Reid Hoffman谈AI、意识与劳动的未来

This is actually one of the things I think people don't realize about Silicon Valley. You start with what's the amazing thing that you can suddenly create. With lots of these companies, you ask, “What's your business model?” and they go, “I don't know.” You're like, “Yeah, we're going to try to work it out, but I can create something amazing here.”

That's actually one of the fundamental—call it the religion—of Silicon Valley, and the knowledge of Silicon Valley that I so much love, admire, and embody.

Reid, welcome.

It's great to be here.

So, Reid, you're one of the most successful Web 2 investors of that era: Facebook, LinkedIn, which you co-created, Airbnb, and many, many others. You had several frameworks that helped you do that, one of which is the seven deadly sins, which we talk about often and love. As you're thinking about AI investing, what's your framework or worldview?

Obviously, we're all looking through a glass darkly, through a fog with strobe lights that make it hard to understand what's going on. We're all navigating this new universe, so I don't know if I have a crisper framework. The seven deadly sins still work because that's a question of psychological infrastructure across all 8 billion-plus human beings.

I'd say there are a couple of things. First, there is going to be a set of things that are the obvious line of sight: a bunch of stuff with chatbots, a bunch of stuff around productivity, coding assistants, and so on. That's still worth investing in, but obvious line of sight means it's obvious to everybody, and so doing a differential investment is harder.

The second area is: What does this mean? Too often, people say that in an area of disruption, everything changes, as opposed to significant things changing. You were mentioning Web 2 and LinkedIn, and obviously part of this is a platform change. You go, “Okay, are there now new LinkedIns that are possible because of AI or something like that?” Given my own heritage, I would love LinkedIn to be that, but I'm always pro-innovation and entrepreneurship, and for the best possible thing for humanity.

What are the more traditional things that haven't changed? Network effects, enterprise integration, and other kinds of things. The new platform upsets the apple cart, but you're still going to be putting that apple cart back together in some way. What is that?

The third, which is probably where I've been putting most of my time, is what I think of as Silicon Valley blind spots. Silicon Valley is one of the most amazing places in the world. There's a network of intense co-opetition, learning, invention, and building new things, which is just great. But we also have our canons and our blind spots.

A classic one for us tends to be: Everything should be done in computer science, everything should be done in software, and everything should be done in bits. That's the most relevant thing because, by the way, it's a great area to invest in. But I thought, “What are the areas where the AI revolution will be magical but won't be within the Silicon Valley blind spots?”

That's probably where I've been putting the majority of my co-founding time, invention time, investment time, and so on. A blind spot on something that's very, very big is precisely the kind of thing where you go, “Okay, you have a long runway to create something that could be another iconic company.”

Yeah. Let's go deeper on that, because we were also talking just before this about how people focus so much on the productivity and workflow sides, but they're missing other elements. Say more about the other things that you find more interesting now.

One of the things I told my partners back at Greylock in 2015, about 10 years ago, was, “Look, there's going to be a bunch of things around productivity and AI. I'll help. You have companies you want me to work with that you're doing? Great, that's awesome. Enterprise productivity, things that Greylock tends to specialize in.”

But I said, “Actually, in terms of getting at the blind spots, there are also going to be things like Madison AI, which is: How do we create a drug-discovery factory that works at the speed of software?”

Obviously, there are regulatory and biological bits, and so it won't be purely the speed of software. But how do we do this? They said, “What do you know about biology?” The answer is zero. Well, maybe not quite zero. I've been on the board of Biohub for 10 years, I'm on the board of Arc, and I've been thinking about the intersection of the world of atoms and the world of bits. You have biological bits, which are halfway between atoms and bits in various ways. I've been thinking about this a lot and about what the possibilities are.

Not so much with a specific company focus as with a “What are things that elevate human life?” focus. That's part of the reason for Biohub and part of the reason for Arc. But then I thought, “Well, wait a minute. Actually, now with AI, you have the acceleration.”

Roughly 10 years ago, I was asked to give a talk to the Stanford Long-Term Planning Commission. What I told them was that they should basically divert and put all of their energy into AI tools for every single discipline. This was well before ChatGPT and all the rest.

The metaphor I used was a search metaphor: Imagine if you had a custom search-productivity tool in every single discipline. Back then, I could imagine it. I could build one for every discipline other than theoretical math. Today, you might even be able to do theoretical math and theoretical physics.

Exactly.

And so, do that. Transform knowledge generation, knowledge communication, and knowledge analysis.

That same thing had me thinking: The biological system is still too complex to simulate. We've got all these amazing things with LLMs, but the classic Silicon Valley blind spot is, “Oh, we'll just put it all in simulation and drugs will fall out.” That simulation is difficult.

Part of the insight that you begin to see from the work with AlphaGo and AlphaZero is that people used to think physical matter was going to take quantum computing. Quantum computing could do really amazing things, but actually, simply doing prediction and getting that prediction right—and, by the way, it doesn't have to be right 100% of the time; it has to be right 1% of the time, because you can validate that the other 99% weren't right, and then find that one thing—is possible.

