AI 领袖揭示 AI 突破的下一波浪潮（FII Miami 2025 现场）| EP #150

Welcome, welcome. We're about to have a conversation, and I want you to listen up. This is the technology that's going to reshape your families, your lives, your businesses, your industries, and your nation-states. The question I'm going to be asking at the end here is: Are you ready? So let's dive in.

I'm going to be going through 2 rounds of questions. I'm embarrassed that we should, in fact, have a 3-hour session for our panel, but we've got minutes, so you'll excuse me as we run through this.

Prem, I love what your company has done, and you're an example of a CEO who takes a company and doesn't 10x it—you 100x Stability AI. What are you doing, and how are you going to impact the world?

Thank you for that. Stability AI is the creator of Stable Diffusion, which launched in August 2022 and changed everything in image-based AI generation. It was the ChatGPT moment for images. There have been over 270 million downloads of Stable Diffusion to date. To give you a sense of scale, the number-two most popular model has been downloaded 9 million times, so it is by far the market leader.

What we're using it for now—my background is in professional film and television—is what I call ultra-narrow AI: fine-tuning our model to work in professional content creation, including film, television, gaming, marketing, and advertising.

You brought James Cameron onto your board, and you have an incredible group of investors. I'm an investor, for full disclosure, and you brought in Eric Schmidt. How far are we from creating reality, given the technology that exists right now?

We're already there with certain workflows. To make this a full reality, we're doing exactly what an artist does when creating a film. What we've seen in other text-to-video models is 1 text prompt and 1 video. That's not how professional content is created. Professional content is created in shot elements, and then those elements are composited together to make the shot.

What we're doing is going step by step through each one of those processes—whether that be rig removal, paint and rotoscoping, or camera match-plate construction—and creating hyper-narrow AI models around each and every step. We're probably about 2 dozen models in, with about 50 to 60 models planned overall.

How far are we from me starring in my favorite episode of Star Trek?

You should be in that now. If it were up to me. But, in all seriousness, in terms of video generation on the fly, where I have a request to create something extraordinary that looks real, I would say that's going to happen this year. I think within 6 to 12 months.

Yeah, within a year. So how does your reality change when you're not sure whether you've created something or someone else has created it? What's possible for you in your businesses and your lives?

We'll come back to you in a moment. Ramin, I'm an investor in your company, so full disclosure over here. Liquid AI came out of the gate from zero to a $2 billion valuation in just about 2 years. At the end of the day, you're enabling private AI capability with your liquid models. I know a number of companies that are fearful—they don't allow their employees to use ChatGPT because they're concerned that OpenAI has access to all the data. What's possible using Liquid AI?

Absolutely. We are a foundation-model company. We are building generative AI systems for enterprises, and we're powering these systems with a new technology—not the Transformer architecture that enabled the new wave of AI, but something built on top of technology that we invented at MIT in Daniela Rus's lab: liquid neural networks.

These are brain-inspired AI systems that we evolved into something more tangible, and now we can create value from this new type of AI. The very special thing about this technology is that the amount of compute needed to pack a lot of intelligence into a device is very minimal.

As opposed to other types of AI, you can get a ChatGPT experience on a phone, locally on a phone or a laptop. In places where privacy matters, from a product perspective, what we do for enterprises is provide a solution wherever there is data sensitivity or a security issue that prevents you from using a cloud solution. Or, if you don't have access to GPU-based infrastructure, this is where Liquid AI can immediately come in and expand access at scale, enabling enterprises to use generative AI.

You just raised a killer round, with G42 as one of the leads—a quarter of a billion dollars. Congratulations on that.

Thank you very much.

You're tracking toward revenues this year that are spiking, which is fantastic. I remember you used your liquid neural networks to fly fighter jets. Can you take 1 second to talk about that?

This was one of the only neural-network architectures that enabled safe applications of AI on a device. The United States Air Force trusted our technology to be the first version of a neural network that could autonomously navigate a fighter jet.

Private AI isn't just about having a ChatGPT experience on a phone. It can power cars, go on a satellite, or go on a jet. The applications are phenomenal. Recently, I talked to some CEOs of an education company that provides tablets to students, and they want to have this kind of experience in that sector.

