发明 Prompt Engineering 的人谈 AI、AGI 与人形机器人：Richard Socher & Salim Ismail

If you were given a couple of billion dollars, you'd be able to build a digital superintelligence. How quickly?

Probably 1½ to 2 years.

Richard Socher, often called the father of prompt engineering, is one of the top 5 most-cited researchers in AI. He's a former chief scientist at Salesforce and co-founder of the AI-powered search engine You.com.

Richard, what's the proper way to phrase your domination in terms of citations?

I have over 200,000 citations. I invented one of the most popular word vectors, got neural networks into the field of natural language processing, and invented prompt engineering.

That's right. Incredible. Richard is the founder and CEO of You.com. His company, MetaMind, was acquired by Salesforce, where he became chief scientist and executive vice president.

Salim, welcome as well.

Good to be here.

A lot is happening this week in AI, and I want to get Richard's extraordinary point of view. I want to start with the launch of Grok 3. If I had to tier all of the activity that's just occurred, I want to contextualize it with the fact that it wasn't very long ago that Elon raised $6 billion. Full disclosure: I was an early investor in xAI. He announced that he was going to create the largest GPT cluster on the planet, make it coherent, and he did that in 122 days, blowing people away.

Were you shocked by how fast he built what he did?

Elon executes. With $6 billion, you can do a lot of damage in AI. We've seen companies like DeepSeek and that hedge fund build amazing models with much less.

In some sense, it's amazing and surprising how quickly they got that far. But in some ways, you can expect this with exponential technologies like AI. If you have enough resources and you go hard, you can move pretty fast.

My standard phrase is, “Don't bet against Elon.” I just saw him last week in Miami. I was there for the FII Summit, and the guy does execute. He's got an incredible team.

I'm curious about how you're benchmarking Grok 3. Apparently, it's outscoring ChatGPT, Gemini, and DeepSeek. How do you rank it?

We already have Grok 2 within You.com, and it's a popular model, although there are others that are chosen even more often by our users.

I think what's interesting is that Sam Altman also talked about how the next generation of models will be almost at the level of a PhD student. But what we notice is that not many people are PhDs or have PhD-level questions in their lives. For more and more people, I think we've reached a level of informational and knowledge needs that's enough for them.

Now you push harder on really difficult tasks like programming. We've seen some exciting announcements today, including Anthropic's new Claude 3.7 model. Programming, science, and research are where the next frontier is for a lot of these amazing models.

Salim, what have you been hearing on the ground?

I'm hearing that Grok 3 is incredible, but the claim that it's outperforming all other AI models seems to be a little more hype than reality. I think it's coming in, as far as I can see from scanning Twitter, or X, a little bit lower than them—but still, it's unbelievable that he's been able to achieve this in such a short period of time.

Richard, I'm fascinated by the fact that you guys do federated AI, because you have access to many models. I'm really interested in hearing more about your model and what you're doing. But on the Grok 3 issue, for me the biggest thing is Elon's ability to achieve coherence across such a large cluster. That part blew my mind, because as far as I could see, every AI expert said you couldn't do it. I'd love to get your take on that.

Not many people have been able to set up a big cluster that quickly. In many ways, that's a combination of hardware and software. A lot of folks like me are more software people, and many AI folks have been spending most of their time in software.

It speaks to Elon's ability to work in both hardware and software, given where he comes from with Tesla and SpaceX, while now scaling everything up and getting all the software components working at the same time. Of course, there are companies like Anyscale and others that make it easier to deal with massive clusters. Anyscale allows you to scale from 5 GPUs to 5,000 GPUs within a few lines of code.

The layers of abstraction are getting higher and higher. Thanks to AI, we're all operating at higher levels of abstraction.

I'm curious about how people can evaluate these models against each other. At the end of the day, I think about human IQ tests as an interesting metric. I was fascinated when Claude 3 came out with an IQ of 101, and then GPT-1 or GPT-3 came in at an IQ of 120. I've been wondering when we'll see something come out at an IQ of 150.

Is that a relevant measure?

IQ has a lot of different dimensions, and intelligence overall has a lot of different dimensions. We briefly talked about that at our FII conference. I don't know if it makes sense to boil it down to one number.

Even the Turing test is essentially broken. The best way to fail the Turing test is to answer questions much better than a human could. If I say, “Write me an app in 30 seconds,” and it can do it, it's AI. If it can't do it, it's human. There are many ways we measure intelligence that are broken.

I'm working on helping the world structure that measurement a little better by understanding what the dimensions are and whether there are upper bounds to some of them, or whether they can just keep growing.

You're providing access to large corporations across most of the AI models. How many AI models do you have on You.com?

More than 40.

More than 40. Amazing. For people who want to understand the largest and most powerful models out there, what's your list of the top 5 or so?

You can't ignore OpenAI. A lot of folks still want to use OpenAI, and especially o1 and o3 are quite popular. We also have a lot of fraud—people trying to create accounts and turn us into a free API, making 10,000 calls in 1 hour. You think, “No one can read that. This is clearly a bot attack.” That happens all the time.

Claude is still very popular too. Claude Sonnet 3.5 is probably one of the best models for programming.

We have our own models, which we fine-tune from open-source models. Then we federate and ask different models depending on where people give the most positive feedback, given the intent they have. We classify the intent: Is it a programming intent? Is it history or medical? Then we route it to different models.

The most surprising thing is how often it changes, and how much mindshare DeepSeek gained in such a short time with almost no marketing budget. It was a very popular model for quite some time.