Literally, it's not a needle in a haystack; it's a needle in a solar system. But you could possibly do that. That's part of what led me to think, “Okay, Silicon Valley will classically go, ‘We'll put it all in simulation, and that will solve it.’” Nope, that's not going to work. Or, “Oh no, we're going to have a superintelligent drug researcher, and that will be 2 years down the line.” I actually think, look, maybe someday, but not soon.

Part of it is also what a lot of people don't realize. Actually, if I'm not going too long, I'll go to the other example that I gave, because you'll love this. This will echo some of our conversations from 10 or 15 years ago.

I'm prepping for a debate this Sunday on whether or not AIs will replace all doctors in a small number of years. The pro case is very easy: We have massively increasing capabilities. If you look at ChatGPT today, you'd go, “For example, advice to everyone who's listening to this: If you're not using ChatGPT or an equivalent as a second opinion, you're out of your mind. You're ignorant. If you get a serious result, check it as a second opinion. If it diverges, then go get a third.”

The diagnostic capabilities are much better knowledge stores than any human being on the planet. So you go, “Well, if a doctor is just a knowledge store, yeah, that's going away.”

However, the question is: What are the things that really do mean “doctor”? It's not like, “Oh, someone will hold your hand and say, ‘Oh, it's okay.’” I actually think there will be a position for a doctor 10 years from now, 20 years from now. It won't be as the knowledge store; it will be as an expert user of the knowledge store.

But it's not going to be, “Oh, because I went to med school for 10 years and memorized things intensely, that's why I'm a doctor.” That's all going away. Great, that part is going away. But there's a lot of other parts to being a doctor.

I went to ChatGPT Pro and used deep research. I went to Claude 4.5 deep research, Gemini Ultra, and Copilot deep research. In all of these, I was doing everything I knew about prompting to try to give me the best possible arguments for my position. I thought, “I'm about to debate AI. Of course I should be using AI to debate.”

The answers were B-minus or B, despite absolutely top-tier prompting. Maybe there are better prompters in the world, but I've been doing this since I got access to GPT-4, 6 months before the public did. So I've got some experience in the whole prompting thing; it's not like I'm an amateur prompter.

And so I looked at this and thought, “Oh, this is very interesting,” and a telling illustration of where current LLMs are limited in their reasoning capabilities. What it did was basically 10 to 15 minutes of 32-GPU compute clusters doing inference, bringing out all this amazing work relative to work that an analyst would have produced in 3 days. It was produced in 10 minutes. Of course, I set it up all in parallel, with different browser tabs going into the different systems, and then ran the comparisons across them.

But its flaw was that it was giving me a consensus opinion about how articles in good magazines, good things are arguing for that position today. And all of that was weak because it was kind of like, “Oh, you need to have humans cross-check the diagnosis.” That was a common theme across this. And I'm like, well, by the way, very clearly we know as technologists that we're going to have AIs cross-checking the diagnosis.

We're going to have AIs cross-checking the AIs that cross-check the diagnosis. And sure, there'll be humans around here somewhere, but that's not going to be the central place to say, “In 20 years, doctors are going to be cross-checking the diagnosis.” Because what doctors should be learning very quickly is that if you believe something different from the consensus opinion that an AI gives you, you'd better have a very good reason, and you're going to go do some investigation. It doesn't mean the AI is always right.

That's actually part of what we're going to need in all of our professions: more sideways thinking, more lateral thinking. “Okay, this is a good consensus opinion. Now, what if it's not consensus opinion?” That's what doctors need to be doing. That's what lawyers will need to be doing. That's what coders will need to be doing. That's what it is. And LLMs are still pretty structurally limited there.

What's funny is my favorite saying is by Richard Feynman: “Science is the belief in the ignorance of experts.” There are so many professions where credentialism is the expertise. It's like, “If this, then that.” It's like, “I have an MD, therefore I know. I have a JD, therefore I know.” And that's why coding is actually a little bit ahead of it, because it's like, “I don't care where you got your degree.” This is kind of ahead of the rest of society.

Now, it's funny: Milton Friedman was once asked, because he was a famous libertarian, “Don't you think brain surgeons should be credentialed?” And it's like, “Yeah, the market will figure that out.” It seems kind of crazy, right? But that's how we now do coding when you're in the world of bits.

But it feels like a lot of the reasons why you have this very not-very-advanced thinking is because so much of it is built upon layers of credentialism. That's a very good heuristic, and historically it has been. If you have a doctor who graduated at the top of their class from Harvard Medical School, it's probably a good doctor. And, by the way, you politically wanted that.

Yes. 30 years ago. Right, right? It's like, “No, no, I need someone who has the knowledge base. Do you have it? Great.” But now we have a knowledge base. Yeah, I totally agree. That was the reason I was saying you would love this, because it echoes our expert discussion.

I thought you were going to get into bits versus atoms, where it's kind of interesting right now. All this high-value work, like a Goldman Sachs sell-side analyst—that's deep research. Whereas folding my laundry, that's $100,000 of capex. So it doesn't work as well as somebody you could pay $10 an hour to. The atom stuff is so hard to actually disrupt.

We're going to get there eventually, but that's where Silicon Valley certainly has a blind spot. It's a capex versus opex thing, or bits versus atoms. The atoms are another part, but that's also the reason why bio—because bios are the bitty atoms.

