The Exchanges

Q of EVs and all different types of models that are available to lots of different people. And that's the way that this is going to get off the ground. And of course the VW group has Audi and Porsche and VW. And is this sort of the way To get those different models to like electrify these models and then sort of make it available to lots of different people?

A Yeah, I mean, I, I couldn't be more excited about the Volkswagen relationship and partnership. We started these discussions more than a year and a half ago, and it, you know, a deal of this size and its complexity certainly takes time. But, you know, Volkswagen as a group is the second largest vehicle manufacturer in the world, um, very close to the largest and in terms of size, in terms of number of vehicles. And, you know, the reason I started Rivian was to have as much impact as possible and to help drive and accelerate the transition to renewable energy and sustainable transportation. So the opportunity to take our technology and help Volkswagen Group Create products that are highly compelling and that'll pull more customers in across a variety of different form factors, price points, brands, markets, you know, across the planet, uh, was really, really enticing and exciting. And of course the technology we developed is, is really strong. Um, and it's, you know, we, I think there's really two companies in the world that have this level of vertical integration and an architecture that's used. Hardware consolidation to massively simplify, uh, the topology of computers, which is us and Tesla, of course. And so, you know, Volkswagen has really leaned in with us. And so the, the 5.8 billion dollar deal reflects just the scale of this. I think often as we've talked about this, it'…

AI assessment note: “help Volkswagen Group Create products that are highly compelling and that'll pull more customers in”

Q or one of the things that people talk about when they talk about EVs is that, I think it's lithium within the batteries, that maybe one day it's gonna run out, or it's really tough to get, or sourcing it, uh, you know, ends up, if you're gonna buy lithium, it's gonna be, you might end up running into, like, some human rights issues. What do you think about that?

A Well, first, a significant misconception and, um, piece of misinformation in the world is that, Uh, batteries like burn through and use up the materials that are in them. So like lithium gets consumed. Um, that's not the case. So the, there's an opportunity for once you get batteries in the system, you're going to continually reuse the lithium. So it's really a closed loop system. And the value of the batteries is too high to, uh, even imagine a world in which, you know, they're just like thrown away. So there's no world in which you're like, have landfills full of lithium ion batteries. It'd be like throwing away. Um, boxes of gold necklaces. Um, so that's the first point, is it, it, it will absolutely end up with a closed loop, really high rate of recycling. And, and an example of this, which is worth calling out, is a lead acid battery, where lead is not very valuable, but it's more valuable than, let's say, plastic, but the recycling rate on lead acid battery is around 99%. Um, so I would, I would say unequivocally, the recycling rate on lithium-ion batteries will be You're very near a hundred percent, 99.9%, maybe 99.99%. So with that said, in the short term we have to extract enough lithium from the earth to support this, what eventually becomes closed loop cycle for batteries. And, um, there's a lot of places that have lithium. Most of them aren't in the United States, s…

AI assessment note: “with lithium, there's, there's lots of places to get it where there's not questions”

Q I'm going to talk about China in a bit, but let's go to Stephanie. I mean, Stephanie, is it not puzzling at all to you that we have, like, again, these kind of bad headlines, but unbelievable growth?

A Well, Alex, I think it's all about positioning. Uh, first and foremost, in the S&P 500, this is about a 6.4% weighting. It's the big weighting, right? If you're gonna own the weighting, you have to have six percent of your portfolio in this thing. And I would say the majority of people, especially in the spring, were way underweight. By the way, myself included. Totally myself included. I've had like a one percent weighting in the beginning of the year. And then all of a sudden, Like, April and May came, and it was gloom and doom and downgrades, and numbers were getting cut, and there was no line of sight of any growth of any kind for Apple in the springtime. And that's actually when I thought, okay, I gotta start adding to this position, because it was so, people are so negative and so offside. And maybe they had a reason to be somewhat cautious, because the last, the prior four quarters, For really nothing to be so excited about. It was a transition for Q, for quarters. So I added to it because it kind of felt like I wanted to be the other side. I wanted to be the contrarian and everybody was on the other side of the boat. So that was number one. Um, and I totally believed at the time that AI would be certainly a positive. Would it be the super cycle? Is 16 going to be the super cycle? I'm still not convinced of it. But it would be the beginning of a change of tone. So that w…

AI assessment note: “Well, Alex, I think it's all about positioning.”

Q decision meetings, a longer line of managers feeling they need to review a topic before it moves forward. Owners of initiatives feeling less like they should make recommendations because the decision Will be made elsewhere. Truly day two type of culture taking place in Amazon. Um, but obviously the, you know, the company has existing assets that are quite strong. So what's your perspective on the Amazon trade right now?

