Why do you care about philosophy? Why are answering these big questions important?
You know, one of the things I sometimes will tell MBA schools is a background in philosophy is more important for entrepreneurship than an MBA. Philosophy is very important to this stuff because it's understanding how to think about very crisply what are possibilities, what are theories of human nature as they are manifest today and as they may be modified by new products and services, new technology, etc. Usually in this show, we talk about actionable ways that people use AI, but a more interesting question is how does AI in general and how might it change what it means to be human? These are really deep, big philosophical questions. I thought you might have a unique perspective on this intersection.
Reid, welcome to the show.
Great to have you. So, I'm sure that everyone listening or watching knows this, but you are a renowned entrepreneur, venture capitalist, author. You're best known as the co-founder of LinkedIn. You're a partner at Greylock. You were a board member and an early backer at OpenAI. And you also have an incredible podcast, Masters of Scale. But perhaps most relevant to this conversation, you also studied philosophy at Stanford and Oxford, and you almost became a philosophy professor, which I didn't know before researching this interview. It's really cool.
Yeah, it was definitely part of it. I've always been interested in human thought and language. I started at Stanford with a major called Symbolic Systems. I was the eighth person to declare that. And then I thought, we don't really know what thought and language fully are. Maybe philosophers do. So I trundled off to Oxford to see if philosophers had a better understanding of it.
I love it. It's funny, I feel like since then, Symbolic Systems has become the go-to Stanford major for curious, analytical people who end up doing startups. So it's pretty funny to know that you're one of the first. So, usually in this show, we talk about actionable ways that people use AI. But underneath that, I think a more interesting question is how does AI in general, and AI in particular, how might it change what it means to be human? How might it change how we see ourselves and how we see the world? How might it enhance our creativity, our intelligence, all that kind of stuff. And these are really deep philosophical questions. As someone who rigorously studied philosophy and probably still thinks about those questions, I thought you might have a unique perspective on this intersection. I think people tend to be either in the philosophy camp or in the language models camp, and people who are sort of in the middle is kind of interesting. I wanted to start with why you care about philosophy. Tell me more about why you care about philosophy, why are answering these big questions important.
So, one of the things that I sometimes will tell MBA schools when I give talks there is a background in philosophy is more important for entrepreneurship than an MBA, which of course is startling and contrarian. Part of that is to get people to think crisply about this stuff. Part of what you're doing as an entrepreneur is thinking about what is the way the world could be. What could it possibly be? Using analytic philosophy language, logical possibility or something like that. It's kind of what is possible. And then partially because these are human activities, what's your underlying theories of human nature about how human beings are now, how they are quasi eternally, and how they are as circumstances change, as the ecosystems we live in change, which is technology, political power, institutions, and a bunch of other things. Philosophy is very important to this stuff because it's understanding how to think about very crisply what are possibilities, what are theories of human nature as they are manifest today and as they may be modified by new products and services, new technologies, etc. People tend to say, 'Oh, that's a philosophical question because it's an unanswerable question,' like the nature of truth or how we all speak and understand languages but don't really know how that works. It's part of the reason why there was the linguistic turn in philosophy that Wittgenstein and others were so known for, which is, 'Well, maybe these problems in philosophy are problems in language, and if we understand language, we'll understand philosophy.' This question around unanswerable questions, but actually science itself is full of a lot of unanswerable questions. It's the working theory as we dynamically improve, and that's part of the human condition. That isn't to say that the same questions today in philosophy are the same questions that Plato and Aristotle and even the Presocratics were grappling with, like truth, knowledge, etc. But some of the questions are also new questions, and the questions evolve. Part of how sciences evolved from philosophy was this question as we get to more specific theories and develop new questions. The same thing is true in building technology, in building products and services, in entrepreneurship. That's why philosophy is robust and important as applied to serious questions. One of the things I wrote my thesis on at Oxford was the uses and abuses of thought experiments. The most classic one is trolley problems. There are both uses and abuses within the methodology of trolley problems. The most entertaining of which, if people haven't watched it, is a TV series called The Good Place, which embodied the trolley problem in a TV episode in an absolutely hilarious way.
That's really interesting. What is the way that people tend to misuse that? I feel like trolley problems are so common in EA discourse and people run into that a lot online.
The fundamental problem is they try to frame it to get an intuition, to derive a principle. They try to frame an artificially different environment. So it's like, 'No, no, it's a trolley and the trolley will either hit the five criminals or the one human baby, and it's default set to hit the human baby. Do you throw the switch or not?' Then when you start attacking the problem, you say, 'Well, how do I know that I can't break the trolley? I could just not make it continue to run.' They say, 'Well, but you know that.' You're like, 'Oh, so you're positing in your thought experiment that I have perfect knowledge that breaking the trolley is impossible.' In your posit experiment, you're positing something we never encounter, or when we encounter, we generally think people are crazy. Like you have perfect knowledge. Why do I know that I have perfect knowledge that I can't break the trolley? What is the right human response to this trolley problem? I'm going to try to break the trolley so it doesn't hit either of them.
Right. That's really interesting.
Right. And you might even say that the problem is that to say you have perfect knowledge that you can't break it, you're like, 'Well, okay.' A, you don't have perfect knowledge, and B, even if you did, maybe it's still the right response. You're trying to get me to say, 'Do I do nothing and run over the baby, or do I do something and run over the five criminals?' Those are my only two options. And you're like, 'Well, no, I could say even if I think I can't break the trolley, that's what I'm going to try to do because that's the moral thing to do.'
I've heard a lot of trolley problems and I've never heard anyone pause at the third option. I love that. That's great. There's something about that where certain thought experiments sort of hijack your instincts and you don't quite reason through all these hidden assumptions. That honestly reminds me of certain doomer arguments. I don't want to go into the full thing, but I think it's a really interesting way to think about it. If I had to summarize what you just said, the value to you of philosophy is thinking crisply about possibilities, thinking about human nature and reality. All of those things are really important for business people. I want to take it another step. Some of those questions that philosophy students or philosophy nerds just sharpen our skills on, there are so many of these big perennial questions like what is truth, what is reality, what can we know. I'm curious if you have a sense, as we start to get into talking about AI stuff, what are those questions where large language models are going to give us a little bit of a new lens on some of those questions, or what are questions where we'll find new ones to ask that are better than previous ones even if they maybe don't answer them?
Well, historically, it's like questions that have led to a bunch of the various science disciplines. Everything from things in the physical world to things in the biological world like germ theory and all the rest. I think it's even true. It's one of the reasons why philosophy is the root discipline for many other disciplines. When you get to questions around how do you think about economics and game theory, or how do you think about political science and realpolitik and the conflict of nations and interests. It's also one of the reasons why, probably one of my deepest critiques of the non-reinvention of the university is the intensity of disciplinary-arianism. It's just the discipline of just political science or just the discipline of even philosophy as opposed to multidisciplinary. Part of the thing that I tend to think is kind of interesting is how much the academic disciplines tend to be more and more disciplinary versus, 'Hey, maybe every 25 years we should think about blowing them all up and reconstituting them in various ways.' That would be a better way of thinking, and why some of the most interesting people are the ones who are actually blending across disciplines within academia. I think that part of it is extremely important. Part of the question in philosophy is, 'How do we evolve the question of what do we know?' Obviously, you evolve the question through, for example, a lot of the history of science is instrumentation, new measurement devices that help with provisioning of theories. That's one of the reasons why people frequently don't think enough about how technology helps us change what is the definition to be human. We have this kind of Cartesian imagination that we are this pure thinking creature, and if we've learned anything, that's not really the way it works. That doesn't mean that we don't think that way to have abstractions to generate logic and theories of the world, but put your philosopher on some LSD and you'll get some different outputs.
That makes sense. So, along those lines, if I step back and squint, I can kind of divide the history of philosophy into essentialism and nominalism for a certain part of philosophy. Essentialists believe that there is a fundamental objective reality out there that's knowable and that there's a way to carve nature at its joints. Nominalists, which would include Wittgenstein, and pragmatists think that truth is more or less relative, or it's about social convention, or it's about what works. There's this ongoing debate between people who think one thing or the other. Do you think language models change or add any weight to either side of that debate?
I think they add perspective and color. I don't think they resolve the debate. There's certainly some question about since they function more like later Wittgenstein or more nominalist, you say, 'Well, does that weigh in on the side of nominalists because of the way they function?' But if you look at how we're trying to develop large language models, we're actually trying to get them to embody more essentialist characteristics. How do you ground in truth, have less hallucination? To gesture at a different earlier German philosopher, Hegel, one of the things I think is part of the human condition is thesis, antithesis, synthesis. You could say we have an essentialist thesis, we have a nominalist antithesis, and the synthesis is how we're putting them together in various ways. I don't even think later Wittgenstein would have said that the world is only language. The deconstructionists went to, 'It is only the veil of language and you have no contact with the world at all.' I think he would think that's kind of absurd. His point was to say that in how we live as forms of life, the way it operates is not a simple denotation. He understood it wasn't just 'the cat is on the mat' or the possibilities of the cat being on the mat, but actually possible configurations of the universe. There was this notion of logical possibility that was described as one language of