And what's the best explanation for why it's so hard to figure out folding laundry, but so easy to figure out deep research? Well, it's actually not that hard to figure out why it's taken us much longer and been much more expensive, because it would have been hard to foresee that in advance. I remember I talked to Ilya about this a few years ago. Why is it that if you read an Asimov novel where it talked about how people would cook for you and fold your laundry, none of these things have happened?

It's like, well, you just never had a brain that was smart enough. This is part of the problem. You could—I mean, yes, you have things like, “How do you actually pick up this water bottle?” And it turns out your hands are very, very well evolved. Why are humans more advanced than every other species? There are 2 reasons.

Number 1 is we have opposable thumbs. And number 2 is we've come up with a language system that we could pass down from generation to generation, which is writing. Dolphins are very smart. There was actually a whole theory that it wasn't just brain size; it was brain-to-body size. So humans were the highest. Nope, not true.

Now that we've actually measured every single animal, there are a lot of animals that have a greater brain-to-body ratio. An elephant or a dolphin—I forgot the numbers—but there are a bunch that are actually more advanced than humans, but they don't have opposable thumbs. And because of that, they never developed writing, so they can't actually iterate from generation to generation. Humans did. And then, of course, the human condition was like this, then the Industrial Revolution, then it went like that, and now it's continued like this.

By the way, this is the reason why, in the last 4 or 5 years, one of the things I realized is that the classic classification of human beings is Homo sapiens. I actually think we're Homo technae. It's that iteration through technology. Yes, exactly. Whatever version—typing, you know—but we iterate through technology.

That's the actual thing that goes to future generations, builds on science, all the rest of it. And that's what I think is really key. A couple of other explanations could be that we have more training data on white-collar work than on picking things up. Or, as some people make this evolutionary argument, we've been using our opposable thumbs for way longer than we've been, say, reading or writing.

Yeah, it's the lizard brain. Most of your brain is not the neocortex, and that's the part responsible for drawing and painting and everything else, which is actually very, very hard. You can't find a dolphin that can draw or paint, and that's probably because they don't have opposable thumbs. But maybe that part of the brain hasn't developed. You have billions of years of evolution for these somewhat autonomous responses—fight or flight—that have been around for a long, long time, well before drawing and painting.

But I think the main issue is that you have battery chemistry problems. It turns out a lithium-ion battery is pretty cool, but its energy density is terrible relative to ATP in cells. You have all of these reasons why robotics don't work, but first and foremost, the brain was never very good. You had robotics like FANUC, which makes assembly-line robots. Those work really well, but they're very deterministic, or highly deterministic.

But once you go into multiple degrees of freedom, you have to get so many things to work, and the capex is like, “I need $100,000 to have a robot fold my laundry.” We have so many extra people who will do that work, so the economics never made sense. But this is why Japan is a leader in robotics: because they can't hire anybody. So therefore, I might as well build robots.

True story: I went bowling in Japan, and they had a robot—a vending-machine robot—that would give you your bowling shoes. And then it would clean the bowling shoes, right? You would never build that here, because you'd hire some guy from the local high school and he'd go do that. Yeah, and much cheaper and actually more effective.

But it's the capex line and the opex line. When they cross, then it's like, “Ooh, I should build robots.” So that's the other thing you probably need. But if the cost goes down, then of course it goes in favor of capex versus opex.

I think there are a couple of things to go into more deeply on the robot side. One is the density—the bits-to-value ratio. In language, when we encapsulated all these things, even into romance novels, there's a high bits-to-value ratio. Whereas when you're going into the real world, there's a lot of, “How do we abstract from all those bits, and how do you abstract them?”

There's another part of it, which is kind of common-sense awareness. This is one of the things that, when I look at GPT-2, GPT-3, GPT-4, GPT-5, it's a progression of savants. The savants are amazing, but when they make mistakes—Microsoft has had this running for years now, agents talking to each other, for, like, “Let's go for a year and do that and see what happens.” So often they get into, like, “Oh, thank you.”

No, thank you. No, thank you. One month later, thank you. No, thank you. Human beings are like, “Stop.” That’s a simple way of putting the context-awareness thing: “No, no, no, no, let’s stay very context-aware.” Even as magical as the progression has been—much, much better data, much, much better reasoning, much, much better personalization, and so on—context awareness is only a proxy for that.

I want to go deeper on your question about doctors reading, because we just released one of your talks around software eating labor. I’m curious what sort of frameworks you have for thinking about what spaces are going to have more of this copilot model versus what spaces are going to be more about replacing the work entirely.

I wish I could use an LLM to predict the future, but I’m going to get a B-minus, apparently. So maybe I’ll answer when I get a B-plus.

I think a lot of it is natural. There’s this anthropomorphic version, which is, “Okay, well, I trust the doctor. Everybody trusts the doctor.” The heuristic is, where did you go to medical school? Apparently, two-thirds of doctors now use OpenEvidence, which is like ChatGPT, but it ingested the New England Journal of Medicine and has a license to that. Daniel Nadler. Good guy.

That seems like there’s no reason not to do that. My “seven deadly sins” version—I’ll simplify it—is that everybody wants to be lazier and richer. So if this is a way that I can get more patients and do less work, of course people are going to use it. There’s no reason not to.