A Look, I think the expectations for Amazon are very low, and the stock is selling at, uh, a discount to the multiple it sold at in the before times in 2019. And they've never, ever gotten their premium valuation back. And, uh, you can't say the same for the rest of the, uh, mega caps in tech. Uh, Meta looks outstanding right now. NVIDIA looks outstanding. Um, Amazon and Tesla are probably the furthest away from their fifty-two-week highs of that group of stocks. Uh, I think Apple made a new high today. Uh, so, so Amazon is absolutely being treated as though there were issues there. Like, I don't think when they go into this next earnings report, people are expecting any sort of, like, upside surprise or anything like that. So, it's kind of in a weird place, but, you know, from my perspective, that's why there's an opportunity. And, uh, You know, we just went through this cycle with Meta. The stock was in a 75% drawdown. People forget it since I think more than tripled. So when, when companies have problems and they address them and they admit to them, it's like the first step of like, Hey, we have to change something here. So I don't know how long it takes for, for Jassy to get, get his act together. But, uh, I think the stock's gonna work.

AI assessment note: “from my perspective, that's why there's an opportunity... I think the stock's gonna work.”

Q Awesome. So let's say you're quite successful here, right? Let's say you're investing in lots of very successful companies. Good problem to have Does that kind of change the composition of what LG is? Because you're gonna have, like, basically stakes and lots of, like, very promising new companies, and so where do you go from there?

A Yeah. So, let me explain to you why we even started this from the beginning, more specifically than just a vision. So, LG is a, uh, is a, uh, is a collection of, uh, many, many large companies. Actually, LG Electronics is just one of them. LG, you know, LG Umbrella, there are about 64 companies. It changes every year plus minus, but there are about 64 companies, including, including LG Chemical, uh, you know, LG Energy Solutions, and, and all these things. Even, they even have a telco in Korea. Um, there are vastly, vast, the vast majority of those companies are hardware and manufacturing focused, ok? Our job, my job, is to create six fifths and six sixth of those multi-billion dollar companies that are very different from the composition of LG that they have right now. Okay. Much more software focused, much more service oriented businesses. So you ask the question, uh, will LG's, uh, you know, business composition change? Absolutely. If you're successful, and that's exactly what it won't happen. Doesn't mean manufacturing is going to go away. It's going to still be there. It's going to grow, but we're going to have additional and new, uh, uh, business lines that it's going to probably grow much faster.

AI assessment note: “will LG's, uh, you know, business composition change? Absolutely.”

Q So talk a little bit about how Ozempic counters this insatiable hunger that we end up being left with from processed foods, and what's your experience, like, been like in the year and five months that you've used Ozempic?

A Well, so if you ate something now, it doesn't matter what it is. After a little while, your pancreas will produce a hormone called GLP one, along with some other hormones. And GLP one is basically your body's natural signals going, Hey buddy, you had enough, stop eating, right? It's the brakes. But natural GLP-ONE only stays around in your system for a couple of minutes and then it's washed away. What these drugs do is they inject you with an artificial copy of GLP-ONE that instead of sticking around for a few minutes, sticks around in your system for a whole week, which has this incredible effect. I'll never forget. Second day I took Ozempic, I was lying in bed just behind where my laptop is now. If I turn my laptop around, you'd see it. Um, and I woke up. And I thought, huh, I feel something strange. What is it? And I couldn't locate what it was, what I was feeling. And it took me about five minutes to realize I had woken up and I wasn't hungry. It had never happened to me before. And I went to this diner just up the street that I used to go to every morning. I'm slightly embarrassed to say I ordered the thing I used to get every day, which was a huge chicken mayo sandwich with loads of chicken and mayonnaise in it. And normally I would eat that really quickly and I would still want some potato chips. And I had like three, maybe four mouthfuls of this sandwich and I, and I wa…

AI assessment note: “What these drugs do is they inject you with an artificial copy of GLP-ONE”

Q And if that happens, I mean, you could abruptly, you could abruptly have people thrown off the medicine, whereas like they can afford the compounded drug, they cannot afford 1300 a month. So what happens if you get abruptly tossed off the drug?

A Well, it's a really important question. So, uh, I want to explain what the drug companies say, and I want to explain why they're probably right, but we need to take it with a small pinch of salt. What the drugs say, what the drug companies say is these drugs are like statins or blood pressure meds. They work for as long as you take them. And when you stop taking them, they stop working. So you regain the weight pretty rapidly. And they have funded studies that, a study that seems to demonstrate the vast majority of people regain the vast majority of their weight within a year. Um, there's a small pinch of salt in that, in that, of course, they have a vested interest in telling us that we have to take them forever because they make money forever. They do seem anecdotally to be some people who take them for a shorter period, use them to radically change their habits, and then seem to stay at a lower weight, although I suspect they're a minority. There'll be more research on this soon, but yeah, if you're thrown off, you'll almost certainly regain the weight.

AI assessment note: “if you're thrown off, you'll almost certainly regain the weight.”

Q there. But the thing that really struck me as the risk was, um, You talk about how it could potentially dull people's, uh, emotionally a little bit to pleasure and reward, and like physical rewards. Uh, for instance, like people start to drink less, um, but there's concern that it might just make everything in life less pleasurable. Uh, can you talk a little bit about why that might be?