This is your educator, your physician, your pilot—everything. It's innate intelligence on your device that you own 100%.

Amazing. We'll come back to you.

Jack, a dear friend. First of all, I have to point out that Jack is dressed for Miami.

We're in Miami, Peter.

Welcome, everyone, to our home in Miami. Thank you for bringing us to Miami, and I have a request for next year. For FII Miami, could we request that the dress code be Miami business? Does everyone agree?

Yes.

And, by the way, Miami business means wearing avocado dress socks. We have avocado dress socks. There are many others to choose from, though.

I do want to do a commercial for Jack's book, AI or Die. It's very subtle, Peter—AI or Die—but it is something that every person should read. It is literally what you need to understand as a leader about AI, and it's written in a very, very readable fashion.

Jack, you are the CEO of SandboxAQ. A is for AI, Q is for quantum. You've got an incredible chairman of your board in Eric Schmidt. You spun out of Google with an incredible seed round. How much did you raise?

We raised $850 million.

Amazing. All right, so what do people need to know about SandboxAQ and the quantum networks that you're producing?

It's a very exciting moment. First of all, it's great to see so many friends on the panel and in the audience. This is an incipient moment for AI. Everyone's excited, lots of businesses are looking at it, but I think we're past, hopefully, the shiny-object phase of AI. Now it's getting serious. Everyone on this panel has very serious offerings that really impact business and affect how Hollywood works and how many major sectors of the world work.

At SandboxAQ, what we realized is that language models are fundamental. For everyone, they're table stakes. If you have customer service—whether you're Delta, Hertz, Hilton, or any company with thousands of customers—you must be using large language models to cut those costs and actually deliver better customer service. I think we all know, though, that customer service cannot get worse than it is now, so it's only going to get better.

We at SandboxAQ decided to go after a different part of the economy: quantitative AI, not language AI. What do we mean by that? If you're Sanofi or another drug company, and you want to create a new medicine for cancer, Alzheimer's, or dementia, each of our families here in this room will unfortunately be impacted at some point in our lives by these diseases.

Language models can help initially, when they scour and look at all the summaries of scientific literature. That's very helpful to give you ideas about what's been done before. But ultimately, if we're talking about building a molecule, we need an AI that isn't trained on social media and cat pictures, but is trained on molecules and atoms. That's fundamental, and that's the AI in which SandboxAQ is the global pace-setter.

Very importantly, you're not talking about using quantum computers to run these quantitative networks; you're using quantum equations on GPUs. What kind of quantum equations are you using on the GPUs?

We're all familiar with Schrödinger's equation and the other equations that we learned about in school. The breakthrough we had was realizing that GPUs were getting so much better. Fundamental to GPUs is the ability to run matrix algebra in parallel.

Imagine a spreadsheet like Excel multiplied by another big spreadsheet: 1 million rows by 1 million columns, and another 1 million rows by 1 million columns. That magnitude of matrix algebra can actually convert the quantum equations—the equations of drugs, treatments, new energy, battery storage, and all of that—to the language of the GPU. That's the breakthrough we had.

When quantum computers come and scale—we just had a great announcement from Microsoft yesterday, and another announcement from Google a few weeks ago—you're going to see these announcements come in great cadence, culminating in a crescendo, Peter, in about 5 to 7 years, when we have great quantum computers. We'll add those to the arsenal. We'll have GPUs and QPUs, or quantum processing units, in 1 mesh cloud hybrid.

But today, Peter, we use GPUs to get the work done with drug companies and Aramco. Aramco, I see, is a sponsor. Our newest announced customer is Aramco in Saudi Arabia. Why Aramco? Because they want to take the hydrocarbons coming out of the ground and convert them to higher-order chemicals using carbon and hydrogen—not low-grade fuels, but carbon composites. Those could be used to make a car lighter, a space rocket lighter, or an airplane lighter for Airbus or Boeing.

This is the kind of transformation that we focus on with LQMs, or large quantitative models, versus the very necessary large language models.

Amazing. Jack is a nuclear power plant, that man.

Jim Keller, Tenstorrent. You're a hardware manufacturer, our sole hardware manufacturer against all of these software geeks. Congratulations on your recent round—a pretty good $700 million Series D round. You've got a lot of capital and the ability to build hardware.

