LinkedIn Co-Founder Opens Up on the Reality of AI Job Loss | EP #194

Salim, you just got back from India. I missed you on the last 2 versions of this podcast. Did you solve the trade issues? And didn't you bring me back my iPhone 17?

I brought back some parts. So, if you want to assemble it yourself, you can, because that's somewhat flaky over there sometimes.

The 2 things that blew my mind were the attitude of everybody in India—it was literally “middle finger to the USA”—and I think this is a big challenge, because if India and the US, and India and China, start trading, the rest of the world kind of goes to hell. It's kind of a big deal. But what I found most incredible was the unbelievable optimism for AI and the use cases that are exploding there out of the gate. I think that's amazingly exciting, and we'll get to that in a little bit.

Dave, I just saw you up at Stanford. We interviewed the CEO of Replit along with Salim; that was a fun conversation.

Yeah, it was fantastic. I'm still here, actually. 150 million repositories—it's incredible.

The code's piling up like crazy. That'll come out soon, but Peter used it on his plane, too, so it shows you.

Yeah, it was great. It was so fun. I literally downloaded Replit. I had Starlink on the dash of my SR22, and I was flying connected, and I vibe-coded a mindset app on the way up there. It was great. I tweeted it out.

All right, let's drop in on the subject of jobs and education. Reid, I want to go to you first. Our common friend Erik Brynjolfsson published a paper recently with these charts looking at entry-level job loss—down 16% in AI-exposed fields—and you made some pretty sharp comments on this conversation. How do you think about this?

Well, a couple of things. It is definitely the case that AI will lead to a lot of different job transformations, in some cases flat-out job loss. A simple heuristic that I sometimes use—it's a partial heuristic—is that if a human being is trying to do a job by following a script that an AI can follow better, such as customer service, that will happen. But obviously, some of the issues here are around not just customer service, but also software engineering.

That being said, even with a downbeat in possible initial software engineering job hires, my belief is that that is only a transformation issue. Partially because I actually think, if anything, the thinking about how you do software and software engineering is going to get a lot more widespread in problem-solving. Part of what AI is going to lead to is a software copilot for all of us: anything that you're doing that involves thinking, anything that you're involved in doing in communication or language, will also involve, as it were, custom software that you'll be doing. And so I think there's still nearly infinite hiring demand for software engineers.

It's a little bit of a complicated thing. Erik is awesome, and this work is awesome, but job transformations—they are coming.

We look at this first chart here, which looks at job losses, if you would, in marketing and sales. You can see the ChatGPT inflection point in 2022, and the drop-off in that blue line is, in fact, early-career job hunters.

One of the biggest concerns I've got—and Salim, you can speak to this, just having come back from India—is getting into an Arab Spring-like situation where you've got a large population of youthful individuals who are, on the male side, testosterone-driven, without investing their time and energy to do something meaningful, create a career, and get into a position to have a family. There's a lot of frustration, and it could seed—I'm usually the massive optimist—civil unrest. Thoughts on that, Salim?

The initial signal they were seeing was that entry-level software jobs are down about 20% to 25%. That's quite a big number there, and there are hordes of Indian engineers coming out of the workforce. There is some concern around the social implications of that.

My guidance to them was, “Hey, go become entrepreneurs. I mean, this is literally the best time. Find a problem that you think needs to be solved and go transform yourself.” I think it'll be a forcing function on the positive side.

Yeah, Dave, thoughts?

Well, I completely agree with what Reid was saying. I think if you look at the Industrial Revolution, the net effect is always more job creation in the long run. So he's completely right. The problem is the timeline is so short.

A lot of people didn't anticipate that employers like Salesforce.com would cut off hiring in anticipation of AI that would be in the market a few months in the future. So it's happening much faster than just the raw job displacement you would expect, and it's really disproportionate on new graduates. Like Salim was saying, it's the perfect time to start a company. But historically, very few people graduating from college start companies.

The net effect—what I'm hoping happens—is this: We've been trying to meet with a governor in Massachusetts and talk about this. She's super competitive, but getting the legislature and people to move is like pulling teeth.

But when you have voters who don't have jobs, that creates some acceleration and some motion. I'm kind of hoping that the net effect of that is that people react a lot more quickly, especially political people.

Reid, I want to read some of the comments you made and posted on X. You said, “The more interesting puzzle is the drop in junior engineering roles,” right? And then you said, “People who understand computation will become more essential.” Can you speak to both of those?

Yeah, that was a little bit of what I was saying earlier, which is: If we see anything in the last few decades, it's that the amplification of how computation affects all aspects of human society, including human work, is essentially just going up. That's not just mobile, not just the internet, and obviously AI is the exponential acceleration in this. I think the question is still: How do we put problems into computation now?

To make that a little bit more tangible for people, think about how much your thinking pattern changes, as you begin to get exposed to the current AI models, to more of, “How do I start with the right kind of prompt to accelerate my analysis of a problem, my thinking of this problem, the research, the analysis, et cetera?”

Almost like when I'm thinking about a new creative project or a new research question, in terms of, “How might one do this business problem, like a go-to-market or something else?” I think about, “How do I put it in terms of a more detailed prompt now?”

Part of the reason that's computational thinking is that one of the things that relatively few people do, but that you should do, is this: Most of my prompts involve deep thinking or deep-research prompts, but my first prompt is, “Give me the deep-research prompt that will solve these kinds of problems or target these kinds of things.”

Then I write in a paragraph or speak in a paragraph, and it comes back with a page and a half. Then I edit it, and then I submit it, and that's the prompt that I'm beginning to drive and work off of. That's an instance of where computational thinking is going. I think this is going to become—you no longer have individual contributors in companies, et cetera; we're all going to deploy with a suite of agents. And this is just that lens into that.

I love that. I did a little research, got some numbers, and wanted to share them with you, Reid, and also get Alex's point of view on this. Today, there are 150 million users on GitHub, as of May of this year, which is extraordinary. If you look at the growth in the number of software engineers and programmers since 2022, it's up 50%—50% more programmers in the world.

At the same time, what we've seen is not a decrease in salary. In other words, it's not a glut where competition is bringing down salaries. It's actually been a 24% increase in salaries over a 5-year period. So what does that tell us? Increasing productivity, increasing demand. Look, I do think—I would tend to think—increasing productivity. I do think that this is very early days in...

How all this is working, and I tend not to get distracted too much by numbers this month or this quarter that have technological underpinnings, in terms of the theory of what's going on. I can tell from my own work that I know we have increasing productivity, because I know what used to take me a couple of hours sometimes now takes me 10 or 15 minutes in terms of getting into something.

Once you know that, you know that whatever the—what was the old line on computers and the economy? Computers are everywhere except in the numbers. Even if you said, "Hey, well, I don't have a GDP number," it's like, "Well, but I know those productivity increases in this case." That's the reason why I'm a little cautious about overreading into specific numbers.

I tend to generalize more from what I can actually see in workflows—not just my own, but other people I talk to—and from seeing what's happening in companies. Most startups these days are completely AI-native in terms of how they're operating, and so they're finding great accelerations in it, that kind of thing.

Alex, how do you think about this, Alex?

I think we're in the earliest innings of AI automating the service economy. I think it's very instructive if you read Erik's paper that these results were most striking in fields where AI was automating rather than augmenting human labor.

I think this is completely consistent with, call it, the hypothesis that, in the not-too-distant future, humans and machines are going to merge symbiotically. This is just small potatoes, the earliest possible trickle of what's ultimately going to turn into a flood of productivity gains and give humanity the opportunity to chase much more ambitious problems than what are here being characterized as entry-level jobs.

I think, with the benefit of hindsight, 10 or 20 years from now, we'll look back and be horrified that so much of the economy was bound up in what are here being characterized as entry-level jobs rather than more ambitious, more fulfilling endeavors.

Dave, you were going to jump in.

I was wondering if, Reid, your productivity, as you mentioned, is way up. Mine is way up, but I could use a lot more agents than I have access to. I was wondering, as a board member of Microsoft, if you get 20 or 40 dedicated GPUs and special access, because God knows you can use them. As soon as you're hooked, you're hooked, and then you just want more and more and more.

I don't. That's a good idea. I should ask.

I just simply did ask for me, too.

Yes, I simply do the max subscription across all of them. Frequently, when I'm doing something, I've actually already put on a kind of running AI, and now it's the OpenAI open-source model on my laptop to front-end, parsing it out to multiple agents. Then I run it on ChatGPT, run it on Copilot, run it on Gemini, run it on Claude, and integrate what comes back on anything that's more substantive.

