being involved in advanced computing, digital biology, and biotech, and cyber physical systems. I want to speak about each of those so you can educate us as to what that means and what you're seeing. But if we think about advanced computing and AI, just to start with, has ChatGPT been as big a deal as, perhaps, maybe the mainstream thinks it is in terms of advancing AI and bringing it into the world?
* The short answer is, "yes." ChatGPT, or generative AI in general, has really expanded things out. Let's talk about it. There's a few things that it's fundamentally changed, right? Let's start with this concept of data synthesis. So before computers even existed, we used to use books. Well, it was oral, and then books. Then you had encyclopedias.
* And then you had computer systems, and then you had the internet. Each of these things, what they fundamentally did was increase the speed at which you could synthesize information. So one of the things ChatGPT does, or any of these large language models do, is that they really inform our ability to rapidly synthesize information.
* Now, when you rapidly synthesize information, think about it like rapidly synthesizing food. From 1900 to now, you grew much taller, you grew much stronger, you grew healthier-- all of this happened. But then food became easier and easier to synthesize-- think fast food, and then you have other things happening, which is heart disease, cancer, other issues.
* So what we've got now is the next evolutionary step in data synthesis. We're able to synthesize things much faster. If you talk to a high school student today, a lot of them are using ChatGPT, and not in ways just to make things easy and cheap, but really to understand and synthesize better.
* The second thing ChatGPT has allowed, and this is pretty profound, is it's allowed us to talk to computers in ways like normal human beings. We've always talked like human beings to each other. Only in the last 20, 25 years, we've started using this sort of weird Boolean search speak. And it's even permeated in our day-to-day lives the way we talk about things.
* We're actually the aberrant generation, if you will-- two generations. The generations before the '70s spoke in normal English. The generations after the 2020s and 2030s onwards will interact with everything again in normal language. And ChatGPT is a fundamental enabler of that, which really does change the way one interacts with any computer.
* It's fascinating. You're right. I think about the term "prompt engineer," and people who learn how to speak to things. And now, we start to speak more like ourselves. As we do more of that and as that expands, how do you see generative AI changing how we do things? The food analogy is really interesting-- that was very productive in the beginning and then it kind of went sideways in other ways. Do you see that happening with AI at all?
* It could. It could. Look, we're about to enter into the very beneficial period of increased synthesis-- think of the early first half of the food synthesis and the ease of access of food sort of approach. We're about to enter a place where people are going to spend even less time trying to collate the data, organize the data, and more time trying to apply the data.
* Let me give some examples. I'll start with example, let's say pediatric medicine. Machine learning is very useful in many areas, but it always needs a lot of data. Here's the challenge in pediatrics-- any one hospital will not have enough data to run machine learning significantly on it.
* So now, what you need is to have groups in McGill correspond with groups in Stanford, correspond with groups at Harvard. The challenge with all of these things is that you don't necessarily have enough data in any one location. With generative AI, you can start creating synthetic data. So synthetic data allows you now to quickly leapfrog this weakness in data and suddenly be able to take advantage of modern machine learning techniques, thereby changing things, like in pediatrics, for example.
* Take another example, Waymo. Waymo has had such a big advantage in self-driving, because it is years ahead of everybody in data collection. Today, with generative AI, people are starting creating synthetic data sets to augment their data set and try and catch up faster. So you heard over the last five years, data is the next oil.
* What you're hearing now is that that itself is being challenged, and part of the reason is generative AI. That's just one example of something that generative AI is able to do for you now.
* Can I ask you-- I know you focus on, you say, advanced computing, digital biology, and biotech and cyber physical systems-- cyber physical systems, what is that?
* So cyber physical systems are physical systems that can scale with software. Typically, you see physical systems being cars or something else. And cyber physical systems are more systems that you buy the hardware once and scale.
* So think of a lot of the applications with space. Space is a perfect example of a cyber physical system.
* OK. And I know another one you mentioned about digital biology and biotech. You were talking a bit about pediatrics and how data can help in that area as well. But what other things are you looking at?
* Well, the ultimate Holy Grail for-- there's two Holy Grails for medicine. One is detect the disease before it really impacts you. And the second is, can I personalize the medication as a result of it?
* What you're seeing right now is the beginning, just the beginning, of a profound shift in medicine. And the real shift is we're now able to predict clinical trials much better, so less of them are going to fail. Since less of them are going to fail, we're going to be able to get more drugs to the market for the same dollar value, which means now we can have a much greater degree of personalization on one side.
