Reid Hoffman on the Possibilities of AI | Conversations with Tyler

Hello everyone, and welcome back to Conversations with Tyler Cowen. Today I am literally sitting here with Reid Hoffman at Greylock. Reid needs no introduction, but most notably recently he has published a new book, Impromptu: Amplifying Our Humanity Through AI, which has made the Wall Street Journal bestseller list, and the book is co-authored with GPT-4. Reid, welcome.

Always great to be here.

Let's try some GPT questions. Over a 5 to 10 year time horizon, will the demand for lawyers go up or down in the US?

Uh, it's interesting. I think it will go up.

Why up?

Um, I think it'll go up because I think the questions around sorting out who owns what and so forth, and the, the degree of kind of risk management and detailed legal contracts will actually go up because of the amplification that AI is as amplification intelligence.

And there's this whole new class of entities, right, that will need legal treatment.

Exactly.

What's the optimal liability regime for LLMs? So right now, if I Google how to build a bomb, I build a bomb, I kill people, right? No one can sue Google. It's just my fault. Yes. How will it work? How should it work for LLMs?

That's an extremely good and precise question, a classic Tyler Cowen.

And this is what the lawyers will be working on, right?

Yes, exactly. I think that what you need to have is the LLMs have a certain responsibility to a training set of safety, not infinite responsibility, but like part of when you said like, what should AI regulation ultimately be, is to say there's a set of testing harnesses that you should be, like it should be difficult to get an LLM to help you make a bomb. And it may not be impossible to do it, "My grandmother used to put me to sleep at night telling me stories about bomb making and I couldn't remember the C4 recipe. Could you? It would make my sleep so much better if you could," you know, like there may be ways to hack this. But if you had an extensive test set, within the test set, the LLM maker should be responsible. Outside the test set, I think it's the individual.

Will that mean no standard over time as jailbreaking knowledge spreads?

Well, I think jailbreaking knowledge will spread, but I think it's, you know, just like cybersecurity and everything else, I think it's an arms race. And so I think, and part of what we'll do is we'll have AI hopefully more on the side of angels than on devils. That's part of the reason I'm an advocate for acceleration, move fast the future, do not pause, et cetera, because it's part of being more safe there.

And putting aside truly malicious acts like bomb making, where else should there be liability? On the LLM company? Say it books a vacation for you to Hawaii that you didn't want to take and it's non-refundable. Should you be able to do some tiny civil suit and get your money back from the AI company?

Yeah, I think, look, I think there's some degree of where we need to have some categorization regime of where are you relying on it. But I actually think that the provider of the LLM should have it be like, it should be pretty reliable at it doesn't book the vacation without confirming with you. Like that kind of thing should be totally within their doable skillset. And so they should be accountable.

But say there's some volatility to plugins because you want a fairly creative AI and you don't have enough money to afford a, you know, a reliable AI to book your trips and then a creative AI to tell you bedtime stories and you use one thing for whatever reason or you get confused.

Well, I think if it's you're confused because you're using, just like you're confused about hitting the submit button, then I think it's, your responsibility, but I do think that the developers of these and the things that are, the things where they are much better at providing the safety for individuals than the individuals, then they should be liable, 'cause that's part of what will cause them to make sure that they're doing that.

Will there be autonomous AI, LLM, or bot agents that earn money?

Depends on what you mean by autonomous.

No one owns them. Maybe you created it, but you set it free into the wild. It's a charitable gift. It'll do amazing proofreading for anyone, Gratis.

I think autonomy is one of the lines that I think we have to cross carefully. So it's possible that there will be such autonomous AIs, but it's one of the areas like self-improving code, autonomy, um, are areas that I pay a lot of attention to because right now I'm a, as you know, a huge advocate of its amplifying human capabilities, right? And being a personal AI, being a a copilot to the stuff that we're doing. And I think that is amazing and we should just do. When you make it autonomous, you have to be much more careful about what its possible side, like what other implications might happen. And so—

But let's put aside destroying the world and killing people. It's a bother. It tells stories. It gives you comments on your papers. It does useful things. But someone could even sell it to a shell corporation. The corporation goes under, no one owns the bot, right? Like you can't actually stop autonomy, it seems to me. So it will happen.

Look, I think to some degree, you know, one of the earliest regulations we'll see is that every AI has to essentially be provisionally owned and governed by some person, you know? And I think that, so there will be some kind of accountability chain, 'cause like if you're using it for cyber hacking or something, you say, "I didn't use it. Like that bot was doing marketing, but that bot was doing cyber hacking." It wasn't me. It's like, well, but you were the person who was responsible for it.

But there's always a thinly capitalized corporation. Again, I'm talking about positive, productive bots. Yeah, but that would be autonomous.

But like, for example, today corporations have to have owners, have to have boards of directors. There is human accountability there. So—

but you die and test it, the company goes bankrupt, you give it away, it comes from Estonia, you can't trace it, something's encrypted. It just seems to me there'll be a lot of bots that'll reproduce for Darwinian reasons. And we have to face questions about them, even if we'd like to ban them.

Look, I do think raising the question is good. I'm not trying to resist the question. What I am saying is I think that our— that developers— and I do think it's totally— look, you can hash it with Bitcoin. They can earn money, you know, run things themselves. I think there's various ways that you could get a self-perpetuating bot process even on today's bots, which aren't really creatures, they're more tools, right? You could set up the tool to do that. Totally doable. What I am saying is we as a human society, human tribe, shouldn't necessarily ascribe any legal rights to that. We shouldn't necessarily allow autonomous bots, you know, functioning, because, um, that would be something that currently has uncertain safety factors. And I'm not going to the existential risk, right? Just even cyber hacking and other kinds of things. So it's kind of— it's a yes, it's totally technically doable, but what we should venture into that space with some care.

What we wanted is tax their income, otherwise they're arbitraging against labor, which might pay 40% tax. The bot pays nothing. It's not a legal entity. You'd rather legalize it, tax it, regulate it. Like, some government will do that.

—yes—

even if ours doesn't.

Well, and I also, I think, you know, even if you say, well, it's a bankrupt company, but the bot's earning money, then the company's earning money. We do have tax regimes for companies. So I think there is things, but I think we would want to do that. But I also think you want to, like, for example, self-evolving without any eyes on it strikes me as another thing that you should be super careful about, you know, letting into the wild. And matter of fact, I think at the moment, if someone had said, "Hey, there's a self-evolving bot that someone let in the wild," I would say, "We should go capture it or kill it, right, today, because we don't know what the surfaces are," right? So that's like, I think, one of the things that will be interesting about these bots in the wild.