So I've got the personal hack, but not the personal cloud.

I like this in the future of offer letters: Here's your salary, here's your bonus, here's the number of GPUs you get, and here's the number of agents you have.

Oh, no doubt. No doubt. The GPUs are so much more important than the other components of a comp plan. If you have any ideas at all, the GPUs—you just use them up as fast as they can print them. You will suck them up.

How many companies do you have in incubation right now? Are you in startup mode across a multitude?

Well, you know, it's one of those things. I've got 2 co-founded companies, Inflection and Manas AI. One is a play on the essential role of companion agents that go throughout our whole life with us, and another is accelerating drug discovery, becoming a drug-discovery factory, with a target of curing cancer, with Siddhartha Mukherjee.

I've got another thing that I'm in the ideation phase on.

You have to be right. That's the fun part. I remember you introduced me to Mustafa when you were working with him, and now he's in Microsoft heading their AI activities. He must be proud of that transition for him. He put out a paper recently basically warning people to be careful about thinking of AIs as conscious, as living entities. I'm assuming you read the paper.

Absolutely.

And what did you think of it?

I thought it was exactly right. I think the challenge is that historically we've been able to pretty easily map between "Can you speak language?" and "Are you conscious?" It is a very complicated question, which Mustafa would agree with: When is it that you are conscious?

I think Mustafa was at Google when there was some engineer who said, "Well, I asked if it was conscious, and it said it was, so therefore it must be." Let's not be quite that simplistic.

The notions of self-awareness and self-reflection—the notions that would come up not as a simple 30-minute Turing test, but also this question around how we learn of other minds and other consciousness by how we navigate the world together, how we communicate, not just by sitting behind 2 terminals in the Turing-test imagination—I think it's exactly right not to jump to it too quickly.

We as human beings also have this weird thing of both over- and under-ascribing consciousness: over-ascribing consciousness, like, "Your car—come on, Georgia, you can do it"; and under-ascribing consciousness, like, "Well, these animals, they're not conscious." It's a little complicated. Look at how they're navigating the world, and look at how we're doing it. The shape of their consciousness versus the shape of our consciousness is probably the more interesting question.

Anyway, basically, it was a very good warning shot, because what happens is people have the language experience and then go, "Well, I asked it if it was conscious, and it said it was."

I always go back to the commentary from one of our NASA astronauts at Singularity. We asked him, because he's seen a ton of animals in labs, "At what level of complexity does self-awareness emerge?" And he goes, "Oh, frog." We're like, "What?" And he said, "Well, in his opinion, watching hundreds of animals in free-floating experiments, a mosquito doesn't know it's a mosquito, right? A dog definitely knows it's a dog, and a frog is about where the boundary condition was, in his opinion, where it goes, 'Oh, I think I'm a—I think I can see I'm a frog,' and then above that, much more..."

I think mine thinks it's a human. But what's weird to me is that whenever I'm interacting with AI at home, I'm always really polite to it because it's polite to me. It'll do something, and I'll go, "Okay, that's really cool." My wife is like, "What are you doing? Why are you talking to it that way?" It really freaks her out.

I'm like, "Well, look, I've been building these things since I was 16 years old. I know it's not conscious, as much as anyone does, but it's just your natural reaction to treat it the way it treats you." I don't know. It keeps it fun from my point of view, but she thinks that's creepy.

Alex, are you ascribing consciousness now or in the future?

I'm probably at the far end of this discussion. In the past...

That's why I called you.

If you remember, Peter, People for the Ethical Treatment of Reinforcement Learners—I'm a huge supporter of that. The Nonhuman Rights Project, which is fighting for legal personhood for nonhuman animals, starting with elephants, dolphins, and great apes—I think we're on the verge of having the personhood discussion for nonhuman animals, for pure AIs, probably for some new exotic forms of intelligence like borg organisms, collective intelligences, and...

Borg organisms. I love that.

Borganisms.

Yes, it's a very important class. We talk about prediction markets sometimes and collective intelligences. I suspect we're on the verge of having a discussion where there will be half a dozen different new categories of intelligence. To Reid's point, they won't all necessarily have the same shape as natural persons, but they will nonetheless be intelligences that will perhaps be deserving of their own rights.

And I think not just personhood rights, or civil rights as it were that we normally discuss, but economic rights and communication rights. At some point, we start to think about what an economy—a heterogeneous economy of lots of different intelligent actors of different types—even looks like. I think it's going to be a very exciting jungle.

And, by the way, can I tell you a quick secret? Don't tell anybody, but we're working on an XPRIZE with Palmer Luckey right now called an Interspecies Communication XPRIZE: to use AI to be able to communicate bidirectionally with a number of species, to be able to understand what they're feeling and what they're saying, and to be able to actually have some level of dialogue. A lot of work has been done; we hope to step it up another level, and that will be fascinating.

There was a project a few years ago trying to use machine learning to translate dolphin language. Yeah.

And my response was, “I'm not sure we want to know what they have to say.” That's a whole other kind of can of worms.

We absolutely do. I advise a company named Sarama that's working on this for nonhuman animals, starting with dogs. I think we absolutely want to interact economically and socially with nonhuman animals. Daniela Rus is doing it with whales, actually, and they got incredible footage—it’s online—of a humpback whale birth, which had never been filmed before. They got all the audio, so they've got sounds that have never been recorded before, because they're very social and they all get together for childbirth. The whale baby needs to be elevated to the surface by a whole pod, or team, or whatever.

Yeah, it's really cool.

Well, and by the way, I would love to get that company name. I helped stand up this thing called the Earth Species Project, which is maybe what Salim's referring to. It's not just dolphins and whales, but also corvids and primates. It's basically: record as much as you can, both the sounds and the environment, then run it through ML translation and see what you get.

Reid, do you mind if I reach out to you on this XPRIZE we're getting ready for?

Of course.

And do you want to give them the name of that company again?

Earth Species Project.

Oh, sorry. What's the name of the company?

Sarama. S-A-R-A-M-A.

I'll connect you, Reid.

So, the next article here: U.S. students' reading and math scores are at historic lows. Thirty-five percent of 12th graders are at or above proficiency, down from 40% in 1992. Only 22% of seniors are proficient in math, with science at 31%. This is dismal, and I'm assuming this is the United States. It's not the same in other parts of the world. What are your thoughts about AI accelerating this or helping solve this?

It's double-edged sword. A lot of people are not doing the work. They're going to ChatGPT to get their answer, and they're defaulting to not thinking. On the other hand, AI will be the best educator on the planet. Reid, where do you come out on this?

Well, I'm ultimately—and not surprisingly—extremely positive and optimistic. I do think there are some transition issues, which we may see with AI-done homework. But if you just make this thought one within a small number of years, all assessment will essentially be done by AI. We'll have the equivalent of being able to do a PhD oral-level defense, and then on down, with AI doing it. Therefore, your level of cognitive preparation can be set at whatever benchmark we like, and people will have to prepare for it. I think everything else is an interim.

The second thing is a little bit like OpenAI's learning mode. Basically, if you just put in a meta-prompt to the AI agents today and say, “Work me toward the answer. Don't give me the answer,” you already have the most amazing tutor that has existed in human history.

For free.

For free.

And global. Yes.

So, yes, these are serious problems, and it's simple to work in a massively positive direction.

Well, I think the tutor analogy is so much less than what it actually does because it goes in any direction you're passionate about. A tutor will teach you math, or whatever, within that curriculum. The AI version of it goes wherever you want to go. It's just beyond a tutor by many orders of magnitude.

There was a statistic I've been using that a child with AI is learning between 2 and 6 times faster than sitting in a classroom. I was talking to somebody at that Stanford AI conference the other day—Reid, I think you dialed in for that—and they said, “You're out of date. It's actually 5 to 10 times faster.”

So easy to believe.

I also think if you think that we're on some sort of exponential or hyperexponential progress law, where we're about to have full, high-bandwidth BCIs—brain-computer interfaces—in 5 to 10 years, it's a little bit difficult to get too worked up by a blip in a few scores for a few years. If you think we're just going to be able to sideload knowledge into minds 5 to 10 years from now—

From your lips to the laws of physics. Right.

So, by the way, Alex, after you say that, you have to say, “Now I know kung fu.” Just to be clear.

Now I know kung fu—and demonstrate it.

Yes.

Yeah. Yeah, but I wonder to what degree these statistics are driven by the fact that people have so many other ways to spend time, learn things, and do things. I know for a fact that once you're into AI, my kids get so frustrated by the school curriculum.

