Top Minds in AI Explain What’s Coming After GPT-4o | EP #130

Welcome to another conversation on AI. I don't think we get enough of this conversation going. I do want to thank Richard Attias and the FII team for really increasing the conversation this year on AI because I think there is no greater topic of import on the financial side, on the leadership side, education side, medical side. It's transforming everything. We have three incredible CEOs here who are representing a variety of different parts of the AI emergence. I'm going to start by asking each of them to take just one minute, introduce themselves and what they're doing and then we're going to jump into where is this going, how fast is it going? How big is it going to get? We'll ask the question what is after ChatGPT. Prem, let's begin with yourself.

Awesome. Thank you. Thank you Peter. I'm Prem Akaraju. I'm the CEO of Stability AI. We are one of the leading open source image, video and 3D models in, in the world and past GPT pictures are worth a thousand words and we're making quite a few of them. And in fact 80% of all the images that were generated by AI last year in 2023 were driven by our model Stable Diffusion.

Amazing. Richard.

Hi everyone. Really excited to be here. My name is Richard socer. I'm the CEO and founder of U.com, y O U Com. It's a productivity engine which is the next generation after a search and an answer engine. So we really make people more productive across a whole host of different kinds of organizations from hedge funds to universities to companies, insurance companies and so on. Publishers, news agencies and almost everyone else in between who has sales, service, marketing, research, analysis and so on. I also run a venture fund called AIX Ventures that invests in early stage pre seed seed AI companies and startups. I've been very fortunate that when I was a professor at Stanford, I had two students who created this cute company called Hugging Face invested in that at a 5 million valuation worth 4.5 billion now. So 4.

That's. That's bragging. That's just straight up bragging. I wish I could brag like that. Dr. Kai Fu Lee hi.

I've been working on AI for about 43 years. I was two at no at the time in college when I started AI. And I think that may have started before my colleagues were born but I actually worked on machine learning and I have a PhD, Carnegie Mellon and, and I have worked at Apple, Microsoft, Google. Some of you may know me as with my books AI Superpowers and AI 2041. My part time job is I run Sinovation Ventures which invests globally and then my full time job is I run 01AI. It's a generative AI company. We build a large language model. We're currently ranked as the third company with the highest performance, only next to the the best models from OpenAI and Google and you can find it online. We're also building consumer and enterprise products. We're based in China but our products are accessible globally. And also we extensively do open source as well.

So incredible. And first of all, Kai Fu is a legend and one of the greatest leaders globally in this field. So very honored to have him on here. Prem, I want to start with you. You very famously were able to recruit James Cameron onto your board. And since Stability is creating video and is creating sort of the future of Hollywood, I am curious about two things. One, did Jim get it right with the Terminator? And secondly, there's been a lot of conversation about the disruption of Hollywood that we're going to have AIs creating the future of all movies, all content and so forth. So you said beyond GPT models, images worth a thousand words. Talk to us about what this future is, what is going to happen in sort of the visualization world of TV in Hollywood.

Love it. So did Jim get it right with Terminator? Let's hope not, I guess. But what a great movie it was. And I love when he actually jokes about it. He says, I told you guys this is coming and now it absolutely is here.

And why would someone like him get involved in Stability?

Yeah, great question. So I had the great fortune of working on Avatar 2 with him when I was the CEO of Weather Digital before I joined as CEO of Stability. And that movie took over four years to make. And that's because it was fully rendered. And I think if you fast forward to five to ten years from now, the vast majority of film and television and visual media as we know it today is not going to be rendered. That's going to be generated. And in fact, in Avatar there were certain shots, there were certain that took 6,000, 7,000 hours of compute time to render one single frame. Thousands of hours that literally can be reduced down to minutes now. So I think Jim just wants a whole lot of life back. And when you think about like the creative process, we all watch films, we watch movies, we love them. From the time we're born to our last memory, it's a commodity we never get sick of. We never not want to watch it. And so there's this insatiable appetite out there in the world to consume stories and to create stories. And I think that we should just accelerate that. The problem with the film production process is time and money. So what he really wanted to do is rip those things out so we can move from a render to a generated model.

Are we going to see a situation where we're ever going to have. AI is generating entire movies because it knows my preferences, what I love, and it's like the perfect movie for me.

You know, personally, I kind of hope not. I don't think that actually the creative process, I think, needs to start with a human. And I think that human needs to dictate these tools in separate agents to actually make that story. And so I'm hoping that you'll probably want to hear stories that other people want to tell you.

All right, well, let's take a different direction then. Sure. Am I going to see Marilyn Monroe and all stars of the past coming back? Is there a need for human actors? If you can generate absolutely lifelike actors and actresses perfectly. I mean, I can't see a situation where they're still around.

