Board of Changemakers: Shou Chew, Jack Hidary, Benjamin Horowitz & others on #AI at #FII8

All right, I have the very unenviable task of doing this in 40 minutes when each of you deserve two hours on your own. I'm going to take notes from the conversation this morning and go for each of you, and wish me luck, please. Travis, I'm going to start with you. Founder of Uber, CEO of Cloud Kitchens. We talk about AI a lot and all about the super fancy, big, billion-person transformations. How is AI going to change business right now and transform and expand and change profitability of businesses? You're doing some amazing things with Cloud Kitchens and AI.

Please, I mean those are two different questions. If you just do, when I look generally at what's going on right now, it feels like we're entering an era of bigger is better. That if you are a large company with a strong business model, that the technology that's coming is actually going to give you competitive advantage over folks that are smaller. The accretion of value is going to the bigger players, and there are probably a lot of very simple things that can be done to create more profit and bigger remuneration on those business models, is my guess. So there's just like blocking and tackling, like workflows, customer support, onboarding, sales, all getting automated will make an already strong business model like supercharged. And I think for the biggest companies, create trillions of dollars of value. I mean, we're doing, you know, my company's doing the future of food. It's real estate, software, and robotics for food. And there's lots of fun practical things that are happening. I mean, they're going to be big, but they're also practical. You know, we have a machine that makes food. Like, imagine if you were to go to Chipotle, that front line of people that is making your bowl, we have a machine that does that. So we have restaurants that are operating with no people or one person just overseeing the machine. But if there's only one person in that room, you know, you got to, we now have that person talking to the machine about how the machine is doing as if they were an employee. So yes, so then it's this is simple stuff, but this is where we're going now. Like we have this up and running in restaurants today where you're asking the robot, hey, is the food warm? Am I running out? This kind of thing. But we just said, well, you know, might as well just give it a personality. So now you can ask like, did the Dodgers beat the Yankees today, which they did. And you can have a conversation and it's sort of, we're quickly getting into that Interstellar moment where you're just interacting in a very human way.

I've got to ask this other question just for fun. If you'd remained CEO of Uber or were still CEO, what would you have done differently? What would Uber be doing now?

Oh man, we could take the whole hour on this one. I'll keep that one short. May I'll give you a couple things. Look, we had probably one of the preeminent AI labs when I left Uber, and we saw the seminal paper out of Google, the attention paper. I mean, you guys know how I rolled when I was doing Uber, like we'd probably do some pretty interesting stuff on that front. I mean, sort of in a more practical, on the more practical side, like we were starting to roll out, I mean, of course we're doing the autonomy thing. I'll put that aside, I think Elon's going to take care of that at this point. But we were starting to roll out almost, we looked at transportation as a high-frequency trading business where we would buy trips from drivers and sell it to riders. We were a market maker. And so we were starting to set up quant teams for each city to do the high-frequency trading of the trips so that we could get the lowest cost, reliable ride, more efficient, gain market share, get more profits, all this. It was just kind of a fun, geeky finance thing.

Right. One in 30 seconds, you've disrupted two huge markets. Your one piece of advice for someone here who wants to go disrupt a marketplace today, how do they, what's your piece of advice for them?

You know, look, I've over time I've tried to distill what it means to innovate at speed and at scale into sort of principles or pillars. They basically are cultural values at my company, and so we've distilled it down to truth, trust, and passion. And again, you know, those are high-level sort of ideas, but doing those really well, each of those really well, is sort of how you innovate at speed and at scale. And it's, you know, there's a lot in there, but like think of the childlike playfulness and curiosity with a teenage rebelliousness with an old man's wisdom. That's a beautiful mindset setup.

Thank you. Yeah, thank you, Travis. Jack Hidary, my dear friend, one of our trustees, CEO of Sandbox AQ. You have your chairman sitting right next to you, no pressure, no pressure whatever. So first of all, please explain AQ, but then what are large quantitative models and what role do you envision they'll play in the current evolution of AI?