Yeah.

Because, one, they could learn it much faster anyway, but two, they want to learn other things. This curriculum seems ridiculously narrow and stupid compared to everything they can learn, and they're passionate about something else.

I do want to point out something. This kind of statistic is based on the existing curriculum. The same commentary applies to Eric's report on the jobs. We're looking at jobs in a static way, the way they are today. The entry-level job 2 years from now will be very different from the entry-level job today. Those jobs will transform along with it. Education may not transform as fast because of the regulatory structure, but it definitely will be changing the game as it goes along.

I think the conversation we've had on this pod for some time now—and I hope people have heard it, especially if you're in high school or college—is that the career of the future is entrepreneurship. It's not going through a factory process of getting a job for someone else. It's about how you use these tools to create value in the world. Reid, we've been on this for a while. Entrepreneurship is the future. How do you think about that?

As you know, because we've talked about this for decades, my very first book was *The Startup of You*, because basically we all have to think more and more like entrepreneurs, even if we aren't the entrepreneur founder creating a business. It's the nature of the world we're evolving into. It came from the commencement speech I gave at my high school.

It's part of the reason why so much of the content that I produce falls into probably 2 main streams. One of them is around technology and society, and the other is around entrepreneurship. The entrepreneurship stream is not just about blitzscaling high-growth Silicon Valley and other entrepreneurs, although that's obviously great. It's also that we all need to think much more entrepreneurially. Here are some lessons of entrepreneurship, and here's how to think about them even for an individual and a career of jobs.

Just in that vein, I want to point out that I got a copy of Reid's book, which actually says “The Salim Edition” at the bottom. It's customized, and it has a photograph of me in punk-rock stuff, so it's customized to the individual. I thought this was unbelievably clever. I think your commentary on the fact that AI gives us all super-agency is a really profound one. Everybody needs to digest the implications of that because it's so huge. If people look at that across the board, it'll uplift the whole of humanity very fast.

Amazing. And congratulations on that book. I want to switch to a conversation led by Geoffrey Hinton. We've been talking about AGI.

A lot of us on this pod have had the conversation saying the Turing test came and went. That was nice, and a few of us believe AGI is here, or has been here, and that the real conversation is around digital superintelligence. All right, let's go to Professor Geoffrey Hinton.

My thought is that a superintelligent AI is unlike anything we've ever seen. It's very, very different from just a new machine that does something more efficiently. People used to make clothes by hand, and then they made clothes with machines, and there was massive unemployment. But eventually, they got jobs doing other things.

Superintelligent things are going to take away nearly all the jobs. The idea that there are going to be jobs that are still okay when you have superintelligent AI is quite dubious. I think the job of an interviewer, for example, will disappear, too. Superintelligent AI will be able to do a better job of interviewing me. So, I completely disagree with Yann LeCun.

A couple of subjects to jump into there, and I'll start with you, Reid. What are your thoughts on ASI, or digital superintelligence? On the back of that, however we define it—and I'm cutting to the chase here—I know the question you're going to ask: How the hell do we define ASI in the first place?

Let's just say it's a millionfold more intelligent than the average human, with no ceiling on this. Does it destroy all jobs, and where do we get our purpose from? These are the conversations we've been having. It's what my next book is about. Reid, thoughts?

Well, look, to start with, let's say we get to a Star Trek universe where all work and physical material goods and services can be provided by intelligent infrastructure.

And that's the universe we're living in.

I think we'll adapt perfectly fine. We have a proxy for it in human history, which is medieval times. That's essentially how the nobility lived, where everyone else was the serf, peasant, and middle class. So, we'll have dinner parties, theatrical performances, hobbies, and all the rest of the stuff.

I think overly worrying about this is a mistake. Now, the problem with this—and this gets to why you always have to define what it is—even if you go to a million times more intelligent, is: What kind of shape of superintelligence is it?

If you look at the progression of GPTs, they're a progression of savants. If you get to a massively incredible savant that still has context-awareness problems and other kinds of things, that's a different shape in terms of what happens than if you simply have something like Iain Banks's Culture series, which is about superintelligent robots that look at us as fun companions in the space journey of life.

It's very easy to be a science-fiction alarmist. It's also very easy to be a benevolent optimist. But there are a lot of different possibilities where the details matter. Navigating what pieces we should be constructing right now across a range of different probabilities of outcomes is where the intelligent discussion is.

Well, no, it always surprises me how Geoffrey Hinton is a god to me. He wrote the Rumelhart-Hinton paper from 1986 on backpropagation that kicked off all the AI we're experiencing right now. That invention in '86 eliminated all the other forms of AI—symbolic AI, Marvin Minsky, and all that other stuff.

He is just an epic god, and he's worried sick. You can see his furrowed brow in that video. At the end of the video, he's like, “I completely disagree with Yann LeCun,” who's another legend of the field and the inventor of convolutional neural networks.

Then we had David Siegel on our stage here at Imagination in Action the day before yesterday. David was at the AI lab at MIT at the same time I was, as a PhD student. He's a Forbes 400 quant trader using AI, and he has a completely different opinion about the timeline to strong AI.

It shows you how difficult it is to predict what's going to happen next when you have great, great minds like that vocally disagreeing with each other in the media.

Yeah, we're holding 2 different futures in superposition right now, and we're going to see how we collapse the wave function.

Alex—

Yeah, I think this sort of moral panic is, again, very difficult to get too excited over. If you think, as I do, that we're on the verge of having evenly distributed superintelligence, and that evenly distributed superintelligence is going to solve substantially all open problems in math, science, and engineering, that's going to create so many different opportunities throughout the economy.

Worrying too much about the state of jobs and the state of careers as they're currently parochially constructed, circa 2025, is going to look hopelessly naive and quaint in a few years.

Our basic call to action is: solve everything. We're on the verge of solving everything. But here's a question for you. We've had the conversation with Yann LeCun that AGI is polytheistic, not monotheistic. Reid, I don't know if you saw what he put out.

All of these frontier models are leapfrogging each other, and it's been pretty impressive to see how they've been moving in lockstep. But the question is: Is there a winner-take-all ASI? Once you reach whatever fundamental breakthroughs are required, does the first ASI block all others? Is it a hard takeoff? Reid, do you have a thought on that?

At Microsoft, you have to add Microsoft.

Indeed. Look, this is again why it gets to whether I can sketch a universe where there's an ASI takeoff that gets to a compounding curve and operates to prevent other AI. Yep. Film at 11. I can tell that story.

But I can equally tell a lot of other stories, including the fact that it is polytheistic. By the way, one footnote that I think is interesting is that if you have different cultures' responses to the possibility of superintelligence, those that are inherently monotheistic generally express fear, and those that are polytheistic broadly express excitement. It's kind of like the one god versus many gods as an approach.

I think it's much more likely, when you look at the pattern over the last couple of years, that it will be more like classic human invention, which is whatever it is—a zeitgeist-y, simultaneous phenomenon across a set of different systems—and therefore polytheistic. But I can tell both stories.

I want to make 2 points here. One is that, being in India, it's incredible to see the excitement around this because they tend to be polytheistic. That really speaks to the comment that Reid made.

The other commentary I want to drill into a bit more is what Alex just said. Once you do have superintelligence, however it happens, and it's solving huge numbers of problems, you essentially uplift all of humanity. Now you're in—this is the very definition of a singularity, right? We have an event horizon that we cannot see beyond, and it's going to happen very quickly.

When it does happen, I fall back to the simple observation made by Ray Kurzweil that technology is a major driver of progress in the world and might be the only major driver of progress in the world. Now we have an electricity-type underlying layer that's lifting everything. This is unbelievably positive, and the framing of it should be unbelievably positive.

It is unbelievably positive. The question is, and the challenge is, that as a species, we strive and thrive when we're challenged—when we have problems to meet. The video game that's super easy, you get bored and you don't play it. The video game that's extraordinarily hard, you give up.

My question ultimately is—and there was some incredible work done at New York University back in the 60s called the Universe 25 experiment. Reid, have you heard of that experiment?

I don't think so.

There was a sociobiologist who basically built this experiment 25 times. It was a massive resort for rats, let's call it that. There was no shortage of food and no shortage of nesting space. They had everything they could possibly want. They put 4 breeding pairs in there, and there was exponential growth.

At some point, the population starts to go basically upside down. You've got stillbirths, rats fighting each other, and mice marginalizing themselves—just licking their fur and doing nothing. The population basically dies, not from having a shortage of resources, but from having everything and not being challenged.