But does it replace that particular thing? Most of the software-eats-labor thing doesn’t actually eat labor right now. The thing that’s working the best is not, “Hey, I have a product where everybody’s going to lose their job.” Nobody’s going to buy that product. It’s very, very hard to get that distributed, as opposed to, “I will give you this magic product that allows you to be lazier.”

Obviously, it’s not framed this way, because “lazy and rich” sounds kind of not great. But “I’m going to let you work fewer hours and make more money” is a very killer combo. If you have a product like that, and it’s delivered by somebody who already has that heuristic of expertise, these are just going to go one after another and get adopted, adopted, adopted.

Eventually, you’re going to have cases like the one that you mentioned where, if you don’t use ChatGPT when you get a medical diagnosis, you’re insane. But that is not fully diffused across the population. It’s barely diffused.

No, I know. You were saying “not fully.” Part of the reason everyone will start doing it—

Yes, 100%.

What’s that? It’s the fastest-growing product of all time, again. It’s barely—

That’s why I’m convinced that AI is massively underhyped. In Silicon Valley, you might not make that claim. Maybe it’s overhyped, maybe it’s evaluation, whatever. But I think once I meet somebody in the real world and show them this stuff, they have no idea.

Part of it is that they see the IBM Watson commercials and think, “Oh, that’s AI.” No, that’s not AI. Or they see the fake AI. They’ve seen ChatGPT 2 years ago, it didn’t solve a problem, and—

Yes, and that’s bad. But I think it’s going to diffuse largely around this lazy-and-rich concept, and that’s where a lot of these things have taken off. I see it less at the very, very big companies because you have a principal-agent problem at the very big companies.

“My company made money or saved money. I’m a director of XYZ. All I know is that I want to leave earlier and get promoted.” How does that actually help me? It helps the ethereal being of the corporation. Whereas at a smaller business, or a sole proprietor, or an individual doctor—if I run a dermatology clinic and somehow I can have 5 times as many patients, or I’m a plaintiff’s attorney and I can have 5 times as many settlements—of course I’m going to use that, because I get to be lazier and richer.

Yeah, 100%. That seems like a good model. By the way, you’re reminding me of another quote from Ethan Mollick that I use often. He’s great: “The worst AI you’re ever going to use is the AI you’re using today,” to remind you to use it tomorrow.

A lot of the skeptics are exactly this: “I tried it 2 months ago, and it didn’t solve this problem. Therefore, it’s bad.” You’re judging it on the present. You have to extrapolate. You don’t want to get too extrapolatory on, “LLMs have this.” You actually have to recognize that the 2 types of people who are underhyping AI are people who know nothing and people who know everything.

It’s really interesting. It’s like the midwit meme where the people in this part of the distribution are correct. Normally, the meme is the opposite: these people are smart even though they’re dumb, and these people are smart even though they’re smart. Everybody here—this part of the curve—is actually correct, because they’re the ones using it to get richer and be lazier.

The other thing I tell people is, if you haven’t found a use of AI that helps you with something serious today—not just “Write a sonnet for your kid’s birthday” or “I’ve got these ingredients in my fridge; what should I make?” Do those, too. But if you haven’t found something for work, something serious about what you’re doing, you’re not trying hard enough.

Yeah, yeah. Isn’t it that it works for everything? For example, I still think if I put in, “How should Reid Hoffman make money investing in AI?” and go try that again, I suspect I’ll still get what I think is the bozo business-professor answer versus the actual game—the name of the game. But everyone should be trying.

For example, when we get decks, we put them in and say, “Give me a due-diligence plan.” If not everybody here is doing that, that’s a mistake, because in 5 minutes you get one and go, “Oh, no, not 2, not 5. But 3 is good,” and it would have taken me a day to get to about 3.

In terms of extrapolation, obviously the last few years have had incredible growth. You were involved with OpenAI since the beginning. When we look at the next few years, it raises the question of whether scaling laws will hold, what the limitations are, and how far we can get with LLMs. Do we need another breakthrough of a different kind? What is your view on some of these questions?

We all swim in this universe of extrapolating the future. One of the things that’s great about Silicon Valley is that you get such things as theories of the singularity, theories of superintelligence, and theories of exponential growth getting to superintelligence soon.

What I find is usually the mistake in that is not the fact of extrapolating the future. That’s smart, people need to do that, and far too few people do. I think I remember liking your post and helping promote it, if I recall. But it’s the notion of, “Well, what curve is that?” If it’s a savant curve, that’s different than “Oh my gosh, it’s an apotheosis, and now it’s God.”

It’ll be an even more amazing savant than we have. But, by the way, if it’s only savants, there’s always room for us. There’s always room for the generalist and the cross-checker and the context awareness and all the rest of it. Maybe it’ll cross over a threshold, or maybe it won’t. I think there are a bunch of different questions here.

That extrapolation too often goes, “Well, it’s exponential, so in 2½ years, magic.” You’re like, “Well, look, it is magic, but it’s not all magic.” That’s the way he’s doing it.

My own personal belief is that the critics of LLMs make a mistake. We can go through all the different criticisms: “Oh, it doesn’t have knowledge representation. It screws up on prime numbers,” and so on. We’ve all heard, “How many Rs are in strawberry?” They’re like, “Oh, see, it’s broken.” You’re missing the magic.