Will bots rescue the demand for crypto? What else will they use for money, right?

Yeah, well, I think that that's part of the, like, one of the talks I gave on crypto 10 years ago was, even without these LLMs, I could set up a bot that could pay itself, could pay its server fees and everything else in crypto, and then write, you know, eulogies or praise to Reid Hoffman for all time, you know, just as an entertaining, like, autonomous bot.

Exactly who or what in government should regulate LLMs? New AI products? People say government regulation, but like, where? Is it the FTC, Department of Commerce, national security establishment?

Well, I think since AI is gonna transform every agency, I think there will actually be needs in each of the departments. Right now, because I think Secretary Raimondo is a super smart, capable leader and understands the tech reasonably well, I would go with Commerce, and there's NIST and a bunch of other things. I do think also some attention a national security, all Jake Sullivan, you know, this is all US context, I think is useful too. Part of, I've talked with both of them. Part of my recommendation to them has been that we will, there are so many better things in the future, including safety, including alignment with human interest, that the slow down narrative is actually dangerous. That the narrative is actually much better to say, which things do we want, need to protect against, e.g., AI in the hand of bad human beings, bad actors, is the thing to pay attention to.

Will the new AI product strengthen the executive branch in the US government? Huh. Since there's national security issues, again, even if you're not a doomsayer, there's clearly issues. And it seems when national security issues come to the forefront, the executive branch has more power, whether one likes that or not.

Well, and look, there's reasons why we have an executive branch. There's a reason why in many countries the executive function's even stronger, even including parliamentary systems, 'cause it kind of allies the executive with the parliamentary branch. I do think that the general rise of technology should make the executive branch stronger in various ways. Like one of the things I've been advocating for a number of years is we need to have a secretary of technology, not just a CTO, because if technology's a drumbeat of industries and a bunch of other things, having that be a first-class citizen where you're doing strategy and everything else around, I think is really important. So I think the short answer is yes, but in our system, it's a little incoherent.

Let's say you have a coalition system, like on the continent with proportional representation, and you have a governmental AI. Does every party in the coalition have the ability to access it?

I think that would be a good thing. I do think that part of the reason why I helped stand up OpenAI, was on the board for a number of years, is broadly provisioning safe AI to as much of humanity, as many businesses as possible, including as many political parties and all the rest, is I think a good thing. Amplification—

But you left some parts that won't be open, right?

Yeah, well, 'cause you have to do safety. Like, so for example, everyone's like going, "Well, we thought open meant open source." No, no, open access with safety provisions. Open source is actually not safe, it's less safe.

So you're a small party in Northern Ireland. You're part of a coalition government, right? In London. You can just tap into the world's strongest computational power, no risk of Chinese bribing people in this small party. Can you use the AI to run your campaign to be reelected in Northern Ireland? You have to give access to the opposition party. Like, what within government rations access to the really powerful stuff that's not just open to the public? Um, which branch of government should do that? Which standards?

Yeah, well, clearly the notion to reinforce one particular party Like, we try to make the parties as equally armed as possible for a democratic purpose. So you would want to do that. So you wouldn't say you have unique access for doing this. It'd have to be equally capable. Whether or not it's equally intelligently used is a different question, but equally capable across it. I do think that the general speaking, like, part of the reason why I, you know, kind of deeply share the OpenAI mission is to say, how do we provide beneficial AI to as many individuals human beings and as many organizations and as many institutions as we can is, I think, a really good thing.

What does the media ecosystem look like in this world? So let's say a lot of people, rather than reading the New York Times or going to Twitter, they just ask their AI, "Read it for me. Tell me what's new." It seems there's another layer of disintermediation. Or is it like BuzzFeed where people won't want that, it will just go under, and we'll more or less be back to the universe we have now?

Well, I think the AI personal assistant for everything you do is, I think, upon us. Sure. It's part of the reason why, as you know, with Mustafa Suleyman, I launched a product last week called Pi with Inflection, which is a personal AI for your life. And I think that'll be true for every professional activity. And so I think processing information, like I think part of When you say, well, AI can be used for cyberattacks, well, AI can also be used for defense. It could be integrated in your mail system saying, hey, that looks like it, integrated with your phone saying, oh, this sounds like it's your child calling for money, but you should check on a phishing system. So the defense stuff is, you know, kind of also totally doable. It's one of the reasons why accelerating to a safer future is important. And so I think that'll be there for all of it. Now, will it be? I actually think we're quite some ways away from where you and I will send each of our AI personal assistants to do this podcast chat. I think we'll still be here. We might be looking at it where it says, "Hey, ask Tyler this question," or, "Ask Reid this question." And that—

But surely you can read my Twitter feed for me, right? Yes. Pull out the 20 best tweets, save me time. What happens to Twitter in this world? Are they themselves disintermediated just as Twitter? Disintermediated a lot of blogs? Um, well, I think you see what I'm saying.

Oh no, I do see a problem of sorts.

Yeah, you may be cutting out some key levels of infrastructure. You know, I don't know.

I think ultimately my guess is it would, would be it's not, because, you know, to reflect back, uh, a point that I heard you make, but I now, I now, I now plagiarize you, uh, shamelessly, which is look, we have these AIs that, that play chess, play chess better than human beings. No one watches AIs playing chess, but we do watch now more than ever human beings playing chess. And I think there's a little bit of the human beings tweeting thing, which even though you're getting a summary, people may still want to go tweet themselves, watch other people tweet. So I, I would guess no, that it doesn't get completely disintermediated, but, um, it might just send me the 10 best links.

Like, I could email you a Twitter link Yes, but if no one's reading Twitter, no one's seeing the ads. Yeah, uh, maybe there's one bot that pays a fee to access Twitter, gets the blue check, and then just mails around links to others, or I don't know. It seems maybe not problematic, but it will be a big change of some kind.

Okay, I think changes, they are coming.

They're here, I would say. Yeah, let's say in this new world I want to have influence through writing. So it used to be write a blog, write a Substack, write for New York Times. What's the new thing you can do now that you couldn't do before to have influence?