Yeah, I think that it's actually quite. It's faster when you're talking about the film production process. It's actually easier to just shoot plates on an actor, just shoot real photography and get their performance. I think there's. That's the visible layer of production. People gravitate toward it a lot. I think that AI will enhance those performances. I think the physicality of a director with a camera and an actor in front of it is a very important part of the creative process. And I don't think that that's going to go away too soon. And in fact, I think about the things that aren't going to change just as much as I think is going to change. But I do think after they take one take, the director's going to say, I got it, because they're going to be able to do what you're talking about, which is manipulate that performance.

May I ask one more question to you before I move on? What is the most dramatic change we're going to see in film and TV 10 years from now as we see digital superintelligence? What's the craziest vision of what we're going to see in entertainment?

I think we're going to see on the magnitude of 5 to 10 to 20x more content being created. I think we're going to see a variation of time where it's going to be a two minute. Like you said, you may want to have 20 minutes before you go to bed, you want to see a movie. You'll have different type of time signatures. And I think that you're going to have an explosion of content creation and explosion of number of artists in the world.

I'm going to come back in 10 years and see if you're right about that.

Okay.

Richard, a lot of your work was instrumental in the early days of bringing neural nets to natural language processing. So what do you see as the next frontier beyond nlp? So explain, if you would, what NLP is and where is it going next.

Yeah, Natural language processing, NLP used to be a sub area of AI, and it has, I think, influenced pretty much every other area of AI. And there are lots of different algorithms you could train. And 2010 I had this crazy idea to train a single neural network for all of NLP. And 2018, we finally really built the first model that invented prompt engineering, where you can just ask one model all the different questions you have. And over time, of course, you can ask questions not just over text, but also over images. And so I think next, one of the answers to the panel's main topic of what's after ChatGPT is that we have many more multimodal models. You'll be able to have conversations over images. You have seamless inputs and outputs in not just the modality of text, but also programming, which is a huge unlock. Visual, videos, images, voice, sound. But one really interesting modality that not many people have quite realized yet is that of proteins. Protein. Proteins are essentially the basic LEGO blocks of all of biology. Everything in our body is governed by proteins. And you can create a protein, just like you can ask a large language model to write a sonnet for you or a poem for your wife. You can ask an LLM to create a specific kind of protein that will only bind to SARS cov2 or only bind to a specific type of cancer in your brain. And what that means is that we'll unlock a lot of different aspects in medicine. So I'm extremely excited about the future of LMS, going into different modalities, and.

We'Re seeing that with DeepMind's products in Alpha Proteo and such. So we had a conversation in back, but I didn't hear the answer. And the question is basically, is there an upper limit to intelligence? And, you know, we've talked about, and we just did a conclave on digital superintelligence and how fast we're going to get there and what does it mean as we think About AI becoming more and more intelligent. Yesterday I was speaking to Elon, he said, okay, 2029, 2030, equal to intelligence to the entire human race. Is it just, you know, a million times more and then a billion times more and then a trillion times more. Is there an upper limit to intelligence?

Yeah, so really interesting question. So just to talk about Alphafold and Google for a second, as you mentioned it, like that was really interesting to understand how proteins fold because that will help you understand how they are likely to function interact in your body. What we did in 2020 is actually create the first LM to generate a completely new kind of protein. And it was 40% different to any naturally occurring protein. And it actually we synthesized it in the wet lab. This was at Salesforce Research. What did it do scientists there? And, and it was an antibacterial lysozyme type of protein that basically has antibacterial properties. And just to put that in perspective.

2020 was really close to COVID 19. So make sure you weren't.

Gotta be careful what you say online sometimes. But what was interesting is that multiple startups have now started from this line of research and I think it's hard for people to fathom like how much that can change medicine in terms of upper bounds of intelligence. It's a really interesting question. Can it just keep going and going, going? I think you have to basically look at the different dimensions of intelligence, right? There's language, intelligence, visual perception, intelligence, reasoning, knowledge extraction and a few others. Physical manipulation and just. I'll show you just one example. So I don't want to talk, could talk about this for hours. But visual intelligence, right there are, you know, for a long time people have looked at just the electromagnetic frequency spectrum of human vision and there, you know, classifying every object on the planet is actually not that hard. And the upper limit is classifying all the objects on the planet. And we're probably going to reach that and we're not too far away from it. But that's just human vision. AI could eventually see all the way down to gamma frequencies and see and try to perceive atoms, right? And there you actually start to hit limits of physics, like quantum limits of like what can actually be observable. And you can go all the way into like seeing massively larger scale things at the universe level. And how many different sensors do you have in that? You can process all of that information and AI could have billions of sensors that go out and then you get into really interesting limits of like the speed of light cone of like Light cone observable. So I could talk about it for hours. It's a really tough subject, but. But in some cases we are astronomically far away from those upper bounds. And in some cases we already get pretty close.