Yes, maybe there are some structural things that, over time—even in 3 to 5 years—will continue to be a difficult problem for LLMs.

But AI is not just the one LLM to rule them all. It's a combination of models. We already have combinations of models. We use diffusion models for various image and video tasks. They wouldn't work without also having LLMs in order to have the ontology to say, “Create me an Erik Torenberg as a Star Trek captain, going out to explore the universe, meeting and making first contact with the Vulcans, and so forth.” With our phones, we could do that now. It would be there courtesy of OpenAI. And, you know, Veo, because Google's model is also very good. But it needs LLMs for that.

The thing that people aren't tracking is that it's going to be LLMs and diffusion models and, I think, other things, with a fabric across them. One of the interesting questions is: Is the fabric fundamentally LLMs? Is the fabric other things? I think that's TBD. And the degree to which it gets to intelligence is an interesting question.

One of the things I think is—I talk to all the critics intensely, not because I necessarily agree with the criticism, but because I'm trying to get to the kernel of insight.

Yeah. One of the things that I loved about a set of recent conversations with Stuart Russell was his saying, “Hey, if we could actually get the fabric of these models to be more predictable, that would greatly allay the fears of what happens if something goes amok.” Well, okay, let's try to do that. I don't think the whole verification of outputs, like logical verification—we can't even do verification of coding. Verification strikes me as very hard. He's a brilliant man; maybe Google will figure it out.

On the other hand, this is a good goal: Can we make that more programmable and reliable? I think that is a good goal that very smart people should be working on. And, by the way, smart AIs—well, that's some of the math side.

If you think about the foundation of the world, philosophy is the basis of everything. Actually, math comes from philosophy. It's called the Cartesian plane after Descartes. You're philosophy-based. You know this, right? So you have philosophy, math, and physics. Why did Newton build calculus? To understand the real world. So, math and physics. Physics gets you chemistry. Chemistry gets you biology, and then biology gets you psychology. That's kind of the stack.

If you solve math, that's actually quite interesting, because there's a professor at Rutgers, Kontorovich, who's written about this a lot. I find this part fascinating, just as a former mathematician, because there are some very, very hard problems. There's a rumor that the Navier–Stokes equations are going to be solved by DeepMind, which would be huge. That's one of the Clay Math problems.

Yeah. But the Riemann hypothesis is not like that. There's no eval.

Yes. Right? Once you get to proofs, it's very, very hard. If you look at the progression of AI, there's the AIME, the American Invitational Mathematics Examination, where the answers are all just integers. The answer is from 0 to 999, and you can keep trying different things. You either get the right answer or you don't, and it's very, very easy to do that. Whereas once you get to proofs, it's very, very hard.

If you solve that, I mean, is that AGI? No, because the goalposts keep changing for AGI.

Yes. But math is just so interesting.

AGI is the AI we haven't invented.

Exactly. That's exactly it. It's the corollary to, if the worst AI you're going to try is today, then AGI is what you're going to have tomorrow, right? It's the same kind of thing.

Math is a very, very interesting one as well, because, again, you have these things. It's not like solving high school math. This is like, if you're able to actually logically construct a proof for something and then validate it. There's a whole programming language called Lean, which is for that. That stuff is also fascinating.

There are so many different vectors of attack, which is the other way of thinking about it. It's fascinating. So, as you just mentioned, Alex, you're a philosophy major, but you're also very interested and deep in neuroscience. Some people say that we'll never create AI with its own sort of consciousness because we don't even understand our own consciousness. We don't understand how our own brain works. Then there's the broader question: Will AI have its own goals? Will AI have its own agency?

Well, consciousness is its own hairball, which I will say a few things about. I think agency and goals are almost certain. There is a question—I think this is one of the areas where we want to have some clarity and control. That was a little bit like the question of what kind of compute fabric holds it together, because you can't get complex problem-solving without it being able to set its own mini subgoals and other kinds of things. So you get goal-setting, behavior, and inference from it.

That's where you get the classic idea: You tell it to maximize paper clips, and it tries to convert the entire planet into paper clips. There's one thing that's definitely old-computer, where there was just no context awareness—something I even worry about with modern AI systems. On the other hand, if you're actually creating an intelligence, it doesn't say, “Let me just go try to convert everything into paper clips.” It's actually not that simple in terms of how it plays.

Consciousness is an interesting question because you have some very smart people, like Roger Penrose, whom I actually interviewed way back when on The Emperor's New Mind, speaking of mathematicians. He says there's something about our form of intelligence—our form of computational intelligence—that's quantum-based, that has to do with how our physics work, and that has to do with things like true tubulars and so forth.

By the way, it's not impossible. That's a coherent theory from a very smart mathematician—one of the world's smartest. It's kind of in the category of there being other people as smart, but no one smarter in that vector. So that's possible.

I don't think you need consciousness for goal-setting or reasoning. I'm not even sure you need consciousness for certain forms of self-awareness. There may be some forms of self-awareness for which consciousness is necessary. It's a tricky thing. Philosophers have been trying to address this—not very well—for as long as we've had records of philosophy. And philosophers agree. I wouldn't think I was throwing them under the bus with this. They're like, “Yeah, this is a hard problem,” because it ties to agency and free will and a bunch of other things. I think the right thing to do is keep an open mind.