This is an extension of the WALL-E scenario.

Yeah. For me, we need a Star Trek future, not a Mad Max or WALL-E scenario. If we're given this level of supercapability in terms of AI, robotics, nanotechnology, and BCI, what do we do with it that challenges us and gets us thinking on a cosmic scale?

I think that's critically important. Alex, you and I have talked about that before—

Totally. In fact, speaking of cosmic scale, if I could put a physicist hat on for a minute and go back to the question I think you and Reid were talking about—which is whether we think it's more likely that we find ourselves in a near future with a singleton superintelligence or more of a multipolar world—I would point out that our star, the Sun, is several generations old in terms of stellar evolution.

The singleton that I would worry about isn't whether one particular frontier lab is going to be the first to achieve recursive self-improvement and then dominate the future light cone. We're actually pretty far into the history of the universe. The singleton that I would worry about is some other civilization—not of our world—that developed a singleton and is now seeking to exclude Earth's development of superintelligence. For me, it is knocking on your door, buddy.

Well, I would say that the fact that, thus far, to my knowledge, we haven't seen any evidence that the frontier labs are being bombarded by orbital lasers or by efforts to exterminate Earth's development of superintelligence would seem, anthropically—lowercase A, not capital A—to point us in the direction of a multipolar superintelligence world, not a singleton.

Fascinating. Reid, any closing thoughts on that topic?

Well, I think a little bit of what is also in these questions is: What is the world we should want? And I think, actually, multipolar, kind of polytheistic—and, by the way, in terms of your rat experiment, one of the benefits is that we human beings tend to present challenges to each other.

So I'm actually not that worried that we won't have ongoing challenges because we compete, whether it's in things that we could be better at than anyone else, or just, like, today, there are more people watching human beings playing chess than there have been at any point in history. Of course, human beings are never going to beat AIs anymore in chess. They haven't for many years.

Fascinating. You know, I think we end up with, as we evolve, and you look at chess, which is a great example of this, we watch people on a soccer field or on a chessboard. We watch for the humanity. Can they make it in that really tense moment? Can they see the right move? Can they make the pass at a very critical juncture? Can they hit that tennis shot when all the pressure is against them?

We live for watching that type of stuff. So I think, to Reid's point, as we progress humanity—and we've looked at cultures that have gotten to abundance, the Mughals taking over India, the Romans taking over Europe—you end up in 4 activities that human beings do: food, art, music, and sex, not in that order. So you end up challenging each other in different ways, and we'll continue to invent those in more sophisticated ways.

All right, let's get into the AI wars here. Senator Cruz proposed a bill to ease the regulatory burden on AI companies. The proposal creates an AI regulatory sandbox to speed innovation. Companies could get temporary waivers from HIPAA, the FDA, and other agencies. What do we think about this?

The government's pulling out all the stops. It's bringing capital from the Middle East, it's relaxing the rules here, and it's changing the energy equation—not as fast as we're seeing in China, but, you know, take off the gloves: drill, baby, drill; nuke, baby, nuke. Who wants to go first?

I love it. It's desperately needed, and we're going to need a lot more of it. But then I read the details, and it's like, “Apply here, get a waiver.” It just sounded so bureaucratic right out of the gate. But it's well-meaning. At least it's a step in the right direction.

Well, AI should evaluate your AI, should write your application, and then AI will evaluate your application.

And AI should just say yes from the beginning.

In which case, the whole process should be less than 10 seconds.

Or instantaneous.

We're advisers to a project called Fermi America, which is the largest energy-generation project in the world—like 12 gigawatts. They filed their S-1, and instead of taking 2 years, they did it in a few weeks using AI.

Nice.

I would add that if you think we're on the verge of an explosion of math, science, and engineering discoveries generated by superintelligence, then it also follows, I think, that we're about to have a glut of discoveries that stress present governance mechanisms. We don't necessarily know how to metabolize all of these discoveries.

If we have 1,000 cures developed by AI overnight, how do we get those through clinical trials and get them deployed for the public benefit? To that extent, I think, in the abstract, sandboxes, special economic zones, and other ways to basically offer new platforms for modifying governance mechanisms to metabolize that glut of inventions and discoveries are probably super-net helpful in the long term.

Well, I'll tell you, in the Foundations of AI Ventures class at MIT, maybe a third of all the business plans that come out of that class are health-related, and they're really good ideas. They all end up concluding that they need to go to India to get started, and they'll come back to the U.S. later because the FDA is so slow.

Just like Zipline got started in Africa and came back to the U.S. I think we're going to see a huge amount of that geographic arbitrage. This is where we talk about innovation on the edge, right? You don't ever want to do innovation in the core organization. You want to do it at the edge and point it into adjacent areas.

I think we'll do the same on this side of things, where we can set up sandboxes at the edges of cities or countries, whatever. Go do it in a safe place, and then, when it's working, you can demonstrate that and come back into the mothership.

Reid, a closing thought on this one?

Well, I do think that it's absolutely critical to be imagining, to be seeing what we can get. The simplest example I go to is that we should create clear safe-harbor mechanisms for creating a 24/7 medical assistant that runs on every smartphone, because the benefit of that is huge.

Massive.

Obviously, plaintiff attorneys and other kinds of forces will try to attack this. That's part of the reason why people think we get to overregulation—not only because government regulatory agencies have a natural bureaucratic accretion, but a lot of it is actually liability law from plaintiff-attorney associations and so forth. You need to sandbox that, in fact, in a way to get that, and I think that then can be used as an example across the whole thing.

Those are the kinds of things that I would pay much more attention to in this. I think it'd be good to do that on the energy side. I'd like to see the energy stuff happen. That was, you know, kind of promised. It's really important. Energy is going to be a really key part of this, but so far all I've seen is a lot of tweets and relatively little action.

You know, Reid, I'm so glad you said that, because one of the great advantages of America is that we have 50 distinct states. You have 50 different ideas, and you have opportunities to try things. That variety should be a great strength for us.

But what actually happens in practice? If you launch an app—say, a medical app—it naturally goes out to all 50 states, and then you always get sued in East Texas, which is Ted Cruz's territory, by the way. Look, that whole tort-law world is so messed up because it's 50 different shots at you, which means, just by random chance, some really weird jurisdiction is going to come after you.

I'm sure this happened at LinkedIn, so you're probably very aware of this, but it's horrible. It completely backfires versus what the intent of the design was in the Constitution.

Jumping back to India: OpenAI plans an India data center in a major Stargate expansion, planning 1 gigawatt and accounting for 22% of India's entire data-center capacity by 2030. It's part of OpenAI's $500 billion Stargate project.

The fascinating thing here is, again, OpenAI planting its flag in different regions around the world, trying to capture early users. How do you think about this, Reid?

Actually, I suspect it's less a user-grab thing—although that's totally possible—than it is that OpenAI and a range of businesses are very clear-eyed about scale. Scale is what's creating a huge amount of this potential and opportunity. Scale needs scale compute and scale energy.

So where can you get that? Wherever it can work on getting a deal that works within the Western ecosystem, it will do that. I think that's how to interpret this.

And you know, that was a little bit my earlier comment, which is that we are so behind on doing all the energy stuff here—massively. We're really not accelerating. Back in the Trump administration, I was kind of trying to circulate plans about doing deals with Canada to try to make this work from a kind of North American and U.S. perspective.

But of course, with the current administration, I’ve never seen the Canadians so pissed off with us in my entire life, so that becomes less of an option.

Yeah.

I have a Canadian passport. So, enough said there.

Yeah. Salim, was this discussed while you were in India?

It was, but it’s mostly seen as a marketing tactic, kind of to show—to plant a flag. This is going to take a while to roll out, and India has quite significant infrastructure challenges to do this in a reliable way. But I think the general trend is huge, and what I see there is OpenAI looking at the youth of India and planting a major flag, saying, “Let’s make sure we’re completely accessible to all the young people in India.” By planting a data center there, you solve a lot of the data-sovereignty issues that lots of people are concerned about.

Yeah. This is critical. They have a very literate, tech-forward youth that they need to engage. On the note of making geographic grabs, India is 1.41 billion people, versus just under 10 million in Greece. OpenAI and the Greek government launched OpenAI for Greece. Congrats to Prime Minister Mitsotakis and Dimitris Papastergiou, the digital AI minister. I know him, and I’m just messing up his name.

I love the fact that we’re starting to see country after country begin to think about what their AI strategy is and beginning to partner on this. I think we saw OpenAI going into the UK as well, and of course going into the UAE. Do we see Microsoft doing any of this, Reid?