Well, Peter, first of all, I'm very happy to see that we still have humans on the panel instead of just AIs. So I think it's just another few, I mixed the AI in this, yes, okay, another few years it's just going to be avatars up here. The mantra of AI die is real, is real. It's not just a phrase. It's happening right now. It's happening on several levels, and certainly companies and countries that do not engage will die. But now the question is, okay, let's all say we all agree on that, which AI, what AI, what kind of AI, what mix of AI? And the marketplace now is starting to offer more than just one slice of AI, and that's what's exciting. So we have, of course, large language models. Many people in this room are and on this panel are involved in that and making huge strides. And we're seeing AI now take on reasoning, and I'm sure Eric and others will also address that. But there's a whole another side of AI that has had less focus, and that's quantitative AI. That's AI that's based on equations and data, quantitative data, numerical data. When we think about starting a biopharma industry in a country that never had a biopharma industry, that was really not even possible five or ten years ago. But now we can come to a country like KSA or other countries and say, hey, let's get that going. But it's not large language models that will win the day there. It's models that have been trained on biology, physics, chemistry, electrons, that kind of interaction. So quantitative, large quantitative models, LQMs, as complementary to LLMs, allow us to do that at scale and speed. And we do that on the same GPUs. We have NVIDIA here on the panel that LLMs use, but we tweak them in a very, very different way. We train it on a very, very different kind of data. The data doesn't come from the internet. The data does not come from downloading Wikipedia and Reddit and social media. Instead, we generate the data from the actual equations that govern our world. And this is now a whole new superpower for humans that we just never had before as a species until three years ago. It was not possible to calculate with any kind of accuracy how one molecule that's meant to be, for example, good for Alzheimer's or Parkinson's or brain cancer or pancreatic cancer would fit, would lock in to that receptor. Now, as of just 36 months ago, with the collaboration of many people in this room actually and on this panel, it is now possible. Yeah. So LQMs are really about another key tool in the tool chest, complementing LLMs, sitting alongside the many LLMs out there, but a core tool that LLMs will interact with going forward in the future.

So you mentioned data, and one of the things that we talk about is the importance of data and how the world may be running out of data to train these. So how should leaders think about the proprietary data generation they have and retention as AI models become more commoditized?

Well, Peter, we all talked about Gen AI. Gen AI is a very real and big thing. But now let's talk about Gen Data, generative data. How do you generate data? Well, when you want big data sets that govern our world, of either quantitative finance to look at portfolio optimization, or again, a new material to build a car that would be lightweight, make the car lighter weight so it's more fuel efficient, we can't turn to the internet for that data. That is not where we're going to get it. Instead, we generate the data from these equations. Those become interesting. Heisenberg, Schrodinger, yeah, exactly. That's where quantum comes in. When people hear the word quantum, I know people get scared, sorry, but quantum is not just about quantum computers. It's about today on GPUs from Alphabet, TPU from Alphabet, GPU from NVIDIA and others. There's many others coming out. We're able to run the quantum equations, that is the equations that actually govern our world, at scale on these GPUs and TPUs. This is new. These chips, by the way, were never designed to do this in the first place. And it's the human ingenuity that is quite fascinating that came up with this. You know, you did an incredible job spinning Sandbox AQ, A is for AI, Q is for Quantum, just for you to remember that, out of X at Google. And you have real products today. Can you just very quickly give us a sense of what Sandbox's products are that you're...

Sure. I'll give some real case studies here at FII. We have, I think, Paul Hudson, the CEO of Sanofi, one of the largest pharma companies in the world, and our two companies actually are announcing just today. We were just on CNBC earlier announcing that we're working together to apply this AQ, to apply this LQM, this quantitative AI, to accelerating biomarker development. That is development of diagnostics that will help all of us, and drugs as well. So that's a real-world case study of how we're using it there. We're also working with companies like Dow and others on the chemical side to say, how do we get new catalysts that could bring us new materials. When we think about the hydrocarbon space, think about the output of a refinery. We say, okay, the high-octane fuels, very profitable, jet fuels, kerosene, profitable, naphtha, but how about the bottom of the stack, Peter? That bottom of the stack, petroleum companies don't make much money on. It's sold to smelters and others and burned in the air. And so when we think about what we can do with AI that understands chemistry, we can now take the bottom of the refinery stack up-value it to things like carbon composites. Carbon composites are what make a McLaren car so powerful and yet so light, Aston Martin, Ferrari, and others. But it's not accessible technology to all cars. By working with the hydrocarbon producers, we can say, how do we now make this kind of technology available on a democratized basis? So these are real-world impacts, be it in biopharma, in chemicals, in fuel, in energy. And Elon was talking about energy and energy storage. The battery technology we have today, Peter, is 45 years old, and we need to leapfrog that. And this kind of AI complementing LLMs, this is where it's going to come from.