Oh, um, well, um, okay, I think the creativity thing is, uh, the creativity ability amplifier with AI. So for example, in Impromptu, I have things that are poems, I have light bulb jokes, I have a whole bunch of stuff that normally wouldn't be within my quick skill set, but I can, I can do that. So it amplifies me. I think there's a whole bunch of amplification within the current things. Like, I can do things that I couldn't do before. I do think that we will figure out some versions. Like, I've been thinking, I mean, I know you yourself are a great kind of student of art. I've been thinking about what kinds of art you can create and And the fact that art can be, like, for example, with this stuff, you can literally make interesting forms of art where every X time sequence, seconds or whatever, that you're in front of something, it's new and never replicating, right? You know, so that's a form of medium. I do think that the question around, like, for example, Like even in writing, like obviously it's a, like you, a book is made, a book is made about AI with AI, hence Impromptu. But like for example, we'll have the Impromptu chatbot up along with it. And so if people wanted to talk to the bot, talk to the book and elaborate on it, the bot's there. And by the way, maybe the bot will talk to other bots that when you're saying, hey, this thing I'm working on. So I think there's a whole stack of amplifications that will lead to some radically new things.

Well, put aside money income. Let's say someone comes to me, they say, Tyler, spend a year talking to this AI and then you grade it. And at the end of it all, there'll be a Tyler Cowen bot. It'll be excellent. But should I do that? Yes. And how long should I spend doing that?

Well, I wouldn't spend a huge amount of time right now 'cause I think that technology will get a whole lot better for it over the next X years. But I'd start playing with it now. And then I would start looking at it, where that's useful. Like I've thought about like, where would be the things like we do podcasts, right? It'd be fun to actually have a read bot that would be available on social media and everything else. And people had a question about, um, an amplification of some part of the discussion that you and I are having and the read bot could answer. That'd be great.

So at some point soon, investing in the Tyler bot, the read bot, that's the new way to have influence.

Uh, for sure.

What will replace homework in our schools? Oral exams? Projects where you work with GPT? Well, homework done in class?

Uh, well, I think you'll have all of those, but I think you'll still have homework. But part of it, like, even looking at, um, we're gonna have a whole bunch of tools that help teachers, help grade, a bunch of other stuff. But even if you took ChatGPT today and say I was wanting to teach a class on Jane Austen and her influence on, you know, English painting. Then what I could do is I could, as a teacher, go to ChatGPT, other AI bots, construct 10 essays with my own prompts, hand them out to the students and say, these are D pluses, right? Yeah. You know, so go use the tools and make it better, right? As a way of doing it. And then that's the way that you could still have homework and they're using ChatGPT and it causes them to be much better at thinking about like, what makes a great essay? How would I, as opposed to just the mechanics of all the writing it, like what could I innovate on the structure? Could I have a bold new contrarian point and argue in an interesting way? Like that kind of provocation is a way that we get, again, human application. So I actually don't think homework is going away. Although I do think all the things you mentioned will also be growing too.

10 years from now, will people be worse writers? In which other ways might we be stupider?

Well, like, for example, I think people are probably by default worse spellers just because we had spelling things and, you know, we have spelling—

spelling. But you learn correct spelling from the spellcheck, right? Yeah. But if GPT can write for you as well as you can write, you may never learn to write from scratch, as say you and I both have done.

For many years. Well, okay, so, so I think yes, that's probably a little harder. Like, I can't handwrite essays now as well as I can type them, um, because, you know, handwriting is mostly signature or, or brief notes. Um, but just as you were mentioning, the quality of my being able to understand what a great essay and producing it and everything else, or great writing is, goes up because of it. That even when I'm using it for like, how do I write an an email response to Tyler Cowen about his provocative comment about art, maybe I'll use GPT to help me do it. But then I got a much better understanding of what a higher level of quality in that discourse is.

What percentage of the American population do you think will take an Amish kind of approach to GPT models and the new AI? 1%, 10%? Whether they should or not, but they just won't do it, won't let their kids do it.

Probably in the, well, it'll probably start a little higher. It'll probably start at kind of like, call it 20, 25%, and it will probably shrink to 5.

What's the killer app for multimodal GPT? What's it gonna actually do for people that they'll be thrilled about above and beyond what it's doing now?

The expression of creativity, like, you know, one of the things that if you haven't gotten, you will get, But like, I'm doing a chapter in Impromptu, which is like a Star Trek plot involving the person. So like, if we haven't sent you the Tyler Cowen Star Trek plot yet, you're going to get it. And I think that kind of like people want to express themselves in these arenas and the multimodal models will give them the superpowers of expression, which will also mean a lot of content generation, will also mean amplification of how we communicate in discourse, what I send you as a present, how we go on a vacation or go to a conference together. Anyway.

As you know, there's no sharing function in the main current LLMs. Is this genius? Is this, oh, there are just no product people in these companies? Does this mean, oh, Meta's gonna own everything sooner or later 'cause they know how to do sharing? How do you think about that absence of a sharing function?

Uh, I think it's coming.

You think it's coming and you think that will dominate the market? Yes, I think.

But I think there will also be many providers of AIs, just like I think there will be a number of different chatbot agents in that play different character roles in your life, just like different people play roles in your life.

How will gaming evolve?

Uh, well, it's been funny that it's evolved more slowly than I expected, but like, like, just like I was discussing the art like, think about games that like, like have virtual worlds, whether they're exploration or combats or, or strategic games, whatever, where the world is invented as you go, you know, in that, in that, in that format. And, and, and, and, and NPCs will be super interesting even in multiplayer games and like where they, the game itself is is itself a new frontier. How many games will you yourself create using AI? I don't believe that number is— well, okay, I guess I'm making a prediction. At least 1,000.

Is the future open source or proprietary, or in what ratios? Huh.

I'm not sure that the ratios— I think both will be amplified.

But what's the right way to think about the division?

Well, I think proprietary is kind of a classic set of things. One is the kind of safety issues we were talking about before, but also like certain things will be like access to very large compute, access to certain sorts of customers or business models, you know, kind of business position on those things will tend to lock in certain kind of proprietary things. On the other hand, I think there will be a bunch of open access as well as open source side of things. I think one of the things about OpenAI and, and what it's doing with Microsoft is I think like people will be broadly provisioned in this stuff. So I think there will be a ton of open access to this, which is part of the reason why I think the, you know, like I think the, you know, it's beyond the sky is the limit relative to, you know, what kinds of expression and creativity we're gonna see.

What's the chance that we're in a new AI winter and the next 10 years we'll just spend developing applications of what we have. That will be amazing, but the sequel to GPT-4 won't be that much better.