Part of keeping an open mind, I think, is that Mustafa Suleyman wrote a very good piece in the last month or 2 on semiconsciousness. We make too many mistakes with the Turing test, which was a piece of brilliance: It talks to us, so therefore it's fully intelligent, and all the rest. Similarly, you had that kind of nutty event with that Google engineer who asked an earlier model whether it was conscious, and it said yes, so therefore it is.

QED. You're like, “No, no, no, no.” You have to not be misled by that kind of thing.

For example, what I actually think is that most people obsess about the wrong things when it comes to AI. They obsess about the climate change stuff because, actually, if you apply intelligence at the scale and availability of electricity, you're going to help climate change. You're going to solve grids and appliances and a bunch of other stuff. This will be net super-positive. You already see elements of it. Google applied its algorithms to its own data centers, which are some of the best-tuned systems in the world, and achieved 40% energy savings. That's just from applying it.

One of the areas I think is this question of what we want children growing up with AIs to be like. What is their epistemology? What are their learning curves? What are the things that play into this? That's something we want to be very intentional about in terms of how we're doing it. If you want to ask a good question that we should be trying to get good answers to, and that you could contribute good answers to, that's a good one.

The most cogent argument that I've heard against free will is just that we are biochemical machines. So, if you want to test somebody's free will, get them very hungry or very angry—all of these things where there's just a hormone, like norepinephrine, that makes you act a particular way. It's like an override.

Yes. So you have this free will thing, but then you just insert a certain chemical and, boom, it changes. Are you saying you're not a Cartesian? You don't have a little pineal gland that connects the 2 hemispheres?

No, I don't know. But it's true. Hunger is, “Yeah, I'm hungry.” That's a thing. And what is the point? If you're developing superintelligence, do you actually want to have this kind of silly override? The reason why people who are perfectly normal sometimes go to jail is that they get very angry.

They do things that are out of character. But it’s actually not out of character if you think about this free-will override of just chemicals going through your bloodstream, which is kind of crazy to think about.

Look, since we’re on a geeky, nerdy podcast, I’m going to say 2 geeky, nerdy things. 1, the classic one is people say, yes, we are biochemical machines, but let’s not be overly simplistic about what a biochemical machine is. That’s like Penrose, quantum computing, et cetera. And you get to this weird stuff in quantum, which is, well, it’s a probabilistic, dual-superpositional form until it’s measured.

Why is there magic in measurement? And is that magic in measurement something that’s conscious? Blah, blah, blah. So there’s a bunch of stuff there. The other thing that I think is interesting, that we’re seeing as a resurgence in philosophy a little bit, is idealism. We would have thought, as physical materialists, that we’d say, “No, no, idealists were disproven. They’re gone.”

But actually, we’re beginning to say, “No, actually, in fact, what exists is thinking, and all of the physical things around us come from that thinking.” Obviously, we see versions of this because I find myself entertained frequently here in Silicon Valley by people saying, “We’re living in a simulation. I know it, you know it.”

And you’re like, “Well, your simulation theory is very much like Christian intelligent-design theory. It’s, ‘I have things that I can’t explain, so therefore, creator.’” “No, therefore, simulation.” “No, therefore, creator of simulation.” You’re like, “No, no, no, but I…” So clearly, I’m not an idealist. But that’s why I see some resurgence of idealism happening.

I suspect—geeky—I suspect we’ll solve for AGI before we solve for various definitions of AGI, before we solve for the hard problems of consciousness.

Yes. I want to return to LinkedIn, where we began the conversation, because we were lucky to—or I was lucky to—work many years with you. We would get pitches every week about a LinkedIn disruptor. Over the last 20 years, right?

Yes. And nothing’s come even close.

Nope. And so, it’s fascinating. I’m curious why people underrated how hard it was. People have this about Twitter, too—other things that look simple, perhaps, but are actually very, very difficult to unseat and have a lot of staying power.

And it’s interesting: OpenAI said they’re coming out with a job service to, quote, “use AI to help find the perfect matches between what companies need and what workers can offer.” I’m curious to see how you think about LinkedIn’s durability.

Look, I obviously think LinkedIn is durable, but first and foremost, I look at this as humanity, society, industry. First and foremost is: What are the things that are good for humanity? Then, what’s good for society? Then, what’s good for industry? And, by the way, we do industry to be good for society and humanity. It’s not oppositional. It’s just how you’re making these decisions and what you’re thinking about.

I would be delighted if there were new, amazing things that helped people make productive work, find productive work, and do it. We’re going to have all this job transition coming from technological disruption with AI. It would be awesome. Of course, it would be extra awesome if it was LinkedIn bringing it, given my own personal craft of my hands and pride at what we built and all the rest.

Now, the thing with LinkedIn—and Alex was with me on a lot of this journey, as I sought his advice on various things—LinkedIn was one of those things where the turtle eventually grows into something huge. For many, many years, the general scuttlebutt in Silicon Valley was that LinkedIn was the butt-ugly, dull, boring, useless thing, et cetera. And it was going to be Friendster. Probably most people listening to this don’t know what Friendster is. Then MySpace. Maybe a few people have heard of that, right? And then, of course, we got Facebook and Meta, TikTok, and all the rest.