Let's go to your chairman next. Can we thank Ruth for spinning Sandbox out? Thank you, Ruth. Good job, Ruth. And what a financial success so far.

Absolutely, it's been amazing. I've never seen a company scale revenue as quickly as you have. It's extraordinary.

Dr. Eric Schmidt, former CEO and chairman of Google and Alphabet. There's nobody on the planet that I think holds your stature in this field. Thank you for all that you've done. The path to AGI. We talk about AGI, it's this blurry line of what is AGI. So do you have a definition for it and why is it so exciting? Why is, you know, Sam saying I'll spend $50 billion, whatever it will take. Talk about that, please.

It's worth understanding what will happen in the next five years, please. And the work that these guys highlighted and others on the panel will highlight are going to generate savants, that is specialized assistants that will work with you in whatever you do. An artist savant, a music savant, a physics savant, and so forth. Those savants will work with you to do research, drug discovery, solve problems, can do many, many things. Why five years? Because today we have all the components necessary. We have planning, we have the ability to do forward and backward reasoning, we can do stepwise reasoning, we can go against objective functions that are much more complicated than they used to be before, and we can generate arbitrary code. In the industry, it is believed that somewhere around five years, no one knows exactly, the systems will begin to be able to write their own code. That is, they literally will take their code and make it better. And of course, that's recursive. I thought that is essentially a change in slope. If you're going like this, all of a sudden it goes like that. It's reasonable to expect that within six to eight years from now, so 2030, right after that, maybe 2032, under current growth rate, it will be possible to have a single system that is 80 or 90% of the ability of the expert in every field. So 90% of a physicist, 90% of the best physicist, 90% of the best chemist, 90% of the best artist. When you have such a thing, you have a non-human that is effectively smarter than any human because no human can dominate all of those fields. Maybe Leonardo da Vinci could, but certainly not now. We don't know what happens when such a thing exists, but we know that that race is really important. There are many, many things that this thing, we don't know what to call it except an AGI, could do. For example, it could analyze cyber threats and develop new ones, or it could protect against them. It could come up with new biological solutions, good ones or bad ones. So there's both a national security component and a worry, but also a notion of a huge step change in human efficiency and productivity. I will assert that we as humans are not ready for the arrival of this. We're just not ready for it.

Yeah, I can imagine all future Nobel prizes in math, physics, chemistry, medicine coming from AI systems. And I should say, by the way, that if you go to Formula 1, you enjoy watching the humans drive around the track. Now, it's obvious that automated cars, way more cars, and so forth, could drive faster, but we wouldn't find that an interesting sport. So here we are in the land of golf. Maybe there will be a robotic golfer that will beat all the LIV top golfers, but we won't look at the robot, we'll look at the humans. So we poor humans will take pity on ourselves. Very biased, aren't we, right? That will be how we choose to entertain ourselves. Very briefly, but this takes an hour worth of conversation, what do you consider the most dangerous outcome around AGI or artificial super intelligence?

There's a huge issue around proliferation, and right now we don't fully understand the rate of proliferation of the mid-tier models and open-source models. There's a consensus at the moment that models that cost less than $100 million to train are probably not that dangerous, and ones that cost more than $100 million are more dangerous. I have no idea why we believe that, but that's the number. So what will happen is at some point there will be proliferation of inexpensive tools that can do significant damage. The most obvious one is in biology.