Fascinating. You talk about work productivity as U.com's objective. What does that mean? And I guess the question is the same. Is there any limitation on work productivity that we're going to see? Given the fact that I can command AI agents and robots to just do anything and everything and just and self improve along the way, it seems like we're going to hit sort of an infinite GDP at some point.

Yeah, there are some areas of AI where AI can actually get into a self training loop. If there's a simulation of something and anything that can be simulated, AI can solve everything in that area. For instance chess, the game of Go, you can perfectly simulate it. Hence the AI can train and play with itself billions and billions of times, create almost infinite amounts of training data, and hence solve every problem in that domain. What are other domains that we can perfectly simulate is programming. If you can, programming languages can be run and then you can simulate the outputs obviously in a computer. And then the AI can get better and better and eventually get superhuman in terms of programming. But where I can't simulate things infinitely many times is in like customer service. Right. You can't have billions and billions of customers kind of ask about all the different things that can go wrong with the product that you're sending. And so in those kinds of areas the limits are going to be on data collection. Can you actually fully digitize a process? I often joke, like plumbers are probably the safest from AI disruption because no one's even collecting data on how to do plumbing. You like scroll somewhere, get different pipes. No one's having GoPro and 3D sensors and robotic arms and so on collecting data for that. So that will take much, much longer. I think in terms of work productivity, a lot of us are going to become managers. A lot of current employees that are individual contributors are going to have to learn to manage any AI to do the kinds of work that they do. And it turns out managing is also a skill. Not everyone is a good manager from day one. You have to really explain to the AI how you do a certain kind of job. And what we've seen, for instance a really large CyberSecurity company called Mimecast is we've had 200 seat licenses using their product and then we did a workshop with them and actually explained to all the different groups like this is what you can do. And someone from marketing can say, well, I usually get this long product description and then I have to describe it for these different industries and the email campaign and I have to write three tweets and three LinkedIn messages, all this stuff. And we're like, well, just say that to this agent. And then the AI agent does it for them. They're like, wow, now it's like six to 20 hours of work every other week. Just got automated by describing this workflow that I used to do manually to an AI agent. And I think that will change pretty much all work and pretty much every industry.

Kai Fu, I can go in a thousand different directions here. First of all, your venture fund synovations, which is how many billions of capital.

Aum we manage about $3 billion.

About $3 billion. And you've been one of the most prolific AI investors. I've had the pleasure to visit you multiple times in China and thank you for your amazing hospitality. You've now become an entrepreneur and you're running both a company in China and a company in the United States. Why did you do that?

Well, because this, this time is for real, right? Imagine, you know, this was my dream.

You've practiced before.

Well, this was my dream when I went to college. That AI was nothing. No one knew what it was, but I felt this was the thing I needed to do. And then we went through multiple winters of AI where there's disillusionment and I had to do other things. And about, you know, seven, eight years ago, we saw with, you know, deep learning, it was became clear it would create a lot of value. So. But at the time, I didn't really see it becoming AGI, so I was an investor. We actually created 12 AI unicorns in Sinovation Ventures. But this time with generative AI, the speed at which is growing is just phenomenal.

So you could help yourself.

Yeah, I felt if I just invested, I'd be missing out. I, I would be in the back seat. I want to be in the, in the driver's seat.

By the way, everybody, I hope you feel the same.

Right?

I. I'm very clear about saying there are two kinds of companies at the end of this decade, companies that are fully utilizing AI and, and everyone else is out of business. And I fundamentally believe that is true. You've written a number of books, AI Superpowers. I commend to all of you. So since that was published, what's the biggest changes in the global AI race? And it is an AI arms race going On.

Well, it is and isn't because the companies in China are largely competing against each other for the China market, and they're generally not.

I don't mean national to national, but it is between companies around the world.