I'll take that.

What kind of hardware are you building? When someone says, “No, I only do software,” how important is hardware versus software today?

GPUs have a real, solid head start on building AI because they had parallel computing. But they're still relatively complicated to program, and the way they handle tensors and things like that wasn't native to GPUs. Now GPUs have evolved to add tensor processors. Tenstorrent builds a native tensor processor that's simpler and easier to program.

We also build it so that the tensor processors natively talk to each other very well. Last year, we open-sourced our software stack. The fundamental math of AI is simple: A equals B times C plus D. It couldn't be simpler at some level. But the scale of it is amazing. When I started building computers 40 years ago, we were doing millions of instructions per second. Now we're doing trillions and trillions of instructions per second, and scaling that takes a special collaboration between the hardware and the software.

When your machines are up and operating, what are they enabling for people in the room here?

Right now, there's a really large family of models. Our mission is to run all the models with very simple, transparent code. With large language models, people say the software is huge. Actually, it's 600 lines of code. It's not very complicated at the program level, but when you go down in the software stack, it can really explode.

By building a native software stack that is tensor-based, communication-based, and open source, people can see exactly how it works and how it runs. I think that's going to unlock a lot of AI applications that are currently hard to program with GPUs.

There's been a lot of debate about open-source versus closed-source AI models. We just saw the R1 model being open-sourced, and we're seeing a lot of conversation in which leaders are saying open source will win. I think there's been extraordinary velocity in open-sourcing. How important is open source, as far as you see it?

I have personal experience working with GPUs where we were trying to solve a hard software problem, but we couldn't because the math library was encrypted or part of the software stack was proprietary. Because we couldn't look all the way down the stack, we couldn't figure out or solve the problem.

Open-source AI is really wild because most of the high-end research is published, and many of the models are open source. Some of the weights are open source, but not much of the infrastructure or the foundation libraries. It turns out to be a mixed bag.

One thing we're going to do is open-source our whole software stack. It's for our hardware, but I encourage people: If you have your own hardware and you want a software stack that works, steal our software. It's a beautiful thing.

I want to make it so that many of the foundation models, the environment, the framework, and the tools to build and train your own models are also open source and available. I think it's really important to democratize the hardware stack and the software stack, so it isn't just a few very large players controlling the AI world.

We have a lot of people from around the world here. Are the machines you're building likely to be used in the Global South more than in North America?

No, we're going to sell to everybody. We license a small AI configuration to go on a television chip, and we're building machines that can train large language models and everything in between.

The other part of our business model—and I think innovation comes from lots and lots of input—is that the software is open. It's been our best hiring strategy, by the way. Our programmers look at our software stack, they like it, and they send a résumé. Or, worse—or funnier—they don't like it and they send me a résumé because they want to come fix it.

I think that's really great. We've licensed our AI and our RISC-V CPU technology to people, and they like it, use it, and send us feedback. We're going to license our AI technology, but also build and sell systems.

Fantastic. How many folks here have heard of McKinsey? Everybody, right? How many folks here have heard of QuantumBlack? Could you raise your hand? You need some publicity, Alexander.

Alexander runs QuantumBlack, which is a 5,000-person software and engineering team inside McKinsey focused on AI.

I think there are basically 2 worlds. There is a beautiful and shiny world on this stage, and we all enjoy the age of AI and the valuations. We have a quite confused audience that hears about it but doesn't see any impact in their lives, whether in their bottom lines or as human beings.

The question is: How do you reconcile these 2 worlds? If you look at the numbers—and take the technology companies aside—the sad number in the last 5 years is 11%. Only 1 out of 10 use cases ever saw the light of production. Everything else is entertainment. We play with it, but it's irrelevant. With generative AI, the success rate is maybe 7% at best.

What we're trying to do in QuantumBlack is bring these 2 worlds together, moving from 11% toward 100%. What we have today, exactly as you alluded to, is 5,000 people working in 50 countries, trying to transform nations and companies. We have 5 R&D centers working on the most precious products for humanity, and roughly 43 products that we deploy globally.