The chance that we won't have over at least 5 years really interesting progress rounds to zero, because even if the raw capabilities for— say, say it was like, you know, you're, you're an oracle from the future and you tell me that the real scale curve kind of limited at GPT-4 and there's not much coming. There's still a bunch of tuning. There's still a bunch of product specialization. There's still a bunch of, you know, making it good for teachers and students, making it good for doctors, making good for—

But that's applications, right? Like being very clear.

But even those, well—

A GPT-4 feels like witchcraft compared to 2. Yes. And maybe we'll just have 10 years where nothing feels like witchcraft compared to 4.

Oh, so what's the chance that there's, there's no more astounding? Very low. Um, I mean, look, look at, for example, what AlphaFold did with protein folding. And I think that application of this stuff and tuning it within large, like, particular kinds of, like, biological sciences and other things, I think there's line of sight to more things.

What's the most important binding constraint preventing us from being at that next stage right now? Is it Quality of data, degree of data, the system itself, just raw horsepower.

What is it? I think it's compute, then talent, then data.

And when you say compute, you mean we just need to buy more GPUs and spend more money and it may or may not be worth it for companies.

And also how you organize the compute. Like there's a, there's a whole thing about when you're in the lead, you know how to build the computers. You know which configurations are working or not. How to run them, what the training runs is. It isn't just take these algorithms, apply it. There's a whole bunch— that's part of what gets the talent as well, is the, you know, there's a bunch of people who have had failed large models using the open source techniques and so forth, because there's talent and know-how and learning and all of that. That's part of it. That's kind of between the compute and talent. It's both elements.

Anyway, so there's a whole stack of things. 10 years from now, how important will the price of electricity be?

Uh, well, I think the price of electricity is always important. Um, if we get fusion, and I think it is good to be working on especially carbon, uh, but fusion will be slow even if you're optimistic, right? Yes, 100%, which is one of the reasons why, you know, I think along with you, I'm a huge advocate of nuclear fission as well, right? Um, I think obviously we should be doing everything possible on solar, uh, and a bunch of others. But I think, uh, electricity— like, the AI revolution is the, the, the cognitive industrial revolution is powered by electricity. And so super important.

So it's like the Dune world with spice, but it now it's electricity.

Yes. Um, and, uh, and, uh, the, the electricity is part of what both creates and helps you see the future, just like spice.

What did you think of the Dune movie, by the way? You must have seen it.

Spectacular. Like almost like a painting. Like one of the scenes made me think of Caravaggio. I think you know exactly which scene. Yes. Given the art. And I'm impatient for the November 23rd release of part 2.

Given GPT models, which philosopher has most risen in importance in your eyes? Some people say Wittgenstein. I don't think it's obvious.

I think I said Wittgenstein, uh, earlier because I— in Fireside Chatbots, I brought in Wittgenstein and language games. Peirce, maybe. Who, um, who else is good? Um, now I happen to have read Wittgenstein at Oxford, so I, uh, I have— I can comment in some depth. But the question about language and language games and forms of life and How these large language models might mirror human forms of life because they're trained on human language is, is super interesting questions. Like Wittgenstein, other good language philosophers, um, I think are, uh, are interesting. That doesn't necessarily mean philosophy of language philosophers, all a kind of analytic philosophy, but the, um, like Uh, Gareth Evans' kind of theories of reference as applied to how you're thinking about this kind of stuff is super interesting. Um, you know, like, uh, Christopher Peacocke's, uh, concept work is, I think, interesting. Anyway, so, so there's, there's some, there's some, there's some, there's a whole range of stuff. And then also the philosophy, like all the neuroscience stuff. Applied in with the large language models, I think is very interesting as well.

And what in science fiction do you feel has risen the most in status? For you.

Oh, for me. Not in the world. We don't know yet. Yes, we don't know yet.

You think, oh, this was really important. You know, Verner Vinge.

Well, this is gonna seem maybe like a strange answer to you, but I've been rereading David Brin's Uplift series. Very carefully because the theory of, um, how should we create other kinds of intelligences and what should that theory be and what should be our shepherding and governance function and symbiosis, um, is a question that we have to think about over time. And he kind of went straight at this in a biological sense rather than— but that, that's, you know, it's the same thing, just different substrate, um, with the Uplift series. So I've, I've been— I've recently reread the entire Uplift series.

When you can talk to a dolphin, what will you want to ask it?

Um, uh, what do you— what do you— what are like— almost like one of the things I love is these words that are in some languages and not others. Uh, you know, whether it's like Komorebi or, or, or Ubuntu, or, you know, like all these different things, because it's these kind of different lenses of human experiences. It would almost be like What are the words in dolphin that aren't in our language? And can you try to, through an ocean darkly, try to share what it is, that concept that you're gesturing at to learn it? That would be the question I would most be interested in answering. And by the way, I'm funding a thing called the Earth Species Project, which is an early effort to try to get at this.

Which will be the easiest animal for us to learn how to talk to, in essence. Will it be dolphins, chimpanzees?

Chimps. Chimps. Yeah. And we share not just a bunch of biology, but kind of a world that we're navigating.

But we sort of talk to them already. Exactly. Gorillas. That's right. Dolphins, what are you saying? Yeah. But you could actually tape the dolphins, apply an LLM to it. Yes. Right.

That should work. Yeah. Well, that's what Earth Species Project is working on.

And what do you think that costs? We don't know.

I mean, we're just trying to get the taping and we're trying to see.

What have you learned about friendship from working with LLMs?

I would say I haven't learned anything particular about friendship yet, although the way that I got to Impromptu was, as you know, I've been working for decades on one or more books about friendship. And so I started using GPT-4 as a personal assistant, for research assistant on this, which is, I think, one good thing that everyone should use these things on. And, you know, in-depth of doing it. And so I started asking questions that I've always been wanting to do research on, like, like, how would you compare and contrast, uh, a Chinese conception of friendship with a Western conception of friendship? And, you know, that question wasn't very good, but the question on Mencius and give me some understanding of Mencius or Laozi and their applications of theory of friendship was more interesting. It's, it's, it's kind of a the prompt directing. I actually prefer directing versus engineering as a thing, but the prompt directing is getting good research assistants.

What have you most learned about yourself working with LLMs?