If you're enjoying this episode, please help me get the message of abundance out to the world. We're truly living during the most extraordinary time ever in human history, and I want to get this mindset out to everyone. Please subscribe and follow wherever you get your podcasts, and turn on notifications so we can let you know when the next episode is being dropped. All right, back to our episode.

Let's look at the Grok 3 benchmarks versus the competition. These benchmarks are on reasoning and test-time compute. Are they relevant and valuable? Everybody wants to know how fast these systems are progressing.

There are 2 interesting insights here. Most normal people don't have highly technical coding, science, and math questions every day in their lives. This is where we're pushing science forward, and that's where the frontier is really exciting.

The other interesting point is that we're looking at test-time compute. It doesn't even make sense anymore to think about a single model's intelligence. Some interesting research has shown that if you simply say, “Wait before you answer this,” and give the model more time to think, the same model does better and gives more accurate answers.

Speed is becoming another dimension of intelligence, obviously overlapping with many other kinds of intelligence. The faster you have to be, the less intelligent your answers are from these models. What that also means is that we may not have to worry about AI running away in open source, because you're going to need a lot of compute at test time if you want the smartest possible answers from these models.

There are a lot of interesting insights here.

I have a big one. As we move toward AGI, I struggle when people say “AGI,” because what does that even mean? I'd love your answer on how you define AGI. If we achieve it, how will we even know?

You also put out a tweet that I found interesting. You said something like, “If you were given a couple of billion dollars, you'd be able to build a digital superintelligence.” How quickly?

Was that a call for funding? Is it, “Everybody listen: Give me $2 billion and I'll give you your digital superintelligence”?

I miss going hard on the research side. When you build products and make revenue, it's amazing and meaningful, but I think there are still a couple of ways the research community is stuck where we can really push things forward.

With AGI, the definitions are so broad. Some people say it's when 80% of work can be automated. That's a pragmatic way of financially defining intelligence. I would say that maybe 80% of all digitized work can be automated, and then maybe 80% of all those workflows. That's already a huge amount of GDP, and it could be a reasonable financial definition of intelligence.

If you're more academically inclined, you have to acknowledge that there are certain kinds of intelligence and types of learning where you want to get faster. Humans can learn something with 1 or 2 examples. We call that learning efficiency. If you're really that intelligent, you should be able to learn with much less data along certain dimensions.

As we define it properly, we're going to have to examine different types of intelligence: visual intelligence, language, reasoning, mathematical reasoning, and social intelligence. Even among AIs, what actions could I take to modify your internal state in order to influence your actions? There are these different dimensions of intelligence.

Knowledge is another dimension, and it's quite unbounded. We can learn more and more about the universe, until we reach physics-based boundaries on how much knowledge we can accumulate based on the light cone around the different sensors we may have.

The full definition probably takes too much time here, but a financial, pragmatic definition—automating a lot of digitized work—seems reasonable.

What's your view of going into the physical realm? For example, Yann LeCun's test is, “Can you make me a cup of coffee?” Now you're getting into robotics. Another test I've heard is, “Can you take an IKEA box and put the piece of furniture together?” Now you're getting into physical manipulation, which is really one of the core rationales for intelligence.

Do you go into that world, or do you stay on the digital side because you can bound it more easily?

Physical manipulation is another dimension, or group of dimensions, of intelligence. At the same time, a deaf person can be very intelligent, and a blind person can be very intelligent. A person with paraplegia can be very intelligent, even though they can't manipulate matter.

We have to accept that these aren't necessary capabilities for a superintelligence. You can have a superintelligence that's purely digital, and it's just different from our intelligence. People who require a superintelligence to have fingers and move around simply haven't read enough science fiction, or aren't creative enough in their definitions of intelligence.

At the same time, I'm loving humanoid robots. The tricky thing is that we often use robots when we want to do certain things many times, very efficiently and quickly, like washing dishes or vacuuming the carpet. Then we give them specialized names: a dishwasher, a Roomba, or a vacuum. We don't call them humanoids.

Salim, you and I have had this debate a bunch, and I'm curious about your opinion—and Richard's—on the whole open-versus-closed AI debate. Do you feel that open source is gaining on closed source? Is that the definitive future?

Undeniably, open source is gaining. When you have this much excitement around something, and it's a product and experience that any normal person can appreciate, there's so much energy going into open source that it's very hard to compete with that in the long term.

The more niche and technical something is, the fewer people can appreciate using that technology. If you're doing ion thrusters for satellites, no one is going to build an open-source model for that with millions and millions of dollars of investment and excitement.

The fact that DeepSeek has been catching up is undeniable. I'm hoping we can eventually build one system where, almost like Wikipedia, people can contribute to it. No one does that. I'm going to have to do that at some point.

I have the same view. We saw this in the software world when Microsoft was running its IIS server and then open-source web servers came along. Open-source web servers absolutely took over. 99.9% of all web servers are now open source, and over time that will always win.

My question then is: We're heading toward open source, but we still have a number of closed-source companies. Are they eventually going to go open source? Is there a winner-take-all scenario here?

There's a good chance that if you're purely a foundational-model company, you'll look more and more like a telecommunications company: huge capital expenditures, very expensive to build, and creating a ton of infrastructure that creates value. But it's unclear whether you can capture all that value yourself.

Thank you for using that analogy. I think that's the perfect analogy here. We're commoditizing and demonetizing all of this. If you look at the demonetization curves in terms of cost per transaction, it's a rapid de-escalation.

In telecommunications, we had a massive amount of bandwidth built out in fiber, cable, and 3G, 4G, and 5G. The value wasn't captured there. It was captured by YouTube, Netflix, and the apps built on top of that. How do you think about that?