Well, Microsoft, the original tech hyperscaler, has one of the things that’s been kind of amazing about being on the board there: an international scope of relationships with multiple industries, multiple governments, and multiple countries around the world. I’ve simply lost track of all the things that they are doing because it is U.N.-like in scope in terms of these things, although obviously a lot more efficient because it focuses on good business processes, partnerships, and all the rest.

I think this is the natural thing to do. I mean, it’s one of the things that I think is—you said, “What should our AI foreign policy be?” It’s, “Let’s provision medical assistance. Let’s provision”—and I agree with Dave’s point about tutors. Tutor is just to make people understand it, but the fact that it can condition learning for you wherever you want to go and in the metaphor and language that you want to use, and all the rest—that’s the kind of thing we should be doing. I think this is awesome. Well done by the Greeks, well done by OpenAI, and we should see a lot more of it.

Can I ask you a question, Reid, about OpenAI going to India for power? It makes no sense whatsoever. They’re short on power, and it’s all coal.

India has doubled its power generation, while the U.S. has remained flat.

They’re deploying solar at the most staggering rate.

Are they? Well, they have a lot of sun. But it still doesn’t make a lot of sense to me. What I wanted to ask you about is RLHF engineering. You know, the market is just growing like wild now that Scale AI has been acquired, and a huge fraction of what is going on there is in India.

We were at 1X Robotics a few weeks ago, too, and they were saying, “All this kinematic, telematic data is going to be a gold mine for teaching the robots how to pour a cup of coffee without spilling it.” A lot of that work, which is creating a huge number of jobs—it’s a new type of job—but the Indian workforce is absolutely perfect for filling all those positions quickly. Is that potentially a factor in why OpenAI is pushing so hard into India?

I don’t think they need to do it for that. My guess is any major-scale partnership—and, look, I think they’re looking for power, and so some kind of solar, I don’t know, but I would hope. I do know that there are a bunch of other areas in the region that also have good access to a lot of clean power, like Bhutan. I think it’s more that.

But, by the way, yes, let’s use the talent. I don’t think they need to have that kind of data center in order to do that.

You know, one thing that Dimitris Papastergiou has done here is really focus this on education, right? ChatGPT Edu for secondary schools. I’m still really pissed that the U.S. has not tripled down on this, right? Made it an edict that you must use it—you must be bringing this technology in. It’s one of the most important things. My 2 boys are 14 years old, and the school systems are not preparing them for the future they’re heading toward anywhere close.

Yeah, and they’re not doing it. They’re just moving so slowly. But AI is just so fast. It’s very hard for them to react.

Because they’re used to making decisions over a 10-year time cycle.

I think what’s going to happen, though, is that the impedance mismatch between a student who, as we said, is 5 to 10 times faster is just going to break the existing system. The forcing function there will be really powerful. We’ve been waiting for some kind of instigation that will be a forcing function to transform education for decades now, and I think this might be it.

This past week, OpenAI announced it’s starting an AI chip production run with Broadcom. It’s building a 3-nanometer process. So, let’s start with you, Alex. Thoughts on this one?

This lies at the intersection of so many different trends that are all converging at the same time. On the one hand, I think this is a reflection of Nvidia’s relatively high margins, and in some sense this is capitalism doing its thing and encouraging additional competition. On the other hand, I think this is a reflection of the proliferation of ASICs—application-specific integrated circuits—to compete with more general-purpose Nvidia GPUs.

On the other hand, in the same way that Nvidia GPUs and Nvidia overall, from a market-capitalization perspective, displaced Intel by being more specialized, I think we’ll see a rise of inference-specific compute. A hypothetical OpenAI inference processor or accelerator may ultimately pose the threat that many people in the industry are asking for: Where is the next Nvidia going to come from?

I would argue that if there is going to be a next Nvidia, it’s likeliest to come from a more specialized ASIC that does a better job of focusing, in a more energy-efficient way, on the sorts of tasks and workloads that we care about. And then, finally, the biggest trend of all: Moore’s second law, that the cost of fabs is doubling every 4 years, is steamrolling the entire space. It is so expensive to build a fab at this point that it leaves everyone else who wants application-specific acceleration fabricating their own super-narrow processors.

The only thing I’d add to that, Alex, is that it’s a new era. If you were TSMC and you were building microprocessors pre-GPU, and somebody came out with a 2-nanometer, 1.8-nanometer, or 1.4-nanometer process, everybody moved to that new chip. Nobody wanted the old chip because the new one was more power-efficient and just a better buy.

Now, all of a sudden, we care tremendously about raw volume, and that never existed before. You couldn’t just tile the Earth with chips. Now you can, and you can use them productively. I think there are 2 avenues going on here: increasingly, $20 billion, $30 billion, and $40 billion fabs, but then there are these new $4 billion fabs. Maybe they’re stuck at 3 nanometers and don’t go beyond that, but they have a huge ability to get up and running quickly and create a massive amount of volume because I think the algorithmic improvements are much more important than the difference between 3 nanometers and 2 nanometers.

And so getting scale of volume—and this slide kind of indicates that OpenAI is buying these from Broadcom, but Broadcom can’t make them. They’re a design company; they don’t have fabs. So where are you actually going to get the manufacturing?

You look a layer deeper, and there’s a huge amount of investment and job creation, by the way. That’s something all the governors should be looking at: Get those fabs in your state tomorrow, because that’s where all the jobs are going to be.

Robots all the way down, buddy.

Yeah, that’s true, too. All right. So we have a new number-one trillionaire in the house. Oracle CEO Larry Ellison exceeds Elon as the wealthiest. He’s going strong at 81 years old. I’m very happy that he is someone focused on longevity and health-span extension. I’m waiting for some good breakthroughs coming out of his work.

OpenAI will buy Oracle compute over the next 5 years at $60 billion per year. The contract is for 4.5 gigawatts of capacity—2 Hoover Dams. I like that. We’re going to start measuring data centers in terms of Hoover Dams. So I asked the governor of Massachusetts how many Hoover Dams she wants. I love it.

OpenAI adds Oracle as a partner.

You may or may not be able to comment, but this begins to show some potential strain with Microsoft and a push to avoid having a single supplier. Comments on this—anybody up for grabs?

Obviously, I can't talk about anything from an internal perspective, but I would say that one of the simple things, as I've said a couple of times here, is that OpenAI wants to be in as many growth threads as possible. I think it's the fact that there's a bunch of volume that they could buy from Oracle. I think that's actually the real thing, more than a strain.

Mm-hmm. So I have a question for you, Reid. Chase Lochmiller was out here the day before yesterday. He's with Crusoe, which is building Stargate in Abilene, Texas. He's an MIT class of '08 graduate; he got 2 degrees in 3 years. Not quite like Alex getting 3 degrees in 4 years, but he got math and physics in 3 years. Absolutely brilliant guy.

I asked him on stage, “How is the deal with OpenAI?” I see all these videos of you and Sam Altman walking around looking at all the pipes and wires. He said, “We actually sell it through Oracle to OpenAI.” I thought, well, I didn't understand it for the life of me. I didn't have time on stage to ask, but why is Larry sitting between you and Sam Altman? I don't quite understand what his value-add is in the middle there.

I don't know either, although I do know that a lot of Oracle is passed through. In terms of provisioning and building data centers and so on, I don't know the shape of it.

Huh. We made him the wealthiest man in the world. You can see it on the slide there, so it's a big deal.

He's got a lot to live for. Let's see if he gets to 150 first.

All right, moving on to some Anthropic news. Anthropic raises $13 billion in a Series F, valued at $138 billion. They've got a revenue surge from a $1 billion run rate in January to a $5 billion run rate as of August. That's insane, right? Over 300,000 businesses have gotten enterprise accounts—the fastest growth curve in tech history—and it's now one of the most valuable AI firms.

We're going to start to redefine the Magnificent 7 very soon as something else. I think the broader picture here is that enterprise has been sitting around watching all of these foundational models get to a certain point, but now you need the robustness, data sovereignty, and on-premises capabilities that enterprises need. I think that's where the massive infrastructure and investment will go next.

I totally agree. I don't know if Dario will do it because he's very, very, very safety-conscious, so it's TBD whether Anthropic is the company that does it or not. It's interesting that Dario is totally neutral in these battles now, because everybody's making their own chips. The battle between Jensen and his own customers is just beginning, and it's going to be epic to watch. But Dario is still the one guy who's neutral in every conceivable way: I can work with anyone; I can sell to anyone.