Yeah, thank you. Something you know a lot about. And I do think that we're going to see extension of the human health span because of AI. We live longer, but it'll be more dangerous in some situations. Hopefully we can moderate that. Ruth, a pleasure. Ruth is President and CIO of Alphabet and Google. You met her this morning. Ruth, there's, I want to quote this and then ask you about it. We've all witnessed companies making grand AI for good pronouncements, for sure. But when push comes to shove and financial performance dictates decision-making, often less lofty ideals are what shareholders demand. So how do you think, you know, I think the world of Google. I think Google has transformed the world in extraordinary ways. How do you think of AI for good versus AI for financial gain in the position you're in and Google as a company?

I think of it as a false choice because the upside from AI is absolutely extraordinary. But if we don't invest to protect on the downside, we'll never have the opportunity to actually invest to capture the upside. And that upside-downside point is two sides of the same coin. So when you think about the upside, it's around accelerating science, it's around social issues, solving things in education, healthcare, as we've been talking about. It's the economic upside. But if there's not a responsible foundation, if we're not doing the heavy work, we don't have a right to have a seat at that table. And that means engaging with regulators constructively. It means investing in our systems internally so that you're protecting from the downside. If you don't protect on the downside, you're going to find resistance, whether it's from the regulatory world or from everybody else. And so they go hand in hand. And to assume that you can pursue one without the guardrails that are critical, I don't think it's long-term sustainable. And so we view them as two interlinked approaches to really maximizing the upside of this extraordinary technology.

Yeah, you've managed through a multitude of ups and downs in the economy, including the 2008 financial crisis. What's important for business leaders here to know about leading their teams through the incredible changes we're about to see? And do you think people realize how much change we're about to see in business and industries and society over the next five years?

I think the most important point to take away from all these conversations is that the art of the possible has fundamentally changed. This is a generational opportunity like we haven't seen before. And you know, I spoke about some of the key proof points in the earlier discussion this morning. Probably one really important one is in life science. I'm so proud of our colleagues who won the Nobel Prize for this with AlphaFold, credited with being the most important contribution to drug discovery. Relevant to your question, Demis Hassabis, the founder of DeepMind, the one who went on this journey to create AlphaFold, when he embarked on it, he said many people said, how is this possible? And his answer was, why not? And so we kept going. And I think the other really important point when you think about this, that I talked about some of the language translation work we're doing. We now translate in 26 languages. The important point is that we added 110 languages in the last six months, allowing 500 million people on the planet. So it's a vertical lift. What that says is you better not delay. You better move right away. It's already been said on this panel, I 100% agree. And the economic upside is profound. So I would say the two most important, you know, there are many important messages, but one of them is the time is now. You have to reimagine what is possible because it really is. And we're seeing it already today. We're seeing it in the manufacturing sector, we're seeing it in education, we're seeing it in so many different areas. In fact, very importantly on this point about language translation, I was recently with the minister of digital transformation of a country in West Africa who said more than 50% of their population is under the age of 19. They knew that the most important thing they could do, and we've already heard comments about education today, the most important thing to do is education. How can you have a country with more than 50% under the age of 19 and not solve that? But they didn't have the teachers to do so. And what's exciting with AI is they can now actually provide education, quality education, to their entire population. And very importantly, in many countries, as in this one, in many of our countries, multiple languages are spoken. And before AI translate, you had to learn French or English in order to get the math books. Now you can learn it in your language. And so to me, when you think about the impact on humanity, it is incredibly inspiring. My dad always said when I was growing up, education is your passport for freedom. It is your passport for life. This is transformative. So one, get on the program right away. And the second really important point is what Demis said, which is why not? You have to radically reimagine what is possible. How you interact with customers, every element of process, risk management, the way to solve healthcare issues, education issues, climate change issues. But the time is now. It's incredible and it's near infinite opportunity, isn't it? Every industry will be transformed. Reimagine the impossible. Reimagine the possible.

Yeah, thank you. Thank you. Ben Horowitz, co-founder and general partner of a16z. Man, oh man, you guys have been on fire in the AI universe. So I'd love to hear your thinking as a leading thought investor. Where is the value going to accrue in this incredible value chain from chips and power and real estate and large models and applications? Are they all equally important to invest in?