Yeah, so you mean Chinese companies. What are their characteristics? So in my book AI Superpowers, I described the American companies are, generally speaking, more breakthrough, innovative. They come up with new things. And then the Chinese companies are better at engineering, execution, attention to detail, doing the grunt work, user interfaces. User interfaces, building apps. So in the case of mobile or deep learning, we saw that Americans invented pretty much everything, but China created a lot of value, arguably more given technologies that were largely invented in the us so now we're in this generative AI, again, invented by Americans, and we're in a unique position where. Where the technology is disrupting itself very quickly in the US and elsewhere. So it arguably is still the age of discovery and us ought to win. But then the Chinese companies are able to watch the innovations, make some themselves, and then do better engineering and deliver solutions. So the company I'm building, 01AI, is doing exactly that. We don't claim to have invented everything or even most things. We learn a lot from the giants in Silicon Valley, OpenAI and others, but we think we build more solidly, faster, execute better. So an example was I talked about how Zero1AI now has, is the third best model modeling company in the world, ranking number six in models measured by LMSIS and UC Berkeley. But the most amazing thing, I think the thing that shocks my friends in Silicon Valley is not just our performance, but that we train the model with only $3 million. And GPT4 was trained by 80 to 100 million, and GPT5 is rumored to be trained by about a billion dollars. So it is not the case. We believe in scaling law, but when you do excellent detailed engineering, it is not the case. You have to spend a billion dollars to train a great model.

This is really important for the audience here because there's a lot of parts of the world that don't have access to, you know, 100,000 H. You know, H. 100 clusters.

Right.

And the question is, oh my God, can I really build a business or a product in. Pick your favorite country with a small number of GPUs.

Yeah.

And I think the constraint on GPUs forced you to innovate.

Right.

Can you speak to that? I think it's really important. We talked about that on our last podcast together.

Yeah. I think, you know, as a company in China, first, we have limited access to GPUs due to the US regulations. And secondly, the Chinese companies are not valued what American companies are. I mean, we're, we're, we're valued at a fraction of the equivalent American company. So when we have less money and difficulty to get GPUs, I truly believe that necessity is the mother of innovation. So when we only have 2,000 GPUs, well, the team has to figure out how to use it. I as the CEO, have to figure out how to prioritize it. And then not only do we have to make training fast, we have to make inference fast. So our inference is designed by figuring out the bottlenecks in the entire process, by trying to turn a computational problem to a memory problem, by building a multi layer cache, by building a specific inference engine, and so on. But the bottom line is our inference cost is $0.10 per million tokens, and that's 1/30 of what the typical comparable model charges.

And where's it going?

Where's the 10 cents going?

Yeah, it's.

Well, the 10 cents would lead to building apps for much lower cost. So if you wanted to build a you.com or perplexity or some other app, you can either pay OpenAI $4.40 tokens, or if you have our model, it costs you just 10 cents, and if you buy our API, it just costs you 14 cents. We're very transparent with our pricing.

Yes, Richard, there's, there's a really interesting paradox called Jevons Paradox from the previous Industrial Revolution. A lot of smart people back then were working on making more efficient steam engines and using that use less coal. They thought, oh, if we make the steam engines more efficient, we're going to need less coal. But instead we needed more steam engines everywhere. And I think that's exactly what's going to happen. We're currently in the Jevons paradox of intelligence. We're just going to use intelligence in many more places. Everyone is going to have their own assistant, their own medical team that like, understands everything about them, instead of being restricted by intelligence being very, very expensive.

Yeah, I totally agree. I want to clarify, I'm not saying there's a fixed workload. We're making it cheaper.

Right, right.

I'm saying we're enabling a workload much, much larger.

I want to ask one closing question to all of you. We have people here who have daughters and sons or nephews or brothers and sisters. What's your advice to someone who is 20 years old listening to this or through this. What's your advice to someone at the beginning of their sort of academic and professional career, given what you know is going on in AI right now?

Prem, I think it's. Don't waste your time learning how to code because I think the new language.

That's fascinating.

I think the new code language is going to be English. And I think that absolutely. Learn as fast as you humanly possibly can on all AI and all AI modalities. And I think if you. And then once you find your passion, I think you're going to then find a very narrow AI to empower you to do what you're really set out to do.

Thank you, Prem.

Richard, I will disagree. I think you should still learn how to program. I think that is how you get to really understand how this technology works at the foundational level and how it becomes less magic and more something that you can yourself modify and construct with. But you need to combine computer science and programming with another passion that you can actually apply all of that intelligence to. And ideally, the younger you are, the more you learn the foundations, math, physics, the sciences.

I think I'm going to cut you off because I'm being. I want to have Kai Fu's final word here.

Okay. I actually agree and disagree with both of you. I think people should follow their hearts, right? If you dream of becoming a fantastic programmer and you can do it, you should do what Richard says. If you think programming is the way to make the most money. No. Then you should follow what premises.

Ladies and gentlemen, please give it up to these three amazing CEOs.

Thank you.

Thank you.

Thank you, Peter.

Thank you.