You can't build Uber without the Internet being everywhere, but Verizon doesn't get a cut of Uber.

That is why, at u.com, we haven't spent a ton of money training models from scratch. We've built a trust layer on top that professionalizes this technology so companies can really use it.

Thanks to DeepSeek, many of our existing and new customers are realizing that they should partner with someone like us. If a new model comes out in 2 months and you're stuck on a 1-year contract with one of the closed-source companies, you can't benefit from it.

That makes a lot of sense, because there's a continuous competition and everybody's racing to the bottom. If you become stuck with a particular model, you have no guarantee that you'll be using the most efficient, lowest-cost model.

We call it future-proofing organizations.

What does a trust layer mean for u.com?

A trust layer is highly connected to data and to helping people learn how to use the technology. We offer certifications so everyone can become a manager of their AI systems and agents.

We incorporate public data better than anyone else because we've been doing it longer than anyone else, but we're also incorporating companies' internal data. Then you can actually start to trust the system.

When you click on citations on u.com, especially in our more advanced research modes, you'll be sent directly to the quote. The browser scrolls down and highlights, “This is where I found this fact.” You can quickly build trust that way.

We taught our models to say, “I don't know.” A lot of models, if they don't find information somewhere on the web, will just make something up. Don't do that. There are a lot of moving pieces to making the system more accurate and building that trust.

Salim, you and I have talked about this when advising companies and investors about investing in AI. You want to invest in companies that have a great connection with their end customers and with data, and assume that the layer in between is constantly going to be replaced with the latest and lowest-cost model.

I think that will be key to success in AI platforms. Richard, it sounds like u.com has done an amazing job creating that layer of abstraction that protects people from the underlying model.

When Peter, you, and I talk to CEOs around the world, one of the huge questions is, “When do you place your chips?” The minute you put your chips down on a particular model, it's out of date in 3 months. You really need platforms like u.com to help with that.

Here's another article from The New York Times. For those listening to the podcast rather than watching it, the headline says, “OpenAI Uncovers Evidence for AI-Powered Chinese Surveillance Tools.”

We've had this incredible back-and-forth with TikTok, and now we potentially have it with DeepSeek as well. What's your view here, gentlemen?

I'm not surprised. My question would be, “How would it not be the case?”

If you download the model and use it in isolation, is it still reporting back information that it's gathered?

You can take the open-source model and still force it to take information from a prompt and from a search-engine backend. That's possible. You can also fine-tune the model to get rid of all the Chinese Communist Party alignment.

Our next story is “Accelerating Scientific Breakthroughs with an AI Co-Scientist.” I love the fact that the Nobel Prize went to Demis Hassabis and John Jumper for the creation of an AI model capable of predicting the folding of a protein.

My expectation, Richard—and you're both a deep scientist and a deep programmer—is that almost all breakthroughs in the not-too-distant future are going to come from AI. We'll attach them to a human so the human can get the Nobel Prize, but the breakthroughs will fundamentally be in materials, mathematics, science, and medicine.

Am I wrong?

100% correct. I'm writing a book on this in my nights and weekends called The Ure Machine, which is the working title. I'm a big believer in it.

What's interesting is that when you ask people around the world what they're most afraid of with AI, most are afraid that it will take their jobs. But in science and medicine, no one wants more jobs. They want more breakthroughs and cool discoveries.

Everyone worldwide is saying, “Let AI do a lot of science.” There's a lot of positive momentum behind it. We'll see more discoveries made with the help of AI, and eventually AI will do most of the work. You mostly need to guide it and tell it what you care about the most. Then it can go off and do more and more in an automated fashion.

This is the area I'm most interested in. If you provide AI with a data set and say, “Formulate 5,000 hypotheses and start testing them,” it can do virtual testing of all sorts of things. I'm incredibly excited about what will come from this.

I love the last bullet here: “Replicated 10 years of antibiotic-resistance studies in just 48 hours.”

Dario Amodei was at Davos and said something I loved: “We're going to see a century's worth of biomedical research in the next 5 to 10 years.” One could imagine that, during that century of biomedical research, we could potentially double the human lifespan.

It's not unlikely that we could double the human lifespan within the next decade.

We'll negotiate where we go from there.

A lot of people who say that people like Bryan Johnson and other longevity researchers are pursuing a bad idea are healthy and aren't currently battling anything. They're like people before the birth control pill saying, “That's not natural.”

A lot of bad things are natural, including murder and lawlessness. Humanity has been pretty good at improving on that natural state. It lacks a certain creativity when people think we can never solve aging and improve health spans.

In 2018, we started the largest project for a large language model for proteins. We published the paper while I was still at Salesforce. We had incredible success. We worked with wet labs and synthesized the proteins, and they were 40% different from naturally occurring proteins.

To put that into perspective, Frances Arnold won a Nobel Prize for what she called directed evolution. It involved random permutations with a lot of experimental science in the loop, identifying when a random permutation improved a particular property and then iterating from there. By the end of her very long process, those proteins were 3% different from naturally occurring proteins. Ours were 40% different.

What taught us that we'd captured the syntax and grammar of these proteins was that they folded properly, had the properties we predicted, and had the properties we wanted. Once you understand the language of proteins, all the medicine will follow.

Larry Ellison, when he was on stage at Stargate, announced the idea that we could have personalized mRNA vaccines against your cancer if you have it.

For me, this is one of the most extraordinary areas: reinventing medicine, curing cancer, curing viral infections, and perhaps curing death. Who knows?

This goes back to Salim's point about AI interfacing with the physical universe. Another friend, Alex Zhavoronkov, is the CEO of Insilico Medicine. He was very early in generative AI and drug discovery.