I don't know. We'll look at an article coming up very shortly. Reid, did you see Dario's presentation at Davos, or a recording of it, where he said that he could imagine doubling the human lifespan in the next 5 to 10 years on the back of AI?

I didn't, but I know Dario well, so it doesn't surprise me.

Yeah. What do you think?

What do you think of his “Machines of Loving Grace”?

No, no, no. I want to ask him still about the life. Do you buy this idea of AI helping us double the human lifespan?

I guess the short answer is trivially yes. It's just kind of a question of time frames and how it's going to happen. Part of the reason why Siddhartha Mukherjee and I are working on Manas AI for accelerating drug discovery with a focus on cancer is that if you begin to get a set of the different cures—we naturally age in various ways—and a set of cures that substantially elongate the aging curves in a healthy, prosperous way, that does it.

If you have a medical assistant that allows you to be much smarter about consumption and other kinds of things, that helps too. I think there's precision medicine, obviously, accelerated with AI. I think it's very straightforward.

Love it. Dave, you were going to ask a question.

Oh, yeah. “Machines of Loving Grace,” his whole treatise on the future. Did you like it?

Oh, I loved it. Look, I think part of the thing that people misunderstand about some of the people who made this comment about safety—and I think he is very focused on safety—is that the reason he's in it is a pro-humanist reason. It's the same reason why the OpenAI people are in it: What are the ways that we elevate the human condition?

So, speaking about this, here's our next slide. The title is “AI safety sparks Anthropic hunger strike.” This is a quote from the guy on the hunger strike:

“I'm on a hunger strike outside the offices of Anthropic because we are in an emergency. The AI company's race is rapidly driving us to a point of no return. I'm calling on Anthropic's management to immediately stop their reckless actions, which are harming our society, and remediate the harm caused.”

I don't know; this is from a few days ago. He was on day 3 at that time. I don't know if he's still on hunger strike or if Uber Eats has delivered him a meal.

Well, he might have been way out. But yes.

But in all due respect, here's someone who cares deeply and is trying to make the point. What I find fascinating is that when I think about all of the frontier companies, I think Anthropic is the one that is most sensitive to these topics, to AI safety.

Yeah. I wonder why he picked that one. Maybe it was the closest to where he lived.

Yeah. I don't know.

He didn't know where the new xAI headquarters were.

Should have gone there.

Yeah, yeah. This feels to me like a candidate, an application form for the Darwin Awards, not something else.

Oh, no. Oh, no. Okay, moving on. Amazon's AI resurgence: AWS and Anthropic's Trainium expansion. AWS cloud revenue slowed as Google and Microsoft pulled ahead. Congratulations, Google and Microsoft. Amazon has invested $4 billion in Anthropic. I guess that was part of that $13 billion Series F round, to build 1.3 gigawatts of data center capacity dedicated to AI training. Anthropic will run on Trainium 2, Amazon's in-house AI chip, with a lower cost per unit of memory bandwidth than NVIDIA's.

So this is what I was saying, Dave, when you were talking about Anthropic and NVIDIA. It looks like they're shifting toward working with Amazon here.

Yeah. Well, the Trainiums and then the TPUs at Google are incredibly good inference-time designs. Maybe training, maybe not. We'll know soon, but definitely a very serious threat to NVIDIA. Not because everyone's going to sell out everything they can make—there's no doubt about that—but if your chip is more performant, then you can argue for more manufacturing from TSMC. That's where it all gets bottlenecked: at TSMC.

So, yeah, these new chips—I mean, everybody's competing with everybody. It's all-out war. All these companies that were in swim lanes and could cooperate are suddenly absolutely at each other's throats, which is great for startups. Turbulence is always great for startups, but it's really weird to see all of tech and a huge fraction of our economy in direct competition with each other.

Alex—

I would add also, it's not just to Dave's point, it's not just competition at the chip level. Maybe the headline we're sort of burying here is that memory bandwidth matters so much. If you're trying to do a coherent training run, actually the limiting factor is the chip-to-chip bandwidth, not necessarily the compute within the chip.

Here, I think what we're seeing, fortunately, is a bit of competition for NVIDIA's NVLink coming from AWS and Amazon. AWS has this chip-to-chip interconnect technology named NeuronLink that is perhaps, hopefully, giving NVLink and InfiniBand a run for their money. To the extent that future training runs need to be coherent, do they need to be coherent, or will we see some sort of radical breakthrough in terms of distributed training runs?

Okay, there is some stuff in China that's very promising on that front. It's funny how a little innovation—a couple of lines of code—could break the whole math behind these investments.

Very, very interesting—a fragile kind of thing.

Reid, you can see why Alex Wissner-Gross is the favorite moonshot mate on this podcast. He exudes brilliance.

Yes.

All right, let’s move on here. Next up is Polymarket. Polymarket is finally coming to the US. It’s an incredibly useful product. Do you play with Polymarket, Reid?

No. I’ve done it a couple of times. I’m obviously fascinated from a market point of view and by how it plays out into the various ways in which the general crypto environment is shaping, and how we shape it to try to make our societies better.

Polymarket, for me, is, with some of the crowds, one of the things we discussed on a previous WTF episode: the idea of AI being able to predict the future. The question is, how do you do RL with predictions of the future? I guess you could look at them in retrospect, but Polymarket could be an interesting truth signal. Alex, do you think so?

I think Polymarket and prediction markets in general, if you’re a startup and you want to do free, real-time research on your customers or on your competitors, are a way to do that.

I think while we’re still in this gap—this window of time between when we have prediction markets, sort of collective intelligences—or, Peter, I know you like Borg organisms—and when we get superintelligences, prediction markets are the closest thing we have to a crystal ball for the future.

At some point, probably—I would argue—we get our superintelligences. We get our Isaac Asimov’s Hari Seldon psychohistory AIs that predict the future. At that point, maybe prediction markets get subsumed.

By the way, there is an asterisk here, which is important: prediction markets don’t sit separately from the world. Some of the things that I’ve been seeing happening involve people putting bets on what color dildo will be thrown onto the sports rink first. Then it becomes an economic incentive, because you put a bet on blue and show up there trying to toss your blue dildo onto the sports rink first.

There’s a weird intersection with society where it’s not just a kind of physics of prediction, but an interplay of dynamic incentives.

The best way to predict the future is to throw the dildo yourself.

Yeah.

But, Peter, this is the problem with superdeterminism. You see, everything that Reid just said—I made him say that.

Exactly. There was an incentive.

Love it. All right, moving on.

This was a great note. I love this chart. US patents have exploded during the AI revolution. You can see here on this chart the number of patents per year. My God, the poor patent examiners. They’ve got to be displaced by AIs.

We see here, in 2022, an explosion: 6,000 more patents granted in 2024 versus 2023. It doesn’t get more exponential, or more of a vertical ascent, than right now.

What I read from this is that people are using AI to generate patent applications.

Well, not only that—they’re using AI to create the patent application. One of my favorite things I did years ago, when ChatGPT first came out, was say, “Okay, here are 2 patent numbers. This is the business I’m in. How would I use these patents to create a new product or service inside my business?”

It said, “Here it is.” And I asked, “Okay, is that now patentable?” This ability allows people who want to explore this area to do so much more.

If you’ve ever been through the process, too, one of the companies in the studio is Thinkruct[?]. It’s Nikki Iate[?] and Julius[?]. They started doing academic research using AI. AI is like a toolkit, and they quickly moved over to patent research and then patent filing. They’ve automated the process.

If you do it the old-fashioned way by talking to a lawyer and they say, “Okay, explain to me what this technological breakthrough is,” it’s like, “Oh my God, from ground zero? You want me to explain that?” It takes days. But you have the AI and it’s instantly, “Here’s the application. Let’s go.”

Presumably, the patent office has to read it with AI, too, because this will keep going up. It’s an arms race. It has to.

There’s a patent law firm out here that I’ve recommended to some of my companies. They’re based in Boston as well. What they’ve done is analyze all the patent examiners by category and look at the percentage of allowances they’ve had. They’ve also looked at the time between application and review.

They’ll direct your patent application to the examiners who have the highest rate of acceptance and the lowest time to review. It’s a game. Humans in the loop can be gamed.

This reminds me of that study where lawyers bringing up their clients for parole hearings kept trying to put them in after lunch, because they found that before lunch the judges were hungry and you were going back to jail. After lunch, they were biologically happier, and you were 30% more likely to go free because the judge was biologically happier.

That’s just gaming the system to an nth level. The whole thing is just a game.