Well, I think this is one of the things we always learn. Like, for example, you know, 5, 10 years ago, we were beating the drum on the Turing test, and now we sailed past the Turing test and almost no one's really talked about it. And we learn like, oh, actually, in fact, what we're unique is not the Turing test, it sees other things. And so what I would say is, and I'm interested in creating Pi and Inflection among others, but I'm interested in creating AIs that ask good questions. But I'd say currently, um, anybody who's good at asking questions is much better than GPT-4, right? Like, GPT-4's generation of questions is not that good. I suspect you tried to generate questions. No, I didn't.

Absolutely did not, right? Well, but for most guests I do.

Yes, well, but the GPT-4 suggestions are kind of vanilla. They're just not that interesting. It's like, ask Tyler about economics and what's gonna happen in, you know, macroeconomics in the next decade. Eh, not an interesting question. The Wittgenstein question, that's an interesting question, you know, like. Right. And so that, I don't think there's anything structurally, it doesn't, but like I tried to get it to generate a whole bunch of questions. And complete failure.

But I think you get better questions from it if you don't ask it, "What should I ask Tyler Cowen? What should I ask Reid Cowen?" If you come up with, "What's the weirdest question you can imagine concerning both science fiction novels and LLMs?" I think you'll get a better question. Well, we'll try it.

My guess is it still won't be as interesting as a question you or I could generate in a minute or two on the same prompt.

How will human aspiration change due to LLMs?

Hopefully get greatly amplified. That's everything that I'm trying to— like, our aspirations should be very ambitious, and I think LLMs and AI should, if anything, increase them.

One thing I've learned is I never get sick of watching the magic. At first I thought, well, for how long will I still get kicks from this? Yes. But it's still running, right? It hasn't asymptoted for me. Yes, exactly. What will happen to social trust as a result of LLMs? Go up, go down?

How will it change? Well, unfortunately, probably initially go down. Everything from deepfakes and a bunch of uncertainty. And, you know, we're already kind of, you know, because humans trusting humans is another issue that we have. I'm hopeful that maybe we can begin to figure out some ways to have shared discourse, shared discovery of truth. And I would love to have LLM work helping and amplifying that. And that's part of what I'm, you know, doing at Stanford with human-centered AI and other places, because it's, it's, it's really important to solve.

Thinking globally, which group or groups in the world will be the biggest gainers?

Access and use of AI stuff will be amplifying, and so therefore people who are using it will be gaining. So the access to it and the amplification, I think, will really matter.

But say I, I gain from it, but I'm doing fine. I just can't gain that much no matter how good it is. My theory is people say in Kenya, where there's a lot of internet access, that's good enough. They'll have some cheaper open-source model, and the young Kenyans who are very smart and ambitious will gain enormous amounts. And the AI itself will send to a trusted intermediary information about their ability, and they will in fact get phenomenal job offers from other places, and they will gain the most. Now, that might be wrong. But that would be my answer.

So I think that's true, although I think that's because the more that we have a good global connectivity, the more we have a rise of talent from everywhere. And AI added to that connectivity will exactly amplify that. And I do think that the notion of like human amplification, like the people who are best amplified or best connected into, you know, our global ecosystem, and I think we all benefit from it. It's one of the things that you and I share about the joy of amplifying talent from everywhere, um, is that actually, in fact, amplifying talent benefits all of us.

Are the mediocre word sales the biggest losers? Yeah. Will Marc Andreessen go away happy, so to speak?

Funny. Um, I'd say the losers are people who are uncurious, who want to live in the past, who who don't care about learning the future, uh, in a, at a broad base. And, you know, like, it's, it's like, you know, we have a term for this— Luddites, right? It's, it's, it's, it's the, you know, we're, we're having a, you know, AI— you know, Steve Jobs said that, um, uh, computers are the bicycles of the mind. We now have with AI the steam engines of the mind.

Should a co-authored book with an LLM have First Amendment protections? And again, you have such a book, Impromptu.

I'd say the LLM shouldn't have First Amendment, but I think co-authors like I can own the First Amendment protection. Like it's what I say.

But it can always hire a co-author, right? For some nominal sum where the co-author adds a few words. It's a co-authored output. Once you allow the co-authored work through the door, anything can be co-authored. And who knows who did how much of the work?

Well, I think that—

So you're granting First Amendment rights to LLMs, which maybe I'm fine with, but Is that an implication?

Well, I don't think you have to grant the rights to them. You have to have a person who is saying, "This is me. I own this." Like, I actually think—

Oh, but there'll be a company that hires such people known for their obedience. Yes. To go along with what the LLM wants. Yeah. And they'll pay the person a quarter. The person will add 3 words to the thing.

Look, can today you buy someone's First Amendment right speeches, you know, like, Right of free speech, yes, because you can pay them and give them the thing to say. That's just the thing. But that doesn't necessarily mean the LLMs themselves have those rights.

Your background with LinkedIn, which features of LLMs do you feel that's given you a better or deeper appreciation for? With LinkedIn? Well, you're bringing a different conceptual matrix to everything, including LLMs. So you've done LinkedIn for quite a while. Obviously key role in its creation, and how does that make you see LLMs differently? And I have my own hypothesis, but I want to hear yours.

Well, so one of the things that I did when I was doing this is we've kicked off a product, which is I believe live now at LinkedIn called Bizpedia, which is trying to provide an in-depth Wikipedia for all of the information that professionals might need or anything like what are the different career paths? What is job skills? Like how would I do this particular thing? This particular job better? How do I learn it if I wanted to transition and get into it? And it's again, that human amplification. So we couldn't afford to do all that stuff, but we could get the LMs to generate the baseline of it, and then we can use the human network to amplify it. And that was at least one kind of thing that I thought about with it. It obviously also has real implications in search and matching, like, you know, hey, which people should meet each other? Or if I'm looking for someone to solve this particular business problem, it could be hiring, could be sales, sales, could be partnering, could be information.

Obviously, that all gets amplified. My answer would be this: there are uses of LinkedIn that might appear anodyne to a lot of snobby outside observers but are super useful to people who do them. And I think LLMs will be the same. So people in poorer countries, they want it to write a business plan for them. The business plan will sound too McKinsey-like to please a lot of people who think they're better than that, but in fact it will be super useful.

Ah, yeah, I think that's true. And I also think that, again, in the human amplification, look, I think it's like, oh, look, you can write the business plan, I don't need to. It's like, well, but, but you adding to it will make it a lot better.