A So I want to stress this is all the scientists who've expressed this concern, expressed that it's speculative and we can't be sure. So some scientists have argued that these drugs seem to be causing depression or even suicidal thoughts in a minority of people who use them. That in itself is contested. It did not persuade the European Medicines Agency, although the FDA in the United States does have a warning about potential suicidal ideation from these drugs. And there's a huge debate about why that might be. Some of it relates to the question about the brain effects. So a different theory about how these drugs may work in the brain and on the brain is that in your brain you have something called your reward systems. The reward systems are what Motivate you to do anything that propagates life, right? We eat because we get rewards from it. We have sex because we get rewards from it. We engage in social behaviors because we get rewards from them. They all make the reward centers of the brain hum. One theory about how these drugs work is they may be dampening your reward systems. So it may be that I no longer Feel the urge to eat a Big Mac over a salad because I get less reward from the Big Mac than I did before. There does seem to be some evidence that there's lots of people report losing their pleasure in eating after taking these drugs. Now, although not everyone, and I actuall…

AI assessment note: “One theory about how these drugs work is they may be dampening your reward systems.”

Q Yeah. I was gonna ask you what the path forward is given that we know that, okay, this is something that's these models are struggling with. So you think that this is the path forward that the way that you just described?

A You look at the innovation that's taking place from, from players out there. Companies like wave in the UK have put together a really amazing, uh, underlying technology that lets the front facing camera be interpreted in a way that the system can communicate to the occupants of the vehicle, what's going on outside the car. And also use that then as a way to determine what is the car going to do? How is it going to steer, accelerate, or brake? So it can interpret that video feed and explain that there's somebody jaywalking, or somebody ran a red light, or there's a child waiting to cross, or whatever it may be. So this generative AI approach is really, I think, going to accelerate the adoption. Very recently at a conference called CVPR, so that's Computer Vision Pattern Recognition. It takes place annually. Uh, last month it was in Seattle, Washington, and there was a competition that they held, so it's a research, uh, conference, over 400 entries into this autonomous driving challenge, and it was basically looking at sensor data and trying to predict the best trajectory for the vehicle moving into the future. NVIDIA submitted and won the challenge. Our research team had developed a new large language model, basically end to end training of that sensor data system for then controlling the vehicle. And so over 400 entries, NVIDIA came out on top with this new large language model…

AI assessment note: “this generative AI approach is really, I think, going to accelerate the adoption”

Q to me that NVIDIA is developing a lot of this technology. Like I went to your automotive section on the website and was like, wow, there's a tremendous amount of models coming out of NVIDIA. I thought it was largely the car makers, like the Waymos or the Teslas that are developing the autonomous technology. So how involved is NVIDIA in developing these models itself? And then who's the customers?

A That's a great question. Um, we work with hundreds of automakers, truck makers, robo taxi companies, um, software startups, the sensor companies, the mapping companies. It really is quite an ecosystem that we've built. We're not creating the vehicles, but we work with those manufacturers, and so we offer the compute hardware. That's our drive platform, so that's the brain that goes inside the car. Our drive OS is the safe operating system that's part of that package. We have a lot of different middleware and libraries that they can use to develop their applications, algorithms, the neural networks. That application layer, though, is generally built by our customers, so Mercedes-Benz or Jaguar, Land Rover, Volvo, Neo in China, and so they can pick whatever parts of the software stack they want, and in many cases, our customers are taking the whole stack and they're developing Um, some of their own algorithms as well. So there might be, uh, a pedestrian detection algorithm from Mercedes running along a pedestrian detection algorithm from NVIDIA. And, uh, and we collaborate on that. So starting, um, through the end of this year, the introduction of the, the new CLA, uh, it's already been announced from Mercedes. That's the new Mercedes model, the C-Class. And, um, So every Mercedes will be built on NVIDIA drive with the software that we've developed and rolled out by NVIDIA. So it…

AI assessment note: “we work with hundreds of automakers... Mercedes-Benz or Jaguar, Land Rover, Volvo”

Q So then how important is simulation and training? Anything having to do with autonomous driving?

A That simulation plays a really, really big role in ensuring the safety of the system. There's really no way that, first of all, driving around collecting data, you rarely are going to see the dangerous scenarios, the hazards, the things that, you know, very rarely occur. You're not going to capture them on your data collection. So we need to use simulation and really what we call synthetic data generation to create those kinds of scenarios. So we can Create fake potential hazards, things falling off of trucks and people running across the street at night, um, whatever it may be, somebody running a red light. And so we can create that data to augment the real data for training the AI. And then we can actually simulate all these dangerous scenarios to ensure that the system will do the right thing. And the benefit of using simulation too, it's repeatable. So we can adjust the software and test something that maybe didn't pass a month ago, but we can run it through the same scenario and say, oh yeah, we fixed that. Um, there may be situations, you know, it often happens that the sensors, um, are blinded by the sun, right? As it's setting, right? The sun is like coming right into the, the eyes of the, the car, into the driver's eyes, into the camera's eyes. And so you only have a few minutes a day where you can actually capture that data as well as test. And so in simulation, it ca…

AI assessment note: “simulation plays a really, really big role in ensuring the safety of the system”

Q can today. And it doesn't sleep. So I guess I'm trying to think why, why, how do you end up getting from a point where a company is trying to get an ROI and all of its investments in this area and reaping those economic gains to a point where there is this, um, AI for all and decentralized version of the technology. Is it just that everybody catches up?