He's built a massive robotic laboratory where AI can come up with experiments and run them 100 times faster than humans. It gets the data, iterates on the experiment, and runs it again. You create a theoretical world and a physical world.

I think we're going to see hundreds of examples like this. The only limit is our imagination and how quickly we can apply these systems. The speed of the technology is now at a level where we can go down pretty much any avenue we want.

Personally, I'm interested in how you reconcile quantum mechanics with relativity. As a physics major, that's my thing. I think AI will be able to figure it out.

I can't believe we're alive right now. People should realize how extraordinarily lucky we are.

Every generation feels like it's alive during the most extraordinary time, whether it was the beginning of flight, electricity, the Internet, or something else. But I think we're too late to explore the oceans and the world, too early to explore maybe different galaxies, but we're right on time to explore superintelligence.

For sure.

The other area besides medicine is materials science. We just saw MatterGen from Microsoft.

Talk about prompt engineering moving into a completely different realm: “Design me a material that's superconducting, includes these elements, costs this much, and can be manufactured.” If we had a room-temperature, ambient-pressure superconductor from that, it would be world-changing.

The nice thing about chemistry is that, unlike biology, you can iterate even faster. There's no living tissue, and you don't have to run FDA trials, so you can iterate more quickly in that loop.

Materials science is at the foundation of everything else.

Salim, what do you think about this one? Satya Nadella said about quantum breakthroughs, “We believe this breakthrough will allow us to create a truly meaningful quantum computer—not in decades, but in years.”

Google and Microsoft are both making progress.

This is enormous. We have to keep in mind that quantum computers are only good for certain classes of problems, so there's that limitation. But the fact that you can create stable environments is huge.

I go back to Hartmut Neven's comment that the existence of a quantum computer may be proof of a multiverse. Your head kind of breaks at that point.

Richard, I'd love to get your take on this, because you go one step further. He says the only way a quantum computer can perform all of those calculations as rapidly as it does is by borrowing resources from a near-infinite number of adjacent universes.

We're doing the computation in parallel universes and bringing the answer back.

I love it. They'll be upset when they find out we're stealing their resources.

I'm super excited. Anything you can simulate, AI can solve pretty much every problem in that domain. It's just a matter of time and whether humans want to put in the effort.

You can simulate Go and chess. Chess is obviously solvable by AI because it can learn in 2 ways: imitation or exploration—that is, supervised training and fine-tuning, or reinforcement learning. When you allow a simulation to train and try billions and billions of things, it can get smarter over time.

What quantum computers will enable us to do, once we scale them up, is simulate much more of physical reality.

My favorite science influencer, Sabine Hossenfelder, put a bit of a damper on this announcement, saying, “We'll see if they can really scale it.” But I'm excited that there are different ways of approaching it, such as trapped ions and neutral atoms.

You hear a lot of quantum scientists dismissing the other approaches and thinking theirs is the best. Then a completely unexpected approach comes along, such as these topological qubits that almost no one had been working on. I love the energy around it, and the fact that some companies have such massive monopolies in their spaces that they have the resources to do 17 years of research before something finally emerges.

About 13 years ago, I had my 2 kids, my 2 boys, and I remember at that moment in time I made a decision to double down on my health. Without question, I wanted to see their kids and their grandkids. During this extraordinary time, where the space frontier, AI, and crypto are all exploding, it was the most exciting time ever to be alive, and I made a decision to double down on my health. I've done that in 3 key areas.

The first is going every year for a Fountain upload. Fountain is one of the most advanced diagnostics and therapeutics companies. I go there, upload myself, digitize myself—about 200 GB of data that the AI system is able to look at to catch disease at inception. It looks for cardiovascular disease, cancer, neurodegenerative disease, and metabolic disease. These things are all going on all the time, and you can prevent them if you find them at inception. Fountain is one of my keys. I make it available to the CEOs of all my companies and my family members, because health is new wealth.

Beyond that, we're a collection of 40 trillion human cells and about another 100 trillion bacterial cells, fungi, and viruses, and we don't understand how that impacts us. I use a company and product called Viome. Viome has a technology called metatranscriptomics. It was developed in New Mexico, the same place where the nuclear bomb was developed, as a biodefense weapon. Its technology helps you understand what's going on in your body, which bacteria are producing which proteins, and as a consequence of that, what foods are your superfoods, what foods are best for you to eat, and what foods you should avoid. It also looks at what's going on in your oral microbiome. I use its testing to understand my foods, medicines, and supplements. Viome helps me understand from a biological and data standpoint what's best for me.

Finally, feeling good, being intelligent, and moving well are critical, but looking good when you look at yourself in the mirror and saying, “I feel great about life,” is important too. A product I use every day, twice a day, is called OneSkin. It was developed by 4 incredible PhD women who found this 10-amino-acid peptide that's able to zap senescent cells in your skin and help you stay youthful in your look and appearance. These are 3 technologies I love and use all the time. I'll have my team link to them in the show notes below. Please check them out. Anyway, I hope you enjoyed that. Now back to the episode.

Microsoft dropped some AI data-center leases. The cancellation of U.S. data-center leases raised concerns about AI infrastructure overcapacity and shifting partnerships. The move sparked industry reactions, including among European energy stocks.

There's been a lot of buildout, and this ties directly to energy. I keep hearing that there's an open checkbook for building out capacity and energy. We're seeing small modular reactors, fourth-generation nuclear, being set up next to these facilities. And I don't get into politics here, but President Trump is saying, “Drill, baby, drill.” We need as much energy as we can get in the United States to support this industry.