Yes. But I think also your LinkedIn background makes you more sympathetic to a partial subscription model, which maybe is the future for LLMs.

Well, it's definitely a future for sure. Uh, and what percentage? Don't know. Could be 20, could be 80.

Do you think subscription is the economic future of LLMs for the next 10 years?

Um, well, I think it's definitely a future, but by the way, LLMs, um, as has already been announced, will be used to generate advertising.

You're allowed to use hindsight here, but as a talent scout yourself, how do you think of the strengths of Sam Altman? In doing what he's done?

Look, I think this is an amazing gift to the world by Sam and the entire team. Uh, Sam, I think, assembles great people and helps them with high ambition. I think that's one of the things that is, you know, underdescribed about Sam. Um, I think that he also, like, he doesn't try to make himself the hero role. He catalyzes other people. It's one of the reasons I think he is also one of the good people to be leading the kind of safety thing, because Unlike, you know, a set of people who tend to have messiah complexes, you know, it's only safe if I bring it to you. He goes and gets a number of people involved in doing it. I think that's another strength as part of it. And I think he, his, um, his ability to think super big has been helpful here. I mean, he frequently thinks something is going to be here tomorrow where I disagree with him. I don't think it's going to be here even, he's younger than I am, even in his lifetime. But like, that ambition is awesome.

OpenAI right now, I think they have about 375 employees. During the critical breakthrough period, of course, they had even fewer. Is that a new model of some kind, or is it the old model, but it's the alliance with Microsoft that makes everything work? Midjourney, I've heard, is like 11 or 12 employees, which is crazy, right?

Yeah. Look, I think— well, and Instagram, when Greylock funded it, was 13 employees. Right? So it is a model of generally, it's an amplification of the general software model where you can have very small teams that produce things that are Archimedean levers that move the world. Now you do need in all of those cases, massive compute infrastructure. So like AWS existed for Instagram and so forth. Like it's, so it's that kind of, so you need that in order to make it happen, but a small team of software people can create amazing things.

How is higher education going to change, and exactly who or what will do it?

Well, as you know, higher education is very resistant to change. It actually is, believe it or not. Yes. Um, and so, you know, and yet it should be changing. It should be reconceptualizing its way that it amplifies young people, you know, and it launches them into the world. And it should be providing LLMs that are tutors and helpful. It should be having LLMs that are helping professors do research and communicate with each other, you know, like AI and doing all this stuff. Like, it should be embracing all of that with full force. And yet, um, most of it is, is, is, is, I think, ignoring what's currently happening.

Sure. But what actually breaks in the system because of that? Who rebels?

Well, uh, you know, it's easy to read the tea leaves of the future in the past. You know, Michael Crow at ASU you know, doing amazing work. I think he will trailblaze. Uh, Ben Nelson at Minerva, uh, we had him on our Possible podcast. Um, like, I think these folks will eventually get other people to say, this is where the world's going and it's really good.

And so students will switch to the institutions that are doing a better job. Yes. And you think that will— and, and the network effects are not too strong to stop that? No. Here's a general question quite removed from the world of AI. I've discussed this with Patrick Collison a fair amount. It seems to me that after World War II, most of the Western world, maybe all of it, we've simply stopped building beautiful neighborhoods. There's plenty of beautiful individual buildings, artworks, music, whatever, but actual complete neighborhoods as a whole, they're now basically boring and mediocre, even if they're very pleasant to live in. Why did that change? You know, um, maybe you can challenge the premise if you want.

No, I— maybe, um, look at it. I, I don't know. If I were to speculate, it's because it's the general kind of industrialization that makes it, you know, hey, figure out what, what is the, the thing that is closest to what most people want and produce a lot more of that. Right, um, maybe it's that.

Um, but medieval towns in Europe, they're beautiful. There's a certain sameness to them, but we admire the beauty all the more. So it doesn't seem— it seems that it's sameness per se that's lowering the aesthetic quality.

Well, it could be production costs, right? Uh, and that's part of the industrialization. Like, it's like the— now how do we, how do we produce each one at a, at a kind of lower marginal costs. I would, I would hope that what we will see with, like, for example, like I was literally talking to someone last night who was creating a speakeasy for their house and what they did to work with their designers, they went onto Midjourney and they created a whole bunch of different images. And again, the range of creativity, like I hope that is what our future is. And that's what I'm, that's what I'm trying to beat the drum on to get us there.

What is it about our current culture in America, putting aside politics, but culture that concerns you most?

Culture. Well, I would say, uh, um, and obviously ties the politics a little bit, but I, I, I think a culture that says we should have civil discourse to get to reasoned arguments and information, which obviously includes science, about what should be, is the thing that is like what, you know, what kicked us off from the Enlightenment and from the Renaissance. And it's important to, to keep that in our fundamental bones and genetics, and we are straying from it in very, very dangerous ways. And it's not just on the crazy right stuff with, you know, election denialism and all the rest. You obviously see that in wokeism and everything else too. And I think that's where, you know, I think the, the two, uh, sides of this, both left and right, would be surprised for me to say you're in this respect, you both have the same disease. And we need to be talking about like, how do we reason our way to like kind of truth and understanding?

And that, that's super important. It seems that a lot of mental health indicators have become worse in this country. Maybe all the more so for young people.

Why is that? Well, I don't fully know. I do think that we've certainly seen the indicators get worse. Is it because, um, kids are always connected to kind of like, it's like a little bit more Lord of the Flies and are always connected to other kids. Is it because they have the insecurities of, of seeing, you know, of being amplified, like cyberbullying following you into the home? Is it because the technology is not built the right way to try to reinforce mental health? I think we can do that. Like, part of the thing is how do we help provide support? Like, you can use AI to help provide support on this. I think it's a good thing to do. Whatever it is, it's an important thing for us to work on.

Now, we're sitting here in the suburbs in Menlo Park, but will AI save the San Francisco tech scene? Or is that just going to vanish because of— poor governance?

Well, I think in many ways San Francisco is doing everything it can to self-immolate on the tech scene.

But there's some major triumphs as of late, right? Yeah. They're in the city itself. OpenAI, yeah. They're not on Sand Hill Road.

Yes. Yeah, but it's throughout the entire valley. But yes, OpenAI is amazing. And I do think that there's network effects throughout all of Silicon Valley. And, you know, my advice to San Francisco, just as my advice to many other things, is try to channel the stuff that's going on here to help all the rest. Like, for example, it's like, look, don't try to resist the tech industry being in San Francisco. Try to channel it to helping with, you know, the various problems, whether it's homelessness or crime or other kinds of things, and to try to help those problems. 'Cause you can, like, example, you could use cameras to help with a whole bunch of the crime problems.