A Alex, there's good news here, um, in that we have seen not a single company, but we've seen five, six, seven, eight companies all in parallel delivering similar generative AI capabilities, and they're all priced either free or like 20 bucks a month, right? No one was so far ahead that it's like, this is a million bucks if you get a chance to use it, right? So the fact that there are multiple players Um, and they're all starting at or near zero is a super positive indication of the future. I don't think you're going to have, uh, people all of a sudden, you know, all eight of them getting into a cabal and saying, okay, let's all charge a thousand bucks a month because we can. Uh, you also have great business models like Google that is, and Baidu, which has, you know, been Available for free around the world. Uh, and there, these companies are making money on secondary and tertiary implications of that free business model. And I think you're going to see that here with AI as well. I think AI will become ultimately the best educator on the planet and the best diagnostician on the planet available on your mobile phone. And I think that educator Again, the best on the planet will be available for free to the poorest child and the wealthiest child. And it won't be like, it's a little bit better for the person who's wealthy. Just like Google for Larry Page's kids is not different than …

AI assessment note: “these companies are making money on secondary and tertiary implications of that free business model”

Q Okay. We've talked so much. We've used the word abundance so much today, and you have this abundance summit. You've talked about getting to abundance. It's one of the things you talk about a lot, but I, for those who haven't heard about it before, what, what do you, what is abundance? Like, what is your definition of it? Do we have it?

A We have it to an extraordinary amount compared to humanity. 50 years ago, a hundred years ago, a thousand years ago. Abundance is access to the fundamentals of life. It's not about a world of luxury. It's about a world of opportunity. I define technology as a force that takes whatever is scarce and makes it abundant. Let me give you some examples. So we used to kill whales to get whale oil to light our nights. That's how we got energy, right? Then technology allowed us to, uh, ravage mountainsides and get coal. Then we drilled kilometers under the ground for oil and natural gas. Now we can use photovoltaics and soon perovskite to convert sunlight into electricity. And there's 8000 times more energy that hits the surface of the earth than we consume as a species in a year. So this energy is there. It's just not in a usable form yet. And so technology takes what's there and not usable and makes it usable. Another example is water, right? We talk about water wars or water scarcity, but we live on a, on a water planet. Right. We have two thirds of our planets covered by water. We've, but unfortunately it's 97.5% salt, you know, two percent is ice and we fight over half a percent. We just launched an XPRIZE. I'm very proud to be chairman of the XPRIZE. We have launched 30 XPRIZE in 30 years, about six hundred million dollars in competitions. And we just launched one hundred and twen…

AI assessment note: “Abundance is access to the fundamentals of life. It's not about a world of luxury.”

Q you have the older, older people, um, who are gonna need, need to be cared for. And, um, it just seems like we're gonna deplete the planet, but you also care about the planet. Obviously you've done this carbon challenge. So how do we, how do we deal with the fact that we could have overpopulation, but also a sustainable, right? Yeah. Massive increases in population, but also sustainable planet.

A So the first thing is to look at the data. And the data shows us that we are about to have not an overpopulation, but a massive underpopulation of planet earth. So if you look at the growth curves, um, 50 years ago, I'll call it the 19 fifties. Uh, it was about an average of 5.7 children per family around the world. And we were in massive growth and books like the population bomb were written and so forth. But the reality is, We are now, we've dropped from 5.7 children per family down to below 2.4 globally in the U S we're below the replenishment level, which is 2.1. And China's way below Japan's way below most of Asia, most of Europe, Italy is dissipating going away. And it likes like .7. Um, the only, Only region in growth is Africa. And as Africa becomes more affluent, as better comms and AI and technology comes there, we'll expect to see that growth level drop as well. And if you've got a massive underpopulation, right, we peak at nine, nine and a half billion and have a very rapid drop off. Um, our population is tied directly to, to GDP towards the ability to maintain our, our world. I think the only countervailing force there is going to be humanoid robotics and AI that's going to allow us to support that aging population. Uh, so I think, you know, Elon goes on and says, listen, have more kids. He's got 10 or 11. I forget how many exactly.

AI assessment note: “we are about to have not an overpopulation, but a massive underpopulation of planet earth.”

Q for you. So basically you can write the prompt of what you want and the code will be developed on the backend, uh, or will help you devote, develop the code on the backend. Is there going to come a point where we're not going to need code at all to build? Because like, if you can just prompt and why do we really need to be in the code?