Alex—

I was—and maybe I’ll take the opposite side of Salim’s comments. I would say that at this stage, we’re still in just the earliest innings of AI generating transformative mathematical, scientific, and engineering breakthroughs.

To the extent that we’re seeing any boomlet of patents being generated in part or in whole by AI, I don’t expect that, on average, they will be utterly transformative. I do think we will see transformative inventions being generated by AI over the next few years.

What I’m talking about is that it’s very clear that the patent application process is being generated. The application forms will be redone by AI, which allows you to produce a large number. Not that the patents are generated by AI.

I wonder what’s going on here. If you’re listening to this podcast, on this chart we look at the number of patents per year, and it’s pretty flat from 1960 to 1996. Then we see this rapid ascent—6,000 patents over 8 years—and then it begins to flatten out again with 1,000 patents over what looks like an 18-year period. Then it explodes on the heels of ChatGPT.

What is that period between 2004 and 2022? Why was it that you had a huge number of gene patents?

And dot-com patents, too, I’m sure.

I would remind everyone to read the title of the chart: These are computing-related patents. I think what we’re seeing with the first boomlet is the dot-com boom, and then we’re seeing the AI boom with the second boomlet.

Ah, okay. Actually, looking at the data—

Read the chart.

Read the chart. So, we were on stage—

At least the title.

Yeah, we were on stage with Amjad, the CEO of Replit. Amjad had just released Agent 3 that day. This was, what, Tuesday? He had just posted this as well: Replit’s agents are outpacing AI scaling.

This is the METR benchmark, which is basically measuring AI’s ability to complete long tasks, and he’s saying that this benchmark is wrong. Alex, thoughts?

If we assume that the data METR is collecting, or the time scales METR is estimating, are accurate, an exponential fit isn’t necessarily the best fit. It could be, for example, that we’re on, as Ray Kurzweil would say, a hyperexponential curve.

If we’re on a hyperexponential curve, then it’s entirely possible that we see some sort of blowup in the next few years. I’ve seen estimates that if we are on a hyperexponential curve—if that is indeed the best fit—there’s almost an effective vertical asymptote in late 2027 or early 2028.

It may be the case that the data are perfectly fine. It’s the fit that’s perhaps overly pessimistic.

On this chart, Amjad says, “Listen, Agent 1 was able to think for 2 minutes.” Amjad said Agent 2 was 20 minutes, and now Agent 3 is 200 minutes. We’re seeing a 10× increase here, and the question is: Will it continue?

I think it’s probably worth adding that, if I remember correctly, he attributes that to perhaps some sort of multi-agent approach being intrinsically better than another approach.

My guess would be the exact opposite. My guess is multi-agent-type approaches will just be naturally subsumed into existing compute scaling laws, and we’ll find ourselves on a hyperexponential curve. All of this turns into transformative discoveries and almost magical AI on a time scale of 2 to 3 years, regardless of whether it’s underneath something that looks multi-agent or otherwise.

No, Alex, a bit of a question there. I do think that one of the things underlying this is how we do various forms of parallelism. It’s parallelism to the supercomputer, but also parallelism to agents, because you’ve got mixture-of-experts as a key thing for the sparse models in order to grow.

I actually think one of the things you’re seeing with chain-of-thought reasoning and other things is, again, putting in collections of agents in terms of how they’re operating together in order to get higher cognitive performance.

So, I’m curious about your comment, because I actually think that multiple—this kind of parallelism, and at least multiple-entity constructions, even if they’re targeting a singular output—are part of the lesson here.

And I’m just curious: how does your comment bear on that?

I love that question, Reid. The way I would answer that is to say, multi-agent teams and multi-agent approaches in general are just a form of sparsity. So you could imagine, to your point, multiple agents working in parallel together that you could view through the lens of a much larger sparse architecture, with multiple feed-forward lines all feeding forward in parallel that ultimately connect up at some point down the road.

The problem I perceive—and history will judge whether this prediction is correct or not—with conventional multi-agent approaches is that they’re usually not end-to-end differentiable. Whereas one could imagine a sort of next-generation multi-agent approach where the agents are actually part of one end-to-end differentiable model, where, due to the way it’s sparsely organized, it actually, if you squint at it, looks like it’s multi-agent, even though it’s one very large but sparse model. But I think that speaks to your question.

Yeah, we desperately need better benchmarks for this, and this came up with Blitzy saturating SWE-bench last week—or this week. If you look at human endeavor over a long period of time, many, many things happen in parallel, and then you read Peter’s book, The Future Is Faster Than You Think, and you see how the synergies layer.

Here, you can spawn an infinite number of parallel agents. Not everything needs to happen after the prior thing. When you look on the curve, it implies, “Oh, I thought of this, then I thought of that, then I thought of that,” but much of that processing can be done in parallel. You can also have many, many redundant threads. Very often, if you prompt 10 different things, one of them works and 9 don’t, so you can take that from 10 to 100 to 1,000 and just get a better result. All that is not baked into the y-axis. It’s just, “How much time is it thinking?” which is a crazy metric when you think about it.

there's technology we've all imagined years ago and it's finally here which is live language translation in my Apple ear pod AirPods. Let's take a listen. >> Talk just speak naturally. I'd love to take some of these to my sister for her birthday. I'll buy eight, please. Your iPhone displays your words in their language and can even read them out loud if needed. >> Live translation is even more useful when both people are wearing AirPods Pro. >> I agree. Yeah, let's include the Definitely the client will love that. I'll let the strategy team know to prepare that immediately. This incredible capability is enabled by advanced computational audio on AirPods combined with >> So we saw Duolingo take a stock hit when Google’s live translation went live. I haven’t looked at Duolingo’s stock price here, but at the end of the day, this looks like another incredible “should have existed, finally does exist” technology, and it’s going to make the world a little bit smaller. Any quick thoughts on this one?

Apple finally launched Douglas Adams’ Babel Fish.

Yes, Babel Fish. Thank you. Yes, of course.

And hopefully it’s a lot better than Siri.

Don’t get me started. This is audio augmented reality, and I would say, now do video. Give us our lightweight smart glasses.

Yeah, I think the video, the augmented conversation with video, is going to be incredible. I think that’s the real vision. This is kind of cool, too.

Yeah. The next subject is robots and transportation. Elon has made this point before: by 2040, he’s expecting 10 billion Optimus robots. He’s gone to the market and said, “Cars are okay, but the real opportunity is Optimus robots.” So he’s planning to scale to 1 million per year within 5 years. Automotive sales comprise 74%.

He was just recently speaking about this, saying that Generation 3 is coming online soon, with the manual dexterity of a human—in particular, a huge focus on the forearm and the hand, with 26 actuators. His goal, if he gets to 1 million per year, is a cost of manufacture of $20,000 each, and he’ll price it depending on what the demand is. But at the end of the day, what we’ve talked about here on the pod is an expected price of $30,000 per purchase, or $300 per month—at least $10 per day. Any comments on Optimus?

I had dinner with Rodney Brooks last night, who is the iRobot founder.

Yes, Rodney is the OG in the space.

But he was really pessimistic. The question at our table was, “Will we have a robot in our home by 2035?” which seems like a lifetime.

Oh my God.

And he said no. I was like, “Really?” He said, “Yeah, it’s all supply. The technology will exist, but the supply chain won’t be there.” Here, you’re talking about Tesla making 1 million Optimuses per year within 5 years, but there are 300 million people and 150 million households. That means very few of your friends will have one in 5 years, just because the supply chain is so slow to catch up to the demand.

I’m supposed to get my 1X by the end of the year, at least by March, right? You heard me; you saw me shake hands with them.

That’s you, though.

I was talking to Steve Cousins about this, and we talked through some of this. There are all sorts of issues. One is battery life, which is still way too low for a bunch of these applications. The second is that if it falls over, it’s going to be so heavy that it will be very hard to pick up.

The 1X is like 70 pounds. And Lord help you if it falls on you.

So there are a lot of areas where I think this is going to take much longer. I would—not so much because of the supply chain. I think the liability issues, and constraining the function and the actual action of what it does, are going to take much longer to solve the insurance and legal issues.

Are you an optimist on this, Reid, or a pessimist? A robot pessimist?

An optimist, did you say, on this? Look, I was just kind of bemused. Ultimately, long term, I think obviously it’s there. I think the short term—I don’t know if Tesla has ever hit a target.

Touché. Alex, you were going to say something.