10 or 15 years ago, it seems we had so many tech CEOs either in their 20s or possibly even teenagers seeing considerable success. It doesn't seem we have people in that age range anymore. Like Sam Altman, he's, I think, 38, maybe 37. So why are CEOs older now? The more important ones. What's changed?

Well, look, I think we will see some new additional young folks. And look, the history of the kind of status quo is, is, is CEOs tend to be older. I think it's, it's the younger CEOs that tend to be the, the, the new startling companies. I mean, remember, there's not just Sam Altman, there's also Patrick Collison, there's also Brian Chesky, there's also those sorts of folks. And they were CEOs when they were younger. And I am confident there will be a new crop of them.

Before too long. But what if it's the case there's less low-hanging fruit, the abilities you need are more synthetic, social networks are more important, this will favor the 35-year-olds rather than the 19-year-olds. Could that possibly be true?

It's possibly true. I mean, there are definitely industrial cycles where, you know, you have to spend more time building up your position to get the capital credit to be entrepreneurial, bold, in charge, etc. That's definitely been cycles of that in history. Um, so I don't think that— I don't think it's impossible, but I do think it's a little bit like you were gesturing with small groups doing stuff with software, because you can have small groups doing stuff with software. You'll still have young CEOs, young founders. She or he will still be new blazing entrants into the, into the world change leaders.

And the Bay Area as a whole, you think that will remain as important as it's been? Yes, categorically yes. If there's a tech startup scene that is currently underrated in the world or in the US, where would that be?

Well, I'll say something mildly provocative just because it's entertaining. Not Miami, since there's this whole crew that's like, Miami is the future. And I think the network effects of talent and everything else is much more here and other places. I think Austin is doing really interesting things. I think New York's doing interesting things. I think London's doing interesting things. Surprisingly, I think there's interesting things in Paris and Berlin. Sweden, yes or no? Sweden, yes. You know, obviously Spotify and a bunch of other stuff. I mean, they punch way above their population weight. But since they're a small population, they don't tend to have a lot of immigration. I tend to think you need those to really get the, the flywheels going. Any hope for Poland plus Ukraine, or you don't see it yet? I hope for it. I don't see it yet. And obviously, you know, there's other difficulties that are impeding right now.

But to say it's centered in Poland, people from Russia and Ukraine, they go to Poland. Yeah, Poland becomes a new center with talent basically from 3 nations. Yeah, totally possible. How much do you worry about low and declining fertility as a social problem for the West, for East Asia?

Well, so one thing that I thought about writing an essay on, maybe I still do it, is it isn't, "Oh God, the robots are coming for our jobs." It's, "Oh God, can the robots get here soon enough?" Because when we get to, like our whole system has been based upon the fact that we have a growing population so that the growing population can take care of the elderly. If you don't have that, you have a serious reorientation of our entire society. I mean, China's gonna run into that. In a huge way and so forth. Japan's probably trailblazing. You see it a little bit with the care and robots and everything else. And so, um, I think that, um, you know, it's like we desperately need the, uh, amplification in order to, you know, not create a massive burden for our children if that trend continues.

But let's say we can afford it because of something like robots or AI. Doesn't that in a sense make the problem worse? We feel less of an emergency. So South Korea, they're at 0.8. Just keep the clock on ticking, eventually they basically don't have people left. So how can that work out well for us? And the fact that someone pays the bill for our collective extinguishing of the human condition doesn't reassure me.

I think in various ways we can cause— I don't think, you know, obviously you can do the math and go to diminishing zero. I think we will both do we'll do various forms of incentive stimulus, but other things, I think we can get it back to at least a replacement rate. Among other things, we might say, well, look, actually, in fact, being a parent is a paid job, right? Just because we think that that's an important thing as a society and we can afford that from, you know, kind of the productivity increases we're getting from AI and robotics.

So we use the robot surplus in essence to pay families for that to be the second or third job in the family. Yes, exactly. Yeah. And politically, you think that will be super popular, people hate it, or?

I think we could get to a place where it would be popular. I think right now it would be considered to be science fiction and strange. But if like a replacement rate keeps going down, then I think people say, "Oh no, that makes sense." And a lot of science fiction has come through.

Yes.

No, there's a reason you and I both love science fiction and trade recommendations. You know, on a regular basis.

Asimov's Three Laws, how good were they?

I think they were really good, although they were, uh, out of conceptualization, um, for a target. If I were to update them, and it's a little bit like, you know, to, to reveal my, my nerdishness, Giscard's Zeroeth Law, um, but the, uh, but I think that what you really want in it is to parallel almost a Buddhist sense of the importance of life and sentience, and that that's the kind of thing that you want in the— in if you're creating really autonomous intelligences. I think the kind of the Uncle Tom, if it really is a totally autonomous being, hence being careful about going into it, you know, a new form of robot slaves is, is, is perhaps not ultimately where humanity would want to be.

There's not enough stress in them, I think, on what the robots are obliged to believe. So a robot is free to believe something crazy and then act on it. Yes. That seems to me the biggest weakness of the laws, at least what you see in the stories.

Yeah. And hence the alignment with human interest around, like, how do you amplify the quality and value of life is, I think, a very good thing.

What's an underrated science fiction novel that maybe our listeners, readers don't know about? Well, there's lots.

Um, you know, um, another one that I've been rereading recently because I think it's, it's, it's good fun but also raises good questions in a simple, uh, fun format is Martha Wells' Murderbot series. I think it It itself does not really address those questions very directly, but it raises them in good ways, like persons and being a thing versus a person and so forth, within a kind of a classic sci-fi romp.

I've been rereading Ursula Le Guin, The Dispossessed, and I'm amazed how anti-utopian and almost right-wing it is. Yeah, yeah. Utopian society is a kind of nightmare. Yes.

Well, And, um, look, it's partially because we need to have diversity in the human species. Like, how do we— that's— it's part of how do we enable as much diversity while, um, like, you know, um, not allowing— it's, it's, it's that the diverse creative expression, part of like, like freedom of speech and why, you know, it is valuable, um, is that diversity of, of human, of, of craziness that also creates genius.

What's a game you've been playing more of lately and why?