Part of the thing with LinkedIn is that it’s built a network that’s hard to build, right? Because it doesn’t have the same sizzle and pizzazz that photo sharing has. It doesn’t have the same sizzle and pizzazz that other networks have. 1 of the things that you were referencing was the seven deadly sins comment.

Back when I started doing that, in 2002—yes, I left my walker at the door—the thing that I used to say was, “Twitter was identity.” I actually mistook it; it’s wrath, right? And it so doesn't have the wrath component of it. And so, with LinkedIn, LinkedIn is greed. Great, because the seven deadly sins are a motivation that’s very common across a lot of human beings.

Rich and lazy.

Yes, exactly. You’re putting it in the punchy way, but simply being productive.

More value creation and accruing some of that value to yourself.

And so I think the reason why it’s been difficult to create a disruptor to LinkedIn is that it’s a very hard network to build. It’s actually not easy. And by staying really true to it, you end up getting a lot of people going, “Well, this is where I am for that.” And now I have a network of people who are here, and we are here together collaborating and doing stuff together. And that’s the thing that a new thing would have to be.

When I saw GPT-4 and knew that Microsoft had access to this, I called the LinkedIn people and said, “You guys have got to get in the room to see this,” right? Because you need to start thinking about what are the ways we help people more with that.

Because you start with, “What’s the amazing thing that you can suddenly create?” This is actually 1 of the things that I think people don’t realize about Silicon Valley, because the general discussion is, “Oh, you’re trying to make all this money through equity and all this revenue.” Of course, business people are trying to do that. But what they don’t realize is that you start with, “What’s the amazing thing that you can suddenly create?”

And part of it is that lots of these companies started with, “What’s your business model?” You go, “I don’t know.” You’re like, “Yeah, we’re going to try to work it out. But I can create something amazing here.” And that’s actually one of the fundamental places of what they call the religion of Silicon Valley, and the knowledge of Silicon Valley that I so much love and admire and embody.

That’s actually a question that I have. I’ll say 1 thing: It’s a huge compliment to LinkedIn—it’s antifragile. Facebook is, “Oh, nobody goes there anymore.” It’s like Yogi Berra: “It’s too crowded; nobody goes there anymore.” It’s like, “Oh, there are too many parents there.” And there’s always been a new one.

How did Snap start? All these other networks started because people didn’t want to hang out with their boomer parents. My kid won’t let me follow him on Instagram, right? So he doesn’t want to use Facebook. So LinkedIn has survived through all of that.

But you referenced something that I think is a very interesting point, which is that back in Web 2, it was: get lots of traffic, get amazing retention, smile curve, and then you will figure out monetization. And that isn’t happening right now.

Yes, it happened with ChatGPT, but it’s $20 a month.

Right? The monetization was kind of built in—very, very clear subscription versus becoming giant and building a giant network. Do you think there will be new ones of those with AI?

Yes. And there will be new kinds of freemium. It’s part of our tool chest now. Part of the reason why it’s trickier, especially when you’re doing OpenAI, is because the COGS are changing a little often.

Yes, right?

No, no, but you just can’t. This is 1 of the reasons why, at PayPal, we had to change to a paid model because, as you know, you were close to us there. We had exponentiating volume, which means an exponentiating cost curve, which means that despite having raised hundreds of millions of dollars, we could literally count the—we could point to the hour that we’d go out of business. Right? Because no, you can’t have an exponentiating cost curve.

So I think that’s 1 of the reasons why some of it has been different in AI, because you can’t have an exponentiating cost curve without at least a following revenue curve.

Right. But it’s almost no fun. It’s like Pinterest. It’s like, “How are they going to make money?” Now it’s a big public company. There were a lot of these during that era, and now it’s like they’re burning lots of money, they’re raising lots of money, but the subscription revenue is baked in from day 0.

Yeah. And that’s the fundamental—

They have to because of the cost curve.

They have to, exactly.

Yeah. So I’m waiting for 1 of these net-new companies that appeals to probably 1 of the seven deadly sins, the new counterpart.

Yeah. Well, I’d be happy to work on one.

Yes, yes. But, yeah, it is fascinating. Some people may have tried different angles on LinkedIn. One that I was curious about a few years ago was this idea: could you get something like LinkedIn’s resumes, but not necessarily references? Resumes are viral; references are like anti-viral or anti-memetic, and people don’t want them on the internet.

If there was a data set that people wanted on the internet, LinkedIn would have done it to some degree. But I think most people who try these attempts don’t appreciate the subtleties. And I’ve actually—I mean, we do have the equivalent of book-blurb references.

Right, endorsements. Yes, endorsements.

We don’t have negative references. Well, by the way, part of the reason why we don’t have negative references is that you have complexity in social relationships. That’s the negative virality point that you were just making. And then you also have complexity—not just legal liability, but social relationships and a bunch of other stuff.

Now, LinkedIn is still the best way to find a negative reference. I mean, that’s actually one of the things that I use LinkedIn to figure out who might know a person.

Yeah. I have a standard email. You probably got a bunch of these from me, where I email people saying, “Could you rate this person for me from 1 to 10, or reply, ‘Call me’?”

The negative what?

Yes, yes. And when you get a “Call me,” you’re like, “Okay.” You don’t even need to take the call.

Yeah, I understand.