Yeah, so I think as people have been saying, it's such a gigantic market. It's the biggest market we've ever seen. That there's going to be kind of money to be made everywhere. I think the issues and challenges are different at the different layers. So if you look at the kind of hardware infrastructure side, chips and data centers and power, the big, there's no question in 10 years we're going to need more of all that. Like there's no doubt at all. But the thing about that is, okay, how does that progress and how do you finance it? So if you look back to the internet, which you recall, we had the biggest bandwidth shortage in the world in 1999 and the biggest bandwidth glut. Akamai was charging whatever they wanted to back then. Yeah, it was the most bananas thing. And how could you possibly have too much bandwidth in 2001? It didn't make sense, right? If you look back historically. But what happened is the bottleneck moved. So the bottleneck was bandwidth in 1999, but then it became, you know, how fast could the server spit out the bits? Did we have load balancers that were good enough? You know, these kinds of things. And then we didn't need the bandwidth because we were stuck in other places. And already we've seen kind of the price of NVIDIA chips this year drop in half. You go, whoa, how is that happening? Well, we have this data bottleneck that's kind of becoming a sort of serious thing. And then like we're going to have a power bottleneck clearly, and other kinds of things. And then we'll have a cooling bottleneck. And so if you're financed with NVIDIA profits, then you're probably good. But if you're financing a data center with a lot of debt, you could lose it. You could get upside down very fast. So you have to really be thoughtful about that. Then if you go to the next layer, the kind of foundation models, the what we call the state-of-the-art models, Anthropic, OpenAI, Gemini, Llama, that's a really, so that's 80% of the software market. Everybody has to use it for infrastructure. So every application, everything calls on it. I'll bet Travis has it as part of his infrastructure. There's no question. So that's just a big and very fast-growing market. But it's really interesting in that the price of a token has fallen a hundredfold in the last two years. The most powerful technology in the world is almost free. Yes, it's almost free. Like the prices are dropping like crazy, but revenue is still increasing. So the size of the market is huge and the price competition is really intense. Like much more intense than you would see in this class of technology at this stage normally. And then we're kind of, we're asymptoting in certain places. So if you look at the growth of GPT-2 to GPT-3.5 versus the growth, you know, in terms of intelligence of 3.5 to 4, 2 to 3.5 is much larger even though we spent way more money going 3.5 to 4. That's because we're hitting a bottleneck. You know, like it's not the end of it. It's not going to asymptote forever, but it does imply that there's going to be an architectural change. And so even though the revenue is so huge in this market, there could be new players that emerge. We could be search before Google when we had 37 search engines and none of them were Google, and we can't even remember who any of them were now because they don't matter. So that's the kind of trickiness there. And then in the application layer, you know, Travis hit on a really interesting thing, which is big companies can optimize very fast with AI. And so as a new company, you have to ask the question, a new application, can I, you know, can I get to market faster than the big company can get a good product? And that's the race. And I'd say, look, in some cases it's the biggest private equity opportunity of all time. In other cases, we made it, just to give you an idea of how fast you can build a good product and take the market, we made an investment in a dev tool company, which is like the slowest growing stuff ever because you're selling to engineers and engineers hate buying stuff because they can build it, of course. And this thing has grown from zero to 40 million in revenue in three months now. Like, you know, like three years would have broken the world record for a freaking dev tool. But that's so they're going to take the whole market before anybody can build anything as good as they have. And so that's certainly possible. And then the other thing that's been...

I'm going to ask you to wrap up on that.

Oh yes, yeah. Sorry, I had too much to say then.

And I want to hear it, believe me. Yeah, yeah. Sorry. But thank you, Shou, as CEO of TikTok, it's a pleasure to have you here. How is TikTok utilizing AI to contribute to the creative economy? And what's the positive impact for the global economy as part of that?

And not at this moment in time, you know, we talked about AI safety. A lot of kids are using TikTok, my family's using TikTok. How do you think about making sure that you've got measures in place for the responsible use of AI?

It is currently true, and I tend to agree with, I agree with Eric on many, many things because he's a lot smarter than I am. But we are seeing some, for example, OpenAI has been posting Sora videos on our platform. I've been following them. Some of them are becoming really interesting. And you can see a path where at some point one of them could go viral and one of them could be so interesting that people feel that it engages them. It is possible.