A You might say we're already at that point to some degree where, you know, when you ask and go chat GPT, ask and go and ask chat GPT a question, uh, you get an answer and you must, when you ask a question to plot the chart, for example, or do a method, method, mathematical, mathematical, um, calculation. It, um, it actually. Generates a Python script that then, you know, plots that data into a chart and it shows you the chart and it still shows you that step where you see in between. Uh, the Python script, but, uh, you, they could just hide that. Um, and you just see the, see the chart output, right? Like in, in many ways, ChatGPT is giving you an answer without you ever having to worry how that was generated. And so, yes, we're going to see computer systems where large language models are just one building block in addition to code, or maybe it's multiple language models and image models and, you know, time series models and whatnot, plus code combined. Um, To, to generate, um, all the output, um, that the developer or the user expects. Um, do we still need engineers then to code? Yeah, because, well, first of all, you know, there's billions of lines of code out there that still have to be maintained. You know, one of the examples I like to get is that most banks are still running Cobalt code. That's a program in English from, uh, invented in the late fifties when Eisenhower wa…

AI assessment note: “You might say we're already at that point to some degree”

Q no. But that is, we're talking about generative AI in particular. So we can go on for days about how AI has been, you know, for feed ranking and computer vision, fine. But the big moment right now is all about generative. And generative has been something that GitHub has written with Copilot to this amazing moment. But that's the question. Is that type of technology in particular transferable elsewhere?

A I think one, one scenario that comes to mind, um, uh, that we are already using at GitHub is, um, support. Um, and so, um, you know, if you look at our support system, uh, today, you actually find, uh, the GitHub support copilot, um, that tries to help you before you, you know, submit your ticket to a human. And we actually see, um, and it generates answers. So it's generally fair. It uses the same last English models to, to stay within the scope of your question. And we see that the number of, um, tickets that get solved that way, um, Uh, is, uh, about 50%. So, you know, 50% of those questions that go through the support co-pilot get solved by the support co-pilot and do not get submitted into, to a human. And so as such, it makes, you know, supporting, uh, our, our developers, um, um, our customers, um, uh, more efficient for us as a company. So I'd say, you know, that's, that's definitely another scenario where we see the, um, um, efficiency gains for us as a company and other, similarly, You know, we have an internal tool called Octobot, uh, you know, like Octocat or our logo, um, my t-shirt, and it, it helps, uh, you know, our folks internally, um, to, to solve IT problems. And our IT team is getting, you know, more than three hours per IT supporter back through that internal tool by, just by, you know, helping, um, uh, uh, uh, employees, you know, to solve their own IT is…

AI assessment note: “one scenario that comes to mind that we are already using at GitHub is support”

Q no. But that is, we're talking about generative AI in particular. So we can go on for days about how AI has been, you know, for feed ranking and computer vision, fine. But the big moment right now is all about generative. And generative has been something that GitHub has written with Copilot to this amazing moment. But that's the question. Is that type of technology in particular transferable elsewhere?

A I think one, one scenario that comes to mind, um, uh, that we are already using at GitHub is, um, support. Um, and so, um, you know, if you look at our support system, uh, today, you actually find, uh, the GitHub support copilot, um, that tries to help you before you, you know, submit your ticket to a human. And we actually see, um, and it generates answers. So it's generally fair. It uses the same last English models to, to stay within the scope of your question. And we see that the number of, um, tickets that get solved that way, um, Uh, is, uh, about 50%. So, you know, 50% of those questions that go through the support co-pilot get solved by the support co-pilot and do not get submitted into, to a human. And so as such, it makes, you know, supporting, uh, our, our developers, um, um, our customers, um, uh, more efficient for us as a company. So I'd say, you know, that's, that's definitely another scenario where we see the, um, um, efficiency gains for us as a company and other, similarly, You know, we have an internal tool called Octobot, uh, you know, like Octocat or our logo, um, my t-shirt, and it, it helps, uh, you know, our folks internally, um, to, to solve IT problems. And our IT team is getting, you know, more than three hours per IT supporter back through that internal tool by, just by, you know, helping, um, uh, uh, uh, employees, you know, to solve their own IT is…

AI assessment note: “one scenario that comes to mind, that we are already using at GitHub is, support.”

Q like something goes wrong because stuff always goes wrong on the internet. They're able to identify it and respond. So if you could take us through the process a little bit, just to give us a sense as to what a network will do in order to be able to keep that connection crystal clear, it would be really interesting. Like what happens when there's a problem? What happens next?

A Yeah. So, so first of all, I'll contrast it with television. Televisions are, are dumb receiving devices, right? They just receive signal and then decode it onto the screen. That's it. There's no, there's no back channel. There's no conversation back of, did I get it? Did it work? Television just assumed that it's gonna work. In streaming, everyone's device that you watch a streaming on, any, any IP delivered device has the ability to send data back. So the good thing is, is now the communication becomes bi-directional, which is critical because anyone who's pushing any business, any business over the internet, whether It's a video product like NBC with the NFL game or even your apps, right? Everyone has the same challenges when they go to buy shoes on an app as you'll hit delays and you'll hit buffering. The good news is all those devices are throwing data back to the source of the people that are trying to serve that digital product. So Conviva essentially captures all of that sessionized data, leaving the device, basically telling NBC during the game what the consumer's experience is as they're watching it. So NBC is looking at the twenty million devices, again, aggregated from Conviva. We collect all the session data from the client side and we ship it back to NBC in a, in a NOP dashboard. You talk about operations, network operations. So they're able to look at their, thei…

AI assessment note: “we ship it back to NBC in a, in a NOP dashboard”

Q making this move. Uh, where it's trying to build for, I think for the first time since GPT for its own model that competes with it. How should we read into what Microsoft is doing there? Do you think it is a loss of faith in open AI? It's a hedge against open AI, sort of an unnecessary move, even though it has such an important partner. What's your take?