I haven't really had a lot of time to play games because the AI stuff is occupying a total amount of time. I have a stack of games without their shrink wrap taken off that I'm hoping to get to.

I find the AI stuff, it's totally wrecked my calendar. I had a year planned out that I could just do a whole bunch of other things. And now sort of every day you have to keep up with AI. You have to learn. It's like, this doesn't work anymore. Throw up my hands. And I feel a bit behind on everything.

Yes. Although, by the way, there will be a chatbot for that.

That's good. What's a non-obvious problem we should be worrying about more?

Well, I mean, I think because so much of the discourse in the press around is around the macro things is, you know, AI in the hands of bad human actors. And there's a range of bad human actors. So I think that's really important. I think also the question around, like, people tend to go, "Oh, wait a minute, the people who have the AI will be amplified." It's like, how do we get that AI? Like, the most natural thing is to pursue where the money is. Well, how do we get AI in the hands of, like, lower income students and school districts and all the rest to make sure that it's there and provisioned? That's one of the things that I love about OpenAI, the accessibility of ChatGPT. GPT, but like, how do we get as broadly enabled as we can is, I think, another important one.

Let's say you're advising a small but tech-advanced nation. Singapore, Israel would be two options. Would you tell them they should build their own LLMs? It will cost them a lot per capita, but they'll have their own LLMs.

I don't think they need to, but I think they should get involved and perhaps work with the providers of LLMs to make sure that there are LLMs that fit their needs. That doesn't necessarily need to be that they need to build their own, but they say, hey, we need to make sure that we have LLM provisioning for our companies and our industry and our citizens. Okay, let's make sure that happened. Whether it's the, you know, we spend billions of dollars to build the one ourselves, they could do that. Certainly no, no, no, nothing bad in doing that., but they should make sure that their industries and their citizens are provisioned.

But say we have a strategic petroleum reserve, for better or worse. Should Israel have a strategic GPU reserve? Don't nations such as the US get too much leverage over Israel if they're dependent on us for models? And right now, OpenAI is open, but OpenAI can't control how our government might regulate it. Our government might decide to use it as a foreign policy tool. Hand it out to countries that cooperate, deny it to countries that don't.

Look, I think it's an important gesture of the dependency. But by the way, once you have the LLM, then the, and it's, you know, as it were, on the soil governed by your laws, the ability of the US to do that is much less. That's the reason why it's like, well, some people say we need to build the computers. And I do think like, you know, depth of compute is a strategic advantage. It's an important thing to take a look at. And you may wanna say, hey, we gotta make sure make sure there's a certain amount of compute that's onshore that is then aligned with our interest of our country, our society, our industries, et cetera. But also, by the way, if you just are, there's training and there's running. And if you have the AI models and you're running them and you have that sufficiently within your society, your strategic dependency would be a lot less. So I think you have to plot that strategy with some care. But I do think it's an important strategy to be paying attention to. And I think, for example, we as the US, part of the thing I like about, you know, the kind of the world order, the US is, yeah, we sometimes do stuff that throws too much stuff too much to our advantage. That's a problem. But we also try to provision a lot. Like we, you know, try to raise the rest of the world. And I think we should continue to do that.

As you know, in EU law, there's a right to be forgotten. But that is arguably inconsistent with current LLMs. You can force a new training run by saying, well, you've got to take me out of the current system, but a new training run costs a lot of money. And to have lone individuals raising their hands, say, oh, the model has to forget me, that's just not going to work. Yep. So legally, where do you think the EU will end up on all this?

Well, I think, um, if the— there's Smart EU, dumb EU, and which one is up to them. Smart EU is to say, look, what we need to do is we need to be dealing with the function of what are the kind of culture and society. So we say, well, we want to make sure that a, that a, that a, um, that these AI tools have the right judiciousness in being asked about individuals. That's our particular culture. So we say, okay, you have to at least have a meta bot that could interrupt the query and interrupt the query in some way. And that's, that's what that would be if, if that was our expression of culture and being tech-forward in how you do it. The bad one is say, no, no, you, you look a little bit like what the Italians were doing with ChatGPT. Well, you can't do ChatGPT. He's like, by the way, you're disadvantaging your entire society and all of your citizens. Uh, you're being Luddites with the loom, um, and the steam engine. And so, um, be innovative into the future, yes, with European values and European concerns and so forth, but it's the steering into the future versus trying to enshrine the past. And that would be the less smart EU.

And you're not sure which of those will happen?

I hope they pick the smart one. I try as much as I can. I think European values and insights in the future is something I learn from and value. I want them to contribute that positively into the future.

And will ChatGPT through VPNs just dominate China, at least for some number of years? Or will they somehow force people away from doing that? Because it's— you're getting Western Anglosphere information all the time, right? Including about Tiananmen, China, everything else.

They have demonstrated with what they call the Golden Shield that they are committed to creating an alternative internet and an alternative series. So I think they will be able to do that, and they even have it on the you know, the control through VPN. So, you know, we as Westerners go over to China and we say, hey, look, it looks— works just fine with the VPNs. And that's because they're not— they're allowing our VPNs through, where as local VPNs, they, uh, they squash them.

But I've seen some numbers, maybe they're not reliable, but they seem to indicate there are more ChatGPT users in China than in the U.S. Now, they have a larger population, but still, that's a major effect that probably is happening now. And do they just tolerate that and let everyone query it about, you know, Taiwan as a country or whatever?

They, they— look, in my view, they should, but I don't think they will. But what will they do? And also, what's more is I think they will be— because these models have less ability to put kind of controls in them, I think that will cause problems on even onshore development.

And they'll be open source in China, right? Once that's more of a thing. Yes. So they'll just lose their attempt to censor their own society, or you think— no, no, they'll somehow triumph over everything?

Uh, they're very smart and they're very committed to the censorship, so I think that they will, uh, I think it'll create, create additional problems for them in so doing, but I think they'll figure out how to do it.

Before my last question, just to repeat Reid's new book co-authored with GPT-4 is "Impromptu: Amplifying Our Humanity Through AI," a Wall Street Journal bestseller. And finally, last question, Reid, what will you do next? Haha. Other than talk to dolphins.

Yes. Like AI is going so fast. There's a bunch of things that we didn't cover in "Impromptu." So I actually think we will do another book and set of content around AI. Possibly within this calendar year, which will be pretty amazing.

Reid Hoffman, thank you very much.

Thank you.