* On the other side, we're really being able to address the differences between your genetics, your proteomics, your interactome, and let's go all the way to your regulum that regulates your whole body-- the combination of that combined with machine learning to be able to interpret these very vast data sets and changing data sets allows you to personalize. So on one hand, you're going to have a lot more drugs that are targeted.
* And on the other hand, you're going to have a lot more personalization, which is a great use. Because in 10 years from now, I think very few of us will be taking non-personalized drugs.
* It's fascinating. Sorry, just I want to, if I could, just circle back on that-- and I was going to ask you about the time frame. You're saying in 10 years you think that everyone will be having much more personalized drugs?
* I think within 10 years, it's going to start. And it's going to be something that we really think that 2035 to 2045 is the period of time where medicine will fundamentally shift. This is a very complex shift, because you can imagine most pharmaceutical companies are used to making very large batches of things, getting one thing regulatory approved.
* Even the way they make things, even their quality control systems will have to change. So the transition is well underway, but it will take 5 to 10 years for the entire infrastructure to adapt to this transition.
* It's fascinating. Bespoke pharmaceuticals for whoever needs them, and I guess made for your genetics. I know another thing that you are also looking at is with regards to location and location accuracy in GPS. Tell us a bit about that in terms of what you're seeing either with your portfolio companies or what you're watching right now.
* Yeah we're looking at a lot of companies in this area. GPS has been a gift that has been given. none of us can imagine a world today without using Google Maps, or Waze, or something equivalent. And without that, the economy would probably be significantly hurt.
* So the question is, if you've ever looked at your Google Maps when you're at an airport waiting for your Uber, or Lyft, or whatever, you watch that dot move back and forth. The trouble with GPS today is that it is not precise enough in terms of it is not fine enough and it is not reliable enough. If you're in a city, sometimes, it's a 20-meter error. And sometimes, it's a 3-meter error.
* And sorry, not to interrupt, Manish, but tell me why that matters. Because I know for someone like me who's walking around, it's like, oh, it's close enough. But tell me where that precision makes a difference.
* Precision makes a difference. So let's just start-- if you're waiting for your Uber, I think many times we've gone to a position where we're not sure exactly which side of the road Uber's going to come on. If you're DoorDash, and you want to deliver, and you want to deliver optimally, the driver going to the wrong house is problematic.
* If you wanted self-driving cars to be able to stay in a lane more precisely, 3 meters is definitely not good enough, because that's the size of the lane itself. So you have many, many use cases where this is really important.
* And what is the technology that will enable that greater precision?
* So one type of technology for this is a new set of satellites, right? So when you launch a whole new set of satellites, you can now get to low-Earth orbit. So satellites traditionally for these things have been at middle Earth or high orbit geosynchronous orbits.
* If you can come to low-Earth orbit, you can now switch to very cheap satellites. Think about what Starlink and SpaceX did for communications in Ukraine. You could do the same thing for GPS.
* And these are techniques that are now coming online that should allow GPS even down to 1, to 2, to 5 centimeters. And we can talk about these very meaty applications, but there are also simpler applications like AR and VR on your phone that became enabled in much greater degree of accuracy once you get to the GPS of that level.
* I could talk to you for hours. I'm not allowed. I have 2 more minutes, Manish. So I'm going to try and ask you-- I know a couple of other ones were federated learning and planetary health, both important to you. So maybe you could tell me which one you want to talk about.
* I'm just going to talk for a minute about federated learning. Planetary health is really crucial, but it's a tough one to do in 1 to 2 minutes. So federated learning-- look, we all use browsers. I think many of us actually appreciate the targeted ads we get.
* We don't appreciate the fact that our data is always being taken and sold, and we don't really want to pay for the internet. So we want to find alternative mechanisms. Federated learning is a way which allows all the data to stay in remote locations and never be taken to a central site, yet you're able to do the analysis at a level that allows you to get the benefit of getting some of the ads, for example.
* So in a browser, you would have privacy protection, but you'd still receive ads. If you're a complex, large company-- let's just say, a Walmart, for example-- you could now interact with all of your suppliers in a way that you could do much better and efficient inventory management without necessarily having to get data from all of them. So the pediatric example I gave in the past is another great example.
* What if Harvard, Stanford, McGill, others could share all their data for analysis purposes but never have to actually share the data itself? So this is a concept called federated learning or distributed computing plus federated learning. It is a relatively new concept. And it's something we're very excited, about because it brings back the concept of privacy back into the game.
[AUDIO LOGO]
[MUSIC PLAYING]