Right. And by the way, sometimes when a person writes back “10,” you’re like, “Really? Best person you know?” But what you’re looking for is a set of 8 9s. And if you get a set of 8 9s, you might still call and get some information, but you’re like, “Okay, I got a quick reference.”

Whereas, more often than not, when you’re checking somebody, you’re like, “Oh, you get a couple of ‘Call me’ responses.” And it’s just that quick because it’s email: one-sentence thing, get back, “Call me.” You’re like, “Okay, I understand.”

We have about 10 minutes left, just a logistics check. A couple of last things we’ll get into. Is there anything you wanted to make sure we covered? But we can do this again. This is always fun.

Yeah, and it’s great.

I’m curious, Reid, as you’ve continued to up-level in your career and have more opportunities—and they seem to compound, especially post-selling LinkedIn—how have you decided where the highest-leverage use for your time is? Or where can you have the biggest impact? What’s your mental framework like?

One of the things that I’m sure I speak for all 3 of us about is that it’s an amazing time to be alive. I mean, this AI, and the transformation of what it means for evolving Homo technae and what is possible in life and in society and work and all the rest, is just amazing. And so, I stay as involved with that as I possibly can. It has to be something so important that I will stop doing that.

Now, within that, part of that was co-founding Monase AI with Siddhartha Mukherjee, who’s CEO, author of The Emperor of All Maladies, and inventor of some T-cell therapies. Getting instruction from him on the FDA process—that’s the kind of thing that makes us all run screaming for the hills. So, that kind of stuff.

But also, one of the things I think is really important is, as technology drives more and more of everything that’s going on in society, how do we make government more intelligent on technology? And so, in every kind of well-ordered Western democracy, I’ve been doing this for at least 20 to 25 years: if a minister or senior person from a democracy comes and asks for advice, I give it to them.

Just last week, I was in France talking with Macron, because he’s trying to figure out, “How do I help French industry, French society, French people? What are the things I need to be doing? If all the frontier models are going to be built in the US and maybe China, what does that mean for how I help our people?”

And he’s doing the exact right thing, which is, “I understand that I have a potential challenge. What do I do to help my people?”

How do I reach out? How do I talk? Sure, they’ve got Mistral, they’ve got some other things, but how do I maximally help with what I’m doing? And so, I’m putting a bunch of time into that as well.

Yeah. I remember seeing your calendar, and it seemed like 7 days a week, with meetings absolutely stacked.

I’ve gone to 6½ days, so I’m glad you’ve calmed down.

One of the ways in which you’re able to do that is, one, you work on important problems, but two, you work on projects with friends.

Yeah, sometimes over decades.

Maybe we’ll close here. You’ve thought a lot about friendship. You’ve written about it. You’ve spoken about it. I’m curious what you’ve found most remarkable or most surprising about friendship, or what you think more people should appreciate, especially as we enter this AI era, where people are sort of questioning what the next generation’s relationship to AI will be.

I actually am going to write a bunch about this specifically, because AI is now bringing some very important things that people need to understand, which is that friendship is a joint relationship. It’s not, “Oh, you’re just loyal to me,” or, “Oh, you just do things for me. This person does things for me.” Well, there are a lot of people who do things for you. Your bus driver does things for you, but that doesn’t mean that you’re friends.

Friends—for example, a classic way of putting it is, “Oh, I had a really bad day, and I show up to my friend Alex and I want to talk to him.” And then Alex is like, “Oh, my God, here’s my day.” I’m like, “Oh, your day is much worse. We’re going to talk about your day versus my day.” Really, that’s the kind of thing that happens, because what I think fundamentally happens with friends is that 2 people agree to help each other become the best possible versions of themselves.

And by the way, sometimes that leads to friendship conversations that are tough love. They’re like, “Yeah, you screwed up, and I need to talk to you about it.” It’s not—I tell you, the whole sycophancy phase in AI, and I was—It’s not that. It’s like, “How do I help you?” But it’s also part of the thing that I gave the commencement speech at Vanderbilt a few years back, and it was on friendship. Part of it was to say, “Look, part of being friends is not just, does Alex help me, but Alex allows me to help him?” And as part of that, that’s part of how I become a deeper friend. I learn things from it. It’s not just helping Alex. That joint relationship’s really important.

You’re going to see all kinds of nutty people saying, “Oh, I have your AI friend right here.” It’s like, “No, you don’t. It’s not a bidirectional relationship.” Maybe an awesome companion—just spectacular—but it’s not a friend. You need to understand that part of friendship is when we begin to realize that life’s not just about us. It’s a team sport. We go into it together.

Sometimes friendship conversations are wonderful and difficult, and that kind of thing. And I think that’s what’s really important. And now that we’ve got this blurriness that AI has created, it’s like, “Shoot. I have to go write some of this very soon,” so that people understand how to navigate it and why they should not think about AI anytime soon as friends.

One thing I’ve always appreciated about you as well is that you’re able to be friends with people with whom you have disagreements, or people with whom you were not close for a few years, but you can reconnect and sort of—

Yeah, that ability is—you know, it’s about us making each other the better versions of ourselves. And sometimes those go through rough patches.

I think it’s a great place to close. Reid, thanks so much for coming on the podcast.

And I hope we do this again.

Excellent.