Are we overbuilding, or are we not even close?

I believe we're overbuilding. I'll tell you why.

You look at DeepSeek and the massive breakthrough it achieved at a much smaller cost. The incremental effort to create the next generation is dropping by 10x every time we go through this. We should reach a point where training can be done very inexpensively and then spend much more time on inference.

The amount of buildout is exaggerated because it's based on the model size people thought they would need 6 months ago, when they started construction. That won't be the case when they finish building. The democratization aspect isn't being taken into account.

I mostly disagree. I've been talking for more than a year—and many other people have recently picked it up—about Jevons's paradox. When we make things more efficient, we actually use more of that resource. I think we're seeing that play out with intelligence.

Everyone will have a personal assistant, a personal health team, and a personal tutor. We'll use all of that intelligence, on top of everything else.

Many human problems are related to not having enough energy. When people say there's a shortage of water, there's obviously no shortage of water. It just has too much salt in it, which is an energy problem. If you have more energy, you can desalinate ocean water and solve that problem. There are deserts where people can't live because there isn't enough water. With enough energy, those problems go away too.

Where I agree with Salim is that when you build a lot of data centers, you also need to have data going into them. You don't want a real-estate crisis where you build a lot of buildings but people don't move into them.

My hunch is that data and energy needs will increase, intelligence will get cheaper and cheaper, and we'll still use more of it everywhere.

Let me distinguish between energy needs, which I think will be enormous, and data centers, which apply that energy in a particular way. I think we'll need less data-center capacity than people expect, but we'll definitely use all the energy we can for desalination and other purposes.

Before we get into Thinking Machines, this article from TechCrunch is about Mira Murati's startup. Over the last year, we've seen a constant flow of leadership leaving OpenAI. That's concerning. I'm not an investor in OpenAI, but if I were, I'd be very concerned.

What's going on there?

I think the doors are very open. The general point is that if you get to that level and you're suddenly the hottest property as an executive or deep researcher at OpenAI, you can follow your passion, find your massive transformative purpose, and build something.

Mira may be doing what she's doing, while other people may be interested in health care or specific applications. They now have the currency to go do that.

A second factor is the speed-and-“move fast and break things” approach Sam has toward building things, which concerns a lot of people. A third group is nervous that we're moving this quickly without adequate wisdom and thought about what we're building.

Richard, where would you place the emphasis among those different factors?

Zooming out a little bit, the fact that California has no noncompetes, while the rest of the United States is moving toward not having them, is tough for companies. Noncompetes aren't enforceable in California.

Research costs a lot of money, but once you show the world that something is possible, it's much cheaper to copy it. It's also easier to take the knowledge of how you've done it in one place and go do it more cheaply somewhere else. You don't have to take any code. The knowledge is in your head.

Overall, for the ecosystem, that's a positive thing. We're going to see cheaper, better, and faster models.

Let's talk about Thinking Machines. Any clue what Mira is going to focus on?

A lot of smart people joined her, including John Schulman, who led the ChatGPT application of large language models. Those models had already been available as APIs, and we had incorporated them into You.com in a search-engine context before ChatGPT came out.

Having amazing people who understand the technology and have ideas for building products is probably very positive. They describe a lot on their website. My hunch is that they're going to explore something. I hope they don't just build another large language model, because there's so much more out there. We'll see.

What fascinates me, Salim, and I'm curious about your view, is that her starting valuation is $30 billion.

Everything's gone up. It used to be millions of dollars, and now it's billions.

I don't know how quickly you can justify monetizing this stuff. I think we're headed for a pretty big bubble on the application side, because the user experience is being demonetized so quickly. Where will the revenue come from? That's the big question over time.

Putting my investor hat on for a moment, the way we think about this is that it's essentially seed-stage risk combined with late-stage returns. As an investor, that expected value doesn't quite work out.

That doesn't mean no one will succeed. It's just that in seed-stage investing, maybe 5% to 10% of companies do something amazing, and 1 or 2 of those companies, through the power law, can return the entire fund multiple times. There are a few such possibilities, but it's really tough. The bar is very high to generate enough revenue to justify these valuations.

Can I riff off that for a second, Richard? When you're investing in AI startups, you have to figure out whether the founder or team has something magical, and whether they can get to market and find product-market fit. That's a big challenge today.

How do you assess those points? Do they need to generate revenue, or do you invest in something with a massive breakthrough and hope the potential eventually yields results?

We've been doing well. Fund 1 is already around 5x TVPI, and it's only about 4 years old.

There are 2 ways to look at it. One is the horizontal, new infrastructure layer. In that category, you have companies like Hugging Face. I was fortunate that the founders were my students when I was a professor at Stanford. I invested at a $5 million valuation, and they're now worth $4.5 billion.

There are a few companies that can break out and become part of this new stack for building software that's fundamentally different with AI. Cursor is another one we're invested in. The founder was actually an intern of mine, and I was very bummed that I didn't get to invest in that company.

Then there are thousands of vertical application companies sitting on top of this new stack. There, we look for deep industry insights and deep AI expertise—teams that understand that their buyer will want a particular feature, rather than teams that simply try a bunch of different things and spend a lot of money.

Are proprietary data sets something you look for or find exciting?

The best companies will have what I call a virtuous data cycle. If they don't already have direct data access, they're building a product where using the product allows them to collect more data.

One reason Tesla is much better positioned, and why we've seen many self-driving-car startups die, is that the startups had to pay for every mile driven by a human to collect data. With Tesla, we all drive the car and give the data for free. In fact, we pay to drive the car and collect that data.