A Yeah. One of my friends who works at DeepMind told me that Microsoft is basically reversing what Google has managed to do over the last few years. And in fact, making the same mistake that Google initially made, which was to have its training distributed, uh, or split up between two different corporations or institutions. For Google, it was a brain, Google brain and DeepMind. And so Microsoft has a company which is in the lead, right? OpenAI. And I guess instead of doubling down on it, they're trying to hedge their bets in this way. I think if you think that open AI is like another product where you have multiple vendors, so you can be sure that if one of them, uh, you know, has, decides to go a different route, you have some leverage over them. That, that might make sense for another kind of product. The thing with AI is if you buy scaling in this picture, that is, you make the models bigger, they get much smarter. Then I don't think it makes sense to hedge your bets in this way. I think you should just double down, give, give one of them a hundred billion dollars and just say like, go make me, go make me super intelligence. You know what I mean? Like, cause then you're just splitting up your efforts and yeah, like it would be much better to have one GPD four than two companies that have, uh, you own two companies that have a GPD 3.5.

AI assessment note: “making the same mistake that Google initially made, which was to have its training distributed”

Q it's going to take to make this stuff work and Are we ever going to get to the place where a lot of these people want to get to talking about like adding more compute and data and energy, and eventually you get to the point where you can train better large language models and see what the scaling law really looks like at its limit. What do you think?

A Yeah, I think compute will be less of a bottleneck than energy. At sort of the seven trillion, I, yeah, I imagine, um, well, the backing up, the reason I think compute will be a less bottleneck than energy is because right now you have one company, NVIDIA, which is making the sort of, uh, GPUs, and other than Google, nobody has a clear competitor, and so the, the thing that was bottlenecking NVIDIA so far is that some of their The components that the need for these GPUs, uh, COOS and HBM, uh, they just weren't able to get enough allocation or get, uh, TSMC to build facilities for these, because TSMC was like, I don't know if we buy all this AI stuff, but, uh, because then they had to make this huge investment into building it out, but now it seems like the, uh, fabs are building it out, and also all these companies have accelerator programs where they're gonna try to ship their own chips, so I think Compute will become more and more available. And that's what Zuckerberg said on the podcast that now the compute constraints are decreasing. Then the question that Zuckerberg pointed to was, well, will there be energy? And the, the key constraint with energy is not necessarily, is there enough energy in the world, but more so for training, is there enough energy in one place? Because to do a training run, it has to usually, at least from what it seems like publicly, the training met…

AI assessment note: “I think compute will be less of a bottleneck than energy.”

Q So how do you do that? Because that's an interesting, you know, your ability to figure out how old people are is an interesting.

A Yes. So, so we are using AI. Uh, we have a, uh, a partner called Yoti that provide a third party API that we are using, and you will enter, you will enter a date of birth when you sign up, but we want to be sure that you enter the right date of birth. So we are going to scan your face and estimate your age to compare it with the date of birth you enter. And if it matches the date of birth, then you can enter the platform. If it doesn't match, then we will ask you to provide more detail, like identity documents, a video of yourself, to ensure that this is you, you are the real person, and this is your real age, to ensure that you will go in the right age group. Because in real life, and when you, you, you, you meet new people, usually they are from the same age.

AI assessment note: “we are using AI. Uh, we have a, uh, a partner called Yoti”

Q My question is, so you've built these models with billions of weights. So how does, how do you include the weights in the models? Talk a little bit about that training process. So if the model is going to understand the relationship between these different entities, what do you do to teach it that?

A Um, so, so what we do is we initialize random weights, and then we, like, start to do something called, uh, gradient descent, which is you predict a word and you compute the error between the word you predicted. The model predicted some, like, random word. You tell it you wanted to predict another word, and then you do, and then you try to teach, you update the weights, you update the numbers inside of the weights to basically converge to higher and higher precision. And so this process usually is what we call, it's, it's what happens during pre-training, but we update those weights, um, in order to minimize the distance between what your predicted value is and what the actual value is. Um, and that process usually does some level of convergence and we train it over thousands of GPUs and trillions of tokens. So if you look at something like, um, the eight billion and, and, uh, seventy billion, They were trained on almost 15 trillion, uh, tokens and tokens roughly you can imagine as a word. So roughly like, um, 15 trillion words, which is an incredible outcome. And it requires thousands of GPUs to train it on. And a GPU is, is, you know, um, uh, roughly the cost of like an Audi is what I call it an Audi A three or something like that, but they're very, very expensive. And so like being able to operate these large scale infrastructure, um, Uh, training jobs is quite a feat of bot…

AI assessment note: “we initialize random weights, and then we, like, start to do something called, uh, gradient descent”

Q Are there any other ways that AI work in your products that I didn't touch on?