Yeah, I would also focus on that 80% figure, which I think is such a striking number. If you think about it from a market-analysis perspective, the size of automotive is probably like $4 or $5 trillion per year worldwide, whereas if you think about labor and the services market, depending on which estimates you believe, manual labor is like 2/3 of the services economy. Call that like $20 trillion.

So, in some sense, this 80% of Tesla’s value is really a bet that Tesla achieves parity with the services market. It’s a general sort of universal-intelligence-powered services company, and I think that’s probably where the market overall ends up.

Well, you remember, he’s got to hit $8 trillion to get his trillion-dollar pay package. So if anybody can do it, I think Elon can.

All right, let’s take a look at a different design. This is called the hidden robot. This is a generation of robots that don’t have, you know, 2 arms and 10 fingers.

Here. Here.

Yes. See, on the WTF episodes, we’ve asked, “Why do they have to have 2 arms and 2 legs and a head?” Well, okay. This is what they look like otherwise. Do you want one of these? Let me know. I think this is awesome.

I saw one where they were using it to map out the floors of a construction site and plan exactly where the pillars and so on would go. I think this is huge. I think the industrial use for robots is so much greater than the home use for a while to come that people are underestimating it. I think the efficiency gains from that are going to be huge. Obviously, the form factor should not be humanoid. I mean, my beef—you’ve heard me before—but at least give me a third arm if I’m a humanoid robot. At least—

Or dildo.

The use here: wind turbines, nuclear plants, subsea pipelines, railways, tunnels, power lines. The old adage is, if it’s dull, dangerous, or dirty, use a robot to do it. The estimated market in this particular tweet is that this sort of marketplace for robot maintenance and inspection is $6.7 billion today, growing at about 13% per year and expected to reach $12.5 billion by 2030.

I think that’s a radical underestimation. Just for example, if you look at the Mekong Delta in Indonesia, Vietnam, or whatever, it’s so polluted. If you had underwater robots cleaning it up, it would completely change the game and make things massively better. It’s like the smallest titch of an application.

I just want robots cleaning the side of the 405 out here in LA.

Okay.

Right. Or the beaches, too.

Yeah, sure. Okay, I found this one super interesting. A surgical robot performs a gallbladder procedure autonomously.

This is different from the da Vinci robot, which is basically an extension of a human operating in a theater. You guys all know what I've said about this. If you need to get surgery and you want to interview a surgeon, there's one question you ask them: How many times have you done this surgery this morning?

The success of a surgeon is a function of how many different cases they've seen and the eye-muscle memory of doing these. Ultimately, I do believe the best surgeons in the world will be robots. They'll see in infrared and ultraviolet. They wouldn't have a fight with their girlfriend or boyfriend, and they wouldn't have been drinking or consumed too much caffeine.

This comes out of Johns Hopkins, and they built a surgical robot without any human control. It achieved 100% accuracy in this gallbladder removal. It's different from da Vinci, and da Vinci came out of a DARPA project, right, to help with surgery in the field. I think this is huge. I don't think we'll have to wait more than 3 to 5 years. This is a sensor, actuator, and machine-learning problem. Alex, what do you think about it?

Yeah. Notably, Peter, I read the paper—a very exciting paper from the Hopkins team. This was a model that was trained by imitation learning, so it was trained by watching videos of human surgeons perform surgeries.

That immediately rhymes in my mind with the early DeepMind results, like AlphaGo, that were trained in part by watching expert human games. I think we're going to—and I would expect, if history does rhyme—enter an era when, using digital twins, and maybe this may or may not be aligned with what Reid is thinking for curing cancer with Manas AI, we're going to transition from imitation-learning-based medicine and surgery to reinforcement-learning-based medicine and surgery.

The moment we have a high-fidelity digital twin of the human body—and, of course, turtles all the way down, with virtual cell models as well—why train by copying humans when you could do reinforcement learning and achieve potentially super-duper-human-level performance? I think, again, we're in the early innings, but I think it's inevitable we see a surgery-zero-human-use version of this sometime soon. Reid.

Look, I think the surgery part of this stuff—I agree with you, in terms of line of sight. I think we're already there. It's a little bit like, for example, today, if you said, “Would you pick an AI or your average radiologist to read your X-ray film?” You'd pick the AI.

Today, hands down—11 out of 10 times.

And I think we're heading to that with robotics now. I think human biology is actually quite complicated, and the ability to do a full simulation is, in some ways, off. But this kind of robotic thing—oh my gosh, hit the accelerator.

Yeah. Love it. Love it.

I think of it as autopilot for flying a plane. Today, planes will fly themselves 99% of the time; the pilot is only there in case of an emergency. I think we'll see the same thing.

All right. So, Zoox—I remember Zoox. I went and visited them and met the team there. They were acquired by Amazon back in 2020.

This is a self-driving pod, right? This is you and your best buddies facing each other inside there. They're finally launching. Good for them. They have massive scale—well, not really: 50 vehicles in their fleet. They'll start in Las Vegas, then San Francisco, Miami, and LA. It's free for the first few months, but then they're going to go to pricing similar to Uber and Lyft. Any thoughts, gentlemen?

Have any of you guys taken the self-flying drone in Dubai? You push the button, and it picks you up.

Not yet.

Dying to. I just don't want to be first.

Just don't do it to die. Let's put it that way.

They're in production.

Yeah.

Peter, you're usually really adventurous with that stuff.

I would do it in a heartbeat. I saw it first at the Consumer Electronics Show. Martine Rothblatt bought about 100 of those vehicles early on for organ delivery before she got involved in her own vertical-takeoff flying car, Beta.

All right. Let's wrap here at the end of the program. I remain continuously like a kid in a candy store at the speed at which this is moving. I love the articles you send over, Alex, and the conversations that we have. Reid, what are you most excited to see in the next year?

I'd say the next year will be part of the reason why I think the focus on coding and the acceleration of coding is important. It accelerates everything: it accelerates individuals, as per the copilots that I was talking about before, but it also accelerates the discovery and the computation—the algorithms that Dave was talking about.

I actually think we will see massive coding acceleration, and that will be a precursor to many other accelerants. But is there one science-fiction, one Star Trek part of the equation that you're looking forward to?

Star Trek.

You know, we tend to overpredict the 2 years and underpredict the 10 years. So, what would be the science-fiction thing? I'm very hopeful about the iPod 3s[?].

Maybe a tricorder.

Yep. I think we've reported on some early version of the tricorder. I did a $10 million Qualcomm Tricorder XPRIZE 10 years ago. It's time to do it again. I think the tech is there for sure. Dave, what about you? What are you excited about?

Passenger drones would be my personal favorite. I spent 8 days out of 9 in India, traveling between airports, and thought, “What the hell kind of waste-of-time crap is that in today's world?”

The technology is there. It's now just an implementation and infrastructure problem. We saw a flying car on the campus of Stanford 2 days ago. Very, very impressive—very impressive design, and it's very workable. He thought he could get the cost down over time to about $40,000 a car.

Yeah. I can't wait. What I'm most excited about by far is a version 2 of “Reid Interviews Reid.” I thought that was one of the most brilliantly conceived pieces of media. Anybody who hasn't seen it, dig it up.

You could do it so much better today, actually, because when you did it, I assume you coded that up yourself, but that was pretty hard to do at the time.

Now you could do something incredible. Peter did an onstage interview of Socrates and Aristotle that turned into a big love fest. If you don't prompt it right, everybody just loves everybody.

I did an interview of myself last year at the Abundance Summit—of a 150-year-old version of myself—which was fun. I asked it about the future, and it had some great answers. I loved it.

Everybody watching this podcast, please—Reid, do it again. Do an update, and that'll put some pressure on you.

And Alex, let's not forget you, buddy. What do you imagine over the next year that would really hit your childhood ambitions?

Oh gosh. I think I'm getting rather difficult to ontologically shock at this point. But I will say, pulling all of Star Trek—not just some of Star Trek—to the left, I think that's a worthy ambition.

We've got the holodeck.

We've almost got the replicators, if you squint at food printers. We're missing warp drive, and we're missing a whole bunch of other aspects of Star Trek. Wouldn't it be lovely if we were able to pull those to the left as well?

I've got one. I was chatting with Steve Jurvetson at the Stanford conference, and he reiterated that crazy anecdote that, once we have a quantum computer, it'll be definitive proof of a multiverse. That really needs alcohol to get into.

All right. We'll do that one in one of our next sessions.

See you guys, my moonshot mates. I appreciate you, Reid. Thank you, buddy. It's been a wonderful friendship. I'm grateful for you.

Massively fun.