That's a perfect example of a virtuous data cycle. You see it in software-as-a-service products like You.com. People give us feedback—“This was a good answer,” “This wasn't a good answer,” or “I didn't like this part.” Those are ways to build an advantage over time.

So I get 2 things from this: Elon owes us money, and to be really successful in AI, you need to have been Richard's intern at some point.

That's right.

That was fun. The other side of AI is one of my favorite topics: humanoid robots. I was building robots when I was in junior high school, but they didn't do what robots can do today.

I'm going to share a short video. This is a robot called Clone. I contacted the CEO, and he'll be bringing his robots to the Abundance Summit next year.

What Clone is doing is essentially creating Westworld. These are muscles and hydraulic systems. The video underrepresents what it can do in terms of moving its hands. They hope to have it walking in the next few months. They're based in Eastern Europe, where they're doing a lot of work.

It's an interesting future for robots. A lot of robots in the United States and China are clunky walkers. They walk, but they don't have human-like emotional expression, cheaper prices, and incredible capabilities.

The second one is, I think, the black horse here, similar to DeepSeek: Unitree. Unitree has some insane videos that look like CGI, where you have four-legged robots that also have wheels, which I think is a clever idea. They're super fast, but they can also jump and climb up things and spin the wheels at the same time.

The question is always, “What's the most amazing use case for humanoid robots?” Why use humanoid robots instead of a tractor or a factory with a bunch of lasers, thousands of arms, and other specialized equipment?

You wouldn't want a bunch of humanoid robots walking over a field, just as we talked about the dishwasher earlier. At the same time, it's not a zero-sum game. There's a lot of cool stuff. I would totally buy a humanoid robot to have things done in my house and clean while I sleep. It doesn't have to be super fast.

The other thing is that everyone is working on the AI version of robotics, like the original Terminator. No one is working on a T-1000. One of my many ideas is to build a T-1000-like robot. I recently brainstormed about it with a brilliant hardware hacker, and he said it could actually work.

Jim Cameron was right, and it's all going to be due to Richard Socher.

I have to say something here. If you want a musculoskeletal humanoid robot, you get a man and a woman, have a baby, and grow the baby.

I struggle with this. If you want a dishwasher, you have a machine that sprays water in a particular way. It looks like a box, and it has trays to put dishes in. The same is true for a vacuum cleaner. Why are we constantly going back to the human form? The human body is frankly not a very efficient design for a robot.

Salim, we've had this debate. I say that if you're going to build a robot, give it 7 arms so it can do many more things. Why make it look like a human?

Go ahead, Richard.

I'm also an investor in Machina Labs. They build massive arms that can form sheet metal, and they work with SpaceX and many others. Whenever you don't want to build an entire factory to make the same large piece of metal millions of times, but you need to make it only 200 times, they're perfect.

They can ship a factory that creates any spare part into the field. It's almost like a blacksmith, but massive. They're also very anti-humanoid.

Again, it's not a zero-sum game. Some people want a beautiful humanoid robot in their house, but we can still have dishwashers and factory robots that are highly specialized and look crazy, with 20 arms. The excitement in robotics doesn't have to be zero-sum.

We have a lot of robot announcements this week. Next up is NEO Gamma.

I think this looks pretty cool in terms of its movements. We don't know how staged it is or how practiced it is. We don't see the 37,000 shots that went wrong, but it looks like a friendly home robot.

One of the questions I ask everybody is, “How many will you own?” When I interviewed Elon and Brett Adcock—Brett is the CEO of Figure, and Elon oversees Tesla—the projection was as many as 10 billion robots by 2040.

I can imagine that. I would have no problem imagining that I would own 2 or 3, maybe 10.

No, not you. You know my struggle with this. One robot moving very quickly is the same as 7 of them. Why does it have to look like a human being? It would be much better with wheels and 7 arms.

I think we're going to end up with the uncanny valley problem, just as we did with virtual reality. It's going to be very disconcerting. I think we'll have the same issue with humanoid robots.

For sure. Science fiction is underrated in showing us the positive possibilities. People will fall in love with their robots, and there are already androids now.

In the short term, we're going to see a lot of people remotely controlling robots and collecting training data that way. Part of the uncanny valley is that you may have someone in India or somewhere else sitting and looking into your entire home, navigating everything, seeing your kids, opening your doors, and so on. You have to be comfortable with that invasion of privacy.

Once the robots get good enough, they could be faster, have wheels, wear shoes with wheels, or have another arm. They could be more modular that way. I'm excited about all of it.

That was NEO Gamma from 1X Technologies. Let's go to the next robot, Figure AI.

For disclosure, I'm an investor in Figure. This is Brett Adcock's company. They recently announced their software. Figure used to have a software and AI relationship with OpenAI, but they ended that relationship and decided to build their own AI team internally with Helix.

The logic is similar to how Tesla got so much data from Autopilot as people drove it around, allowing the company to create incredible models. Figure AI will get a lot of data, train its AI in the physical universe, and move forward.

I hope they come up with a separate name for the robot, because calling the company Figure and the robot Figure is confusing. Let's take a look at their video.

Instead of having 4 arms, you have 2 robots collaborating. It's called collaboration.

I think this is going to take much longer to work out than people realize. But it's fantastic to see the speed at which things are moving forward. Ten years ago, when we first looked at robots, it was hard to imagine they would get this far. They were so clunky.

The use cases and application areas are where it will be decided. My Roomba still can't clean a room without me moving all the furniture around for it.

Robotics has done a phenomenal job. If we can constrain the environment a little more, it becomes much easier. That's why self-driving is also a fairly constrained environment. Highways look similar, and road signs are standardized.