A Sure. Uh, we have a few. Uh, so we, uh, use AI in shine on our photos app. To, uh, to do things like deduplicate to overall, uh, understand, um, and, and understand when you might be doing something that's share worthy, either on a photo level or on an album level. You mentioned earlier, the suggested albums. It's one of my favorite features. The fact that the AI can figure out where I tend to take interesting group photos and where I don't and make those kinds of suggestions to me is something that relies on AI. In our events app, we're using generative AI, so different type, um, to make beautiful invitations, and we do have gotten rave reviews for our invitations. They are, my view is they are the best out there in terms of really helping build a mood and a theme and anticipation for your event. Uh, and in things like Sunshine Contacts, we're deploying, ah, text recognition and pattern recognition to overall do extraction around signatures, so we can figure out which of your contacts is this person, and we can ultimately pull in their professional information, phone numbers, ah, and all of those pieces to really make your contact that much more up to date, ah, and enriched.

AI assessment note: “we use AI in shine on our photos app... In our events app”

Q And that's now part of like, uh, the press releases that they can optimize, but they've been doing it forever. Is there anything about generative AI that's exciting for you as something, uh, running a business that has a lot to do with efficiency? Of course, some relationships, but You know, can you build on top of like a GPT model and do anything that you couldn't have done beforehand?

A We are, yeah, we, we, we've actually using, um, OpenAI's APIs as well as a company called Adept, uh, which their CEO used to be the head of engineering for OpenAI, um, and so they, but we're using these, the thing that's working for us is that the fact that we've taken, you know, a complex transaction like shipping a container from door to door and broken it down into a series of small discrete tasks, uh, in our workflow system that we build ourselves, then those tasks are simple enough that This generative AI can complete these tasks. Um, and that, I think if you just told open AIs, I'm sure that if you just said, hey, ship this container from here to here, it would, it would hallucinate. It wouldn't do anything useful. Um, but when you say, hey, extract this data from this website and put it into this data format, into this database, or parse this carrier contract, we get these container, uh, these contracts in Excel format with like 20,000 rows and tabs and all these if then Subject to charges. Uh, and actually, uh, AI's quite good at doing the simple things like this and putting that into your database in a structured way, so it's saving us a lot of money, time, being more accurate than humans. Uh, we're finding major progress. I think our, our team is, uh, not all of this is gonna be generative AI, to your point. Some of it's like more traditional ML. Some of it's just lik…

AI assessment note: “we've actually using, um, OpenAI's APIs as well as a company called Adept”

Q because they have gone from this company that everybody like hailed as like the, you know, the beacon of innovation, 20% time. Uh, which was I think before your time, but to one that really can't get its act together shipping. Is it, is it that they don't have the right talent? Is it the process there? Um, is it leadership? What would you say is going on within Google?

A I certainly don't think it's because of talent. Um, I mean, you have to be insane to think you have more talented people than Google. I don't even think OpenAI has that, even though, like, they have a massive concentration of talent density today. Um, I think it's more to do with the bureaucracy, uh, and the culture. And, uh, leadership. These are the three main reasons why they're not able to do it. But in our case, um, the argument goes even further, which is probably the fourth thing and the most important thing is even if they wanted to do a product like perplexity and roll it out on google.com, that domain, the ultimate domain of the internet, they cannot do it that fast. They cannot do it even slowly. It's very hard because, uh, The bedrock of their business model is sending traffic to other people, uh, sending, like, link clicks, link views, so that's why there's two terms in their business model called cost per thousand views, CPM, and cost per click, CPC, and as an advertiser, you're bidding on certain keywords based on the CPM and the CPC, and there's an auction model, and like, It arbitrarily drives up, and Google Analytics gives you, like, which keywords are getting a high CPC and CPM, and the number of views and things like that. So it's basically, that is their business model. Like, the, the, the, the TenBlueLinks user interface. That's literally where they've bui…

AI assessment note: “I certainly don't think it's because of talent... more to do with the bureaucracy”

Q down, uh, the pike this year, just like the way that they're talking is that it sounds like some serious advances are going to happen this year. Do you have any sense as to what those might be? And I don't understand like what, what could, What could inspire such talk? Is it much? I mean, if it's a slightly better GPT model, okay, but it sounds like something more.