We work with many colleagues to bring their innovation into the day-to-day operations of enterprises around the globe. McKinsey is known for producing slides and many other funny things, but, to be fair, it has reinvented the consulting profession twice already. This is the third attempt, and I'm humbled to be here to try to reinvent it.

One of the things I say is that by the end of this decade, there are going to be 2 kinds of companies: those that are fully utilizing AI and those that are out of business. Do you agree with that?

I agree with it, but I think the problem is slightly different. Today, most of these transformations—and I think everybody here has tried to do a digital transformation, or at least declared it to their boards and shareholders—have a low success rate because we're trying to infuse technology into, by definition, broken processes and into the old process.

What I truly believe is that, instead of learning the technology, as all the colleagues here on the stage are doing, and stepping back to reinvent something and understand a completely different reality that operates on a very different cost structure and different social rules, we try to force it into the old system. Therefore, it fails.

If anything, I truly believe that we are at the end of the age of mediocrity. Whatever mediocre standard could be handled by a machine, a machine will handle. We're actually at the beginning of the age of creativity, because the notion of how you create something with technology that is a commodity becomes much more interesting.

Prem, back to you. What's the most important thing the audience here needs to take away from the work Stability AI is doing, from your perspective on AI as an enterprise creativity tool?

We've all seen the controversy around AI in the entertainment industry. The industry even went on strike for over a year, and then the guilds settled and cut deals with the studios.

This is really no different from what happened in 1927, when movies went from silent to talking. There was great controversy at that point. People on Broadway thought talking was for Broadway, and movies needed to be silent. Obviously, that was proven wrong.

Color took forever. It took decades to catch on, and then, finally, in the 1960s, it did. Now it's unthinkable, except as an artistic choice to be in black and white. Digital transformation was the same way: Everybody fought it at first. They're confusing headwind with tailwind, I think, is probably the best way I can summarize it.

When digital kicked off in 2000 instead of film, by 2017, I think 98% of all films were made digitally. The lasting statement for the film industry is: Don't look at AI as a headwind. Look at it as a tailwind.

Do we see Stability becoming a creative agent so that every individual can become a creator?

Absolutely. We're going to see it in Avatar 3, 4, and 5.

Avatar 3—Jim is editing now, so I think it's done. Hopefully, it'll come out in December.

Definitely, I think in the later Avatars and others, you're hopefully going to see a lot of our tools in there.

All right. Congratulations on the success. Prem came in as CEO about 8 months ago.

8 months ago.

He has revolutionized the company. There was a huge legacy of models and capabilities, but he has really driven it in an extraordinary way. Thank you, and congratulations.

Ramin, the world needs another LLM—why?

What the world needs is, as was detailed very nicely, to make AI useful. It doesn't matter what runs AI. We're in this amazing period of time when, at every scale, AI can bring value.

I can see a future where everything is going to be integrated into our society. It's not that we need a different type of LLM; we need to do it right. We have a motto in our offices: Every engineer at Liquid AI is designing AI. We call it “ML Done Right.”

That means we don't need to consume a lot of energy to build AI systems, and we don't need to use a lot of energy to host AI systems. What we're doing is democratizing access to AI, thinking about it in the cheapest possible way. Hosting a foundation model on a phone or on a device with Liquid AI costs $0 because it doesn't run on a GPU anymore.

Is everything in my home, in my car, and in my office going to be AI-enabled?

Correct.

What's your world going to look like when everything is intelligent—every device you're touching, talking to, or thinking about is intelligent?

I think it's going to be amazing. The humanoid robots that run in our homes in the future aren't going to have to be connected to the cloud. They're going to have their own local AI brains, where they can be safe, so Elon can't actually start the robot revolution.

Jack, you see the future and you're leading it. It's not 10x; it's 100x. As we start to see LQMs and quantum computers coming online, is the world ready for how much is going to change in the next 5 years?

I think it's going to be a fascinating next 5 years. But, Peter, if I can give 2 ideas to share with the audience to help absorb what's about to happen:

First is the power of small teams. What I recommend to every one of us, and what we're practicing, is that armed with LLMs, LQMs, and these new AI tools from the panelists or from others at this conference and elsewhere, small teams can change the world.