Houses have very little standardization. You're right that it will be very difficult. The companies that can get 1 use case nailed down—one that's important and large enough—will have a huge advantage. But it is harder than most people think.

It will be very capital-intensive. Then the question is whether you can be a fast follower out of China and say, “This is how they do it. We'll reverse-engineer it,” leapfrogging the expensive research stage.

I'll go to my favorite use case, which is going to be a while away: getting one of these humanoid robots to say, “Go change the baby's diaper.” There are so many things that can go wrong with that.

I still love the idea of walking into a room and finding the robot holding the baby by one foot.

I can't do an episode without Bitcoin. Richard, are you a believer in Bitcoin? There's a faith component here. When I say believer, are you a holder?

I have a tiny bit here and there. I'm invested in a fund that does a lot of crypto-related things, just to have a little exposure. But I mostly want to focus on AI and find crypto a bit of a distraction. I'm not deeply involved in it.

When focusing on AI, agents will need mechanisms for transacting financially. Let's take it slightly sideways to cryptocurrencies and the ability of AI agents to do business with one another. What do you think about that?

It makes sense, but they can also use credit cards. We'll have AI making credit-card purchases fairly quickly.

I was dismayed when I tried to play around with the technology and encountered gas fees. The fees were so high that I thought, “This is almost a credit-card fee. It already costs a lot of money.” That doesn't seem right.

They need to lower the prices so the transactions themselves are almost free.

There's a whole stack. You have Bitcoin with expensive transaction fees and proof of work, then proof of stake. As you get closer to the end use, you need less security.

If you're storing jewelry in a bank vault, you need a lot of security, but you don't make many transactions. With a debit card, you can have much less security, and the transactions may be limited to $50 each. You can lower the security in exchange for volume. I think that's what we'll see in the crypto world as well.

How nervous do you get when you see the price of Bitcoin now?

I'm encouraged by what's happened over the last few days. Two things happened. One was the Bybit hack, which was the biggest hack ever. In previous years, that would have caused a massive collapse in the crypto world, but it was barely noticed.

The second was the response from the exchange. The CEO said they would make everybody whole again very quickly. That gives me encouragement that robustness is being built into the ecosystem and gives people confidence going forward.

The Trump meme coin didn't help the crypto world at all, which is unfortunate. That's life—you get what you ask for.

Did you buy it?

No. You can see it's going only in one direction.

You mentioned the Bybit billion-dollar hack. Can you unpack it for us?

One cold wallet storing a lot of Ethereum was hacked and suffered a massive withdrawal. If you're the hacker, you want to move the money into anonymous places and watch the transactions, because crypto is fairly traceable.

There were appeals to Ethereum co-founder Vitalik Buterin to roll back the transactions before the hack and essentially undo it. But washing all the currency out will be very difficult. Everyone is watching the wallets carefully to find out who did it.

I'm encouraged by the response from Bybit and its leadership, saying they're going to keep everybody whole. The fact that they had enough reserves to do this is encouraging.

In general, what we've learned in crypto is that you don't want to keep major wealth on a centralized exchange for exactly this reason. A lot of people lost money on Mt. Gox early on. You keep it offline and use exchanges for trading, but not as a store of value.

I use a Trezor or a Ledger—essentially a thumb-drive wallet—but I panic every time I plug it into my computer. It's nontrivial and very tricky.

This goes to the whole usability issue. When a technology goes from deceptive to disruptive, usability becomes 10 or 100 times better. Steve Jobs made the smartphone usable, and it took off. Coinbase made purchasing Bitcoin usable and user-friendly, and that took off.

But the rest of crypto is still a hot mess. Anyone who tries to buy or trade an NFT knows how sticky it is. To execute a smart contract, you have to be a level-14 geek just to touch the technology.

The tricky thing is that credit cards work because you're insured. If someone steals your credit card and you see a bunch of purchases, you can say, “That wasn't me,” and the bank gives you your money back.

With decentralization, you decentralize the risk, the security, and the liability. Each user is responsible for their own wallet. People simply aren't sophisticated enough to deal with all the cybersecurity threats.

Let's switch to MicroStrategy, now called Strategy. Michael Saylor was my roommate and fraternity brother at MIT, so we go way back. He's extraordinarily brilliant.

I was recently with him in El Salvador. I was there speaking with Carlos Slim, Michael Saylor, Marc Andreessen, and Ben Horowitz. Michael gave a massively compelling 90-minute presentation to a room full of billionaire family offices.

Every time I hear him, I think, “I should mortgage my house, sell everything, and buy Bitcoin.” He's dangerous to listen to.

One point for those who are nervous about this is that the strategy is to hold and buy on the dips. I need to verify this, Salim, but I wonder if trying to buy into and out of Bitcoin is problematic. I seem to remember that most of the gains last year were made on about 5 trading days.

That's historically accurate. In any given year, Bitcoin accelerates at some point, and a very small number of trading days account for 80% of the upside.

The problem is that you don't know which 5 days they are. I've managed to spectacularly miss 4 out of the 5, and then I bought on the other side of it and it went horribly wrong.

It's very tricky. What I tell people is: Buy as much as you can and close your eyes for 10 years.

If you can.

Strategy acquired another 20,000 Bitcoin for about $2 billion. Those are extraordinary moves.

I wish he had incentives to give 90-minute presentations to everyone to buy more Bitcoin.

He does. For sure, it's compelling.

If you wanted someone to be the prime evangelist for a technology, the articulation he brings to the table would be hard to beat. You could spend a lot of time trying to find a better one.