A I mean, I don't know what, what has been said publicly. Maybe you, you hear something more privately, but, uh, I guess I'll just go by what Sam Altman told Bill Gates, and I was listening to it like two days ago, which is, um, they'll have a lot more reliable models. So Uh, right now, sometimes, one in a 10, or one in a hundred completions of a GPT-IV model, uh, is hallucinatory, can, can say arbitrary things. It's not reliable, it's not deterministic programs, so I think they'll address that. Uh, they'll also address the fact of the models being, like, more multimodal. Like, right now it supports images and input, but the most general version will support audio, video, as inputs and also as outputs. Well, uh, interface with text. I think they'll make some advances there. I think the third thing they'll make some advancements on is, um, reasoning. Uh, whatever reasoning GPT-IV is able to exhibit today is already amazing, but it's still pretty limited, and I think they'll make some more progress there in terms of multi-step reasoning and, like, doing back and forth thinking. You want models to be able to, like, think for a while and then come back and give you some things, right? So they, they probably make more progress on those things, and I guess the fourth thing that, um, Altman mentioned was, like, just, like, Reducing the costs. Oh, sorry. Increasing the personalization of…

AI assessment note: “I'll just go by what Sam Altman told Bill Gates... they'll have a lot more reliable models”

Q With anthropic, um, what do you, what do you think about the, uh, the Claude model and does it hold a candle to GPTs?

A So far not to be very honest, but I'll tell you one thing. They, there are people in the world who actually like Claude a lot more than GPT in terms of response styles and eloquence. They like Claude a lot more. It's more natural. It feels more like talking to a human than to an AI, if you talk to Claude. So there are, we have, we offer Claude as an option on the perplexity product for pro users. And I know a lot of people who still use Claude instead of GPT-IV, even though GPT-IV is a more capable reasoner, because they like the way Claude responds. So Anthropic certainly has something. Now, Can they create a model better than GPT-IV? I definitely think they can, and if they don't, they're kind of doomed, in fact. This year, in 2024, if they don't create a model better than GPT-IV, all the funding and the, like, you know, the race is not being put to good use. But I actually think they will end up creating a model better than GPT-IV this year. Like, it, it, it's sort of almost guaranteed to happen. So I believe it's going to happen with CLAW-III. Now, Does that mean OpenAI is in trouble? I don't think so either. I'm sure there's a GPT 4.5 or five that will stay ahead. So it really is going to be a cat and mouse game there where Anthropics playing catch up and OpenAI is ahead through multimodal capabilities, reasoning capabilities, and things, things like that. Now the question…

AI assessment note: “So far not to be very honest, but I'll tell you one thing.”

Q How about naming? Yeah. Do you want to name one?

A Yeah. Let me take an example, our segment anything model, which is a little bit different than our, than our Lama model. Um, but, but I think has been the one that has been just incredibly impactful in terms of people quickly building on it. Our, our segment anything model is one where you take an image and it gives you a detailed segmentation. Of that, uh, of that we released it back in April, including a lot of, um, tools and data to, to go along with it, and, and within days we had people who had built up applications essentially for, um, Conservation applications, so being able to track down some species who may be endangered, using that to, to follow them. We had people use it for the treatment of medical images, so segmenting cells from some, some of these images, and it's been wonderful to see that, that explosion, that explosion of work. Um, on the language side, we also saw many people build up all sorts of, Um, different tools. And in particular, the, the work that we're most excited about is, uh, the work on efficiency, to be honest with you. Um, there's so much that we can do to make these model more compact and, and efficient and, and, and running, um, really, really fast with low energy. And I think that's one of the things that I've been most excited about seeing. There's lots of other applications too.

AI assessment note: “Let me take an example, our segment anything model”

Q mean, it does seem like what you're saying though, is that this Landscape is going to be a lot more competitive than it has been previously. I mean, you have a company like Nvidia that added, what, six hundred billion dollars to its market cap in one year. Like, there are going to be others that are going to be trying to get in. Does that sound right to you?

A Absolutely. You know, and I sort of put them into two classes, uh, Alex. There's going to be those that build their own, you know, and that's what you see Amazon, Microsoft, Google are doing. They're going to say, hey, I'm going to own this and do this myself. And then there's going to be the general providers in the marketplace, and that's going to be Intel, AMD, Nvidia, I think will be the three big ones in that space. So there's going to be do it ourselves. We're going to own the full stack of hardware and software. You know, which is the big cloud guys. And then there are those who say, Hey, I'm going to sell my chips to everybody. And I expect those would be the big three.

AI assessment note: “Absolutely. You know, and I sort of put them into two classes”

Q Okay. So wait, which, which part are you going to compete in then both or both?

A Yeah. You know, we're going to, you know, cause I want to be a foundry to what Amazon does, what Microsoft does, what Google does, and I'm going to sell my chips and I'm going to sell my chips to the enterprise customers who want to do this with their data on premise. As well as to the big cloud guys as well. And today, you know, biggest customers for Nvidia today are probably Microsoft who's putting up their big farms, but they're saying, Hey, no, I'm going to build my own ship. I'm going to build Maya so that I do just like what Google is doing with their own TPU as well. I want to own that margin and I'm going to do it on my architecture as well. So Intel, I think uniquely has two bites of the apple here to pursue.

AI assessment note: “I want to be a foundry... and I'm going to sell my chips”