Thank you. Jay Puri, EVP Worldwide of NVIDIA. You know, NVIDIA has become synonymous with the AI revolution that we're living in. Thank God for those early video game players, huh? How should companies and countries think about AI to drive growth and prosperity? We have a number of countries and leaders represented here, and what do you advise them to do right now? Because it's going to be fundamental to their existence, to their future economies. You agree?

Absolutely. Yeah, I mean, AI is basically about creating intelligence. So how can something about creating intelligence not be fundamental to the economy? And so, you know, and the other thing, of course, is the pace at which AI is advancing is frankly just unbelievable. There's just no expletives that you can use to describe that. Just two years ago we had, I call it all-new stock price highs. What's that? Exponential growth in NVIDIA's stock growth, right? So only two years ago we had ChatGPT and generative AI, and we were very impressed with the fact that, you know, you could get next token prediction and it would answer. And now today, as everybody is talking, you know, the AIs are becoming very sophisticated and able to reason and reflect. And in fact, they don't just do one-shot reasoning, they can reason a lot and solve very complex problems. This is the agentic AI. And then the way you interface with them is not necessarily a chatbot, but you can have human avatars. So, you know, we are going to have to learn how to live with the digital humans as teammates, if you will. And then going forward, we are moving into things like physical AI, where of course Elon talked about robots and humanoid robots, but all kinds of robots, even factories and warehouses and everything will be robots that will be built into the digital world, and they will all obey the laws of physics and so on. You will not be able to tell the difference between the digital and the physical. And you will use AI to optimize, you know, everything in the digital world, and then you will implement it in the physical world. And then it's kind of the AI flywheel, because whatever you do in the physical world then provides data to improve what happens in the digital world, and so on. So what I'm saying is, you know, with this type of what's going on, what is it to your point, you know, what is the condition that you have to create for you to be successful in using AI, whether you are a company or you're a country? And the condition is, of course, AI is about taking the information that is embedded in the data that is there, right? That's where the intelligence is embedded, and AI allows you to discover that using these sophisticated models and everything we've been talking about. So to be able to do that, what everybody needs to do right away is to build the instrument to allow you to do that, and that is what we call the AI factory. Okay, an AI factory is custom-built to take raw data, process it into generative models and so on, and then allows you to produce these monetizable tokens at scale, right? And so these factories, once you build them, can change everything. But they're not easy to build, as Elon was saying. It's extremely difficult to build these AI factories. And I think one of the things that we at NVIDIA have been doing, of course, we have been involved with every one of these factories that has been built. So we have tried to make it as simple and straightforward as one can, and make them by taking all the learnings that we've had, codify them into reference architectures and so forth, and see how quickly and replicate these factories, right?

And let me ask you a closing question here. The GPUs that NVIDIA has been producing, there are many companies chasing you, no one at the scale. How long do we see the GPU as the dominant architecture in AI? Do you imagine, is it the rest of this decade?

Well, I think one of the key things that people miss is NVIDIA is not just about the GPU. We are about an accelerated computing platform. So what does accelerated computing mean to you? Okay, so what accelerated computing is not about just, it is, you know, we need general-purpose computing, which is what we've relied on for a long time, sort of was using Moore's Law, and it kept advancing, say, twice as fast, you know, say every year or so. But that type of advancement in computing, even if that was to go on, and Moore's Law is not happening anymore, in 10 years you'd only be able to advance computing by 100 times. To do AI, we have advanced computing for a particular domain by using GPUs by a million times. That's the reason why deep learning is possible and the types of models that you're doing. And to do that, you have to innovate across the whole stack, not just at the GPU, but all the software, refactoring it, parallelizing it, looking at the unit of computing as not a server but a data center, and so forth. So, you know, that's what we think. It's not about a particular chip. It's about an accelerated computing platform that you can be true to, which is based on CUDA, and people can continue to innovate in AI, whether they're building models or they're doing inferencing or whatever, and they can be sure that there is backwards and forward compatibility. And so, you know, it's very much more than about a chip. And so I think that's what we have focused on, and that's what we hope to continue doing for the long term.

Ladies and gentlemen, six minutes per person is nowhere near enough. Thank you all for your patience and for your wisdom. Grateful. Let's give it up for our panel here.