He's incredible.

Richard, open forum: What have been the most amazing events, breakthroughs, technologies, or companies you've seen in the last few months?

We just covered quite a few. I saw Agentforce. I did a podcast with our friend Marc Benioff. Agentforce 2 is coming on strong.

Peter, what do you think about the whole agentic world?

I'm a huge fan. When you think about the kinds of sequences large language models can handle, they are essentially very large neural sequence models. They can be trained on any kind of sequence, both through imitation and exploration.

In 2018 and 2019, you started working on large language models for protein sequences. That gave us biology. The very obvious other sequence is a sequence of actions.

I'm very excited. We already have more than 50,000 custom agents built on the you.com platform by our users. You can select which large language models you use.

Give us examples of the agents people would use. What are the top use cases?

Suppose you're in marketing. Every 2 or 3 weeks, you get a large PDF with a group of new features, along with a website describing the features that product engineering has shipped. You're tasked with writing 2 email marketing campaigns for specific industries and 3 LinkedIn messages. You also have to search the web and compare the new features with the competition, so you don't claim that your product has something the competition already has.

We've talked to marketers who say, “Describe that process to an agent on you.com.” The next week, when a new document arrives, you drag and drop the PDF into the agent, and it goes through all those steps. It writes the LinkedIn messages and email campaigns for you, and you're done.

Journalists use it to research new topics. Suppose they're writing an article about advances in prostate cancer. They need to go to 50 different sources, read a series of research papers, and put everything together. You can specify the kinds of sources you want, such as medical journals, and the agent handles much of the research and synthesis.

Venture-capital firms use agents in a similar way. When they receive a new data room, they go through 10 steps: net-dollar retention, CAC-to-LTV ratios, and so on. You describe those steps and drag the entire data room into you.com, and it goes through the process.

Whenever you're dealing with knowledge work, you can automate a tremendous amount.

Can you create an agent that says, “Go out there and raise me $1 billion in venture capital, find the companies that are going to be unicorns, invest in them, and send me the bank-account information at the end”?

That would be epic.

My definition of agentic AI is a white-collar job description.

The next level is when agents start taking actions for you, such as booking flights. But we may see a trough of disillusionment, similar to what happened with the Rabbit R1.

In its demo, it said, “I want to book a flight with my 4 kids to London on these dates,” and then everything was done. I thought, “There's no way that was real,” because there are so many details. Maybe I want the hotel close to certain sites.

Your preferences change over time. When I was a poor graduate student at Stanford earning less than minimum wage, I would have been willing to take a 10-hour layover to save $200. Now I spend thousands of dollars extra to have a direct flight.

The agent needs to know those subtleties: How long are you willing to wait, and how much extra will you pay? Personalization is still needed to make these agents work. But with knowledge work, you can already automate a lot.

Richard, why haven't we seen an agentic version of Jarvis that watches your tasks and says, “Last time you booked this, you always did that. Are you sure you don't want to do it again?” It could track you, learn from your patterns, and represent you more easily. I would have expected to see that by now.

Have you seen anything like it?

You would have to give it permission to listen to your phone calls, read your emails, and watch everything you do. There are 2 or 3 blockers, none of which are impossible to fix.

First, you're not allowed to record other people without their consent. Many countries will sue you for that. California and Europe have restrictions as well.

Second, Microsoft tried to launch something that watched everything you did on Windows, and people went crazy. They said, “No way are you going to send a screenshot of everything I do.” People do private things in their browsers and don't want to share all of that.

You need to build an enormous amount of trust. Many AI-first startups don't yet have the users' trust to collect all that data.

Apple may eventually be able to do it. Apple cares about privacy, and people may be more likely to trust Apple with everything they do on their phones. Eventually, we'll get there.

The fourth issue is that we're going to have more AI agents surfing the web than people. That's a massive change for how the Internet monetizes.

There are a few companies that make money selling physical goods, like Amazon. But even those companies are getting more involved in the main category, which is advertising. Your AI assistant doesn't get distracted when it's trying to book a flight for work to Utah by an advertisement telling you to take your next vacation.

Expedia and even Amazon make a lot of money from ads. If AI assistants start ignoring all of those ads, it changes how the Internet is monetized. Those companies will try to block operators and AI agents from getting work done.

You can have the intelligence, but the surrounding infrastructure will slow adoption.

Richard, who are your main customers at You.com? Who should check out your site, and how should they do it?

You can just go to You.com.

Our biggest customers include cybersecurity companies like Mimecast. We also have many publishers that use You.com to improve internal efficiency for journalists or let readers ask questions on their websites and receive citations from articles within their own networks. That can keep users on the site longer.

I want every journalistic outlet to eventually have its own GPT version that answers questions about its articles. You could imagine articles having personalized follow-up questions. If you've never understood why the Hutus and Tutsis were fighting, and you read an article about that conflict for the first time, the outlet could show you explanations, background stories, and additional context.

We're building that for media and publishing companies. We also have universities with 30,000 students going live on You.com. All the students can use it, and professors will realize that students can drag and drop an assignment into the system and receive a perfect answer. Universities will have to rethink their assignments.

We're excited about that, as well as a range of consumer companies that want search APIs to power the infrastructure of their large language models, answers provided for them, or both. We have API customers ramping up significantly, and revenue is increasing a lot. It's been great.

It's been a pleasure getting to know you and building our friendship. Salim, as always, thank you for making time.

I used to feel like I had a grip on what just happened. Now it's happening at an insane rate, and I can't imagine what next year will bring. Incredible week in technology.

Richard, Salim, thank you for joining me.

Thanks for having me.