If you're a big company, portion off a team of 10 people and say, “You're going to be in this new area. You have this mission. Go.” Your moonshot teams.

At SandboxAQ, just to give you 1 practical example, there's a big issue when you try to fly a plane. Now there's no more GPS if you try to go to parts of Europe. There's no more GPS if you go anywhere in the Gulf region—Saudi Arabia, if you're landing in Riyadh or Dammam, or if you're landing in Abu Dhabi or Dubai. It's being jammed and spoofed. It's also out in the Indo-Pacific area, where the PRC, China, is blocking GPS.

An 11-person team, armed with this kind of AI and some quantum sensors, solved the problem. It's now flying on the United States Air Force. Small teams are the order of the day. As managers and leaders, this is what we're doing, and this is what I think more people will start to realize.

The second thing I would leave you with, Peter, is that I know a lot of people are still concerned about AI and its implications. Let me also say that what we're concerned about is people not embracing AI fast enough to solve the big problems of our current society.

Let's tackle the big diseases that have plagued and challenged us for 40-plus years. Let's bring battery storage to a new level, going beyond lithium-ion and beyond the current chemistries. This is where we need to focus more: embracing AI responsibly, of course, but making sure we lean in.

I was pleased to see at the Paris AI Summit we all just came from that there was a lean-in attitude, rather than 2 years ago, when people were saying, “Should we even touch this stuff?” Small teams, and let's lean in and solve the big problems in society.

Amazing. Thank you, Jack.

Jim, what do you want people to take away from the work you're doing? What should they remember, and how should they utilize the technology you're building?

AI doesn't have to be unbelievably expensive, unbelievably big, or unbelievably proprietary. That's not required. The computational hardware is fairly straightforward, and we want to make it available to lots of people so they can use it.

I think there is going to be a big up-level in how we build and write software and build machines. It shouldn't take 2 years to build a computer. We want to pull that down. It shouldn't take $10,000 to buy a single chip. We're going to take that down drastically.

How much cheaper are the systems you're building?

Our target is 5 to 10 times cheaper than the current systems. Then we have a roadmap to continue to make them better.

The other piece is that you have to take a big-swing approach to using AI, but start small. I'm asking my software team to double their productivity this year. Everybody is starting to use the code generators and code helpers. We're building our own tools to check the quality.

Just start working on it and get used to it. You're right: If your system is broken, patching up the broken system isn't quite right. But getting a real feel for it, using it, and then starting to iterate on how your system works is really important.

Everybody should dive in and embrace it, but we don't have to solve world peace first. I would like to make my code have a few fewer bugs.

Alexander, last words from you. Who typically comes as a customer to QuantumBlack, and what is your value proposition? Is it, “We're going to understand your problems and solve them”?

Basically, first of all—and it goes back to the question of what needs to be done—the customer is the chairman or the CEO. Unless the number one person in the organization, or the head of state, is really interested in this problem and willing to invest his or her time, don't waste your time. It's not going to work. You're never going to do it right.

Unless you have buy-in from the very top of your organization and are prepared to make yourself AI-first, it won't work. One of the biggest challenges a lot of companies have is that you're not competing against your typical companies. You're competing against the startup that is AI-native from the beginning.

It starts with your own literacy. I think this room grew up under the paradigm that unless I can explain something in 2 minutes, I'm probably incompetent. It's all right, but we need to speak the same language. The first thing is to go and learn that language.

While we could claim that AI failed in many things, you could clearly see the drop in AI usage during the summer. You ask yourself why: because all the kids are out of school or university. AI was the best tutor in the world.

So, first of all, use AI to educate yourself. That's number 1. Number 2, this is a leadership challenge, because to go fully in and transform the enterprise, you need to get the data right. That is never right. You need to change your architecture, which essentially means changing the politics within the organization, and we don't like to change politics.

Then, hopefully, you hire good people, but you also have to teach the rest of the organization what this is and how to ensure that you embrace it. All of a sudden, you have a team with human beings and certain agents. How do you operate that team yourself? You multiply the likelihood of success.

Unless you believe and go fully in, it's risky, because you put your career and your company's future on the line. You cannot go small. You need to go big to succeed. That's what we're trying to do: use AI as a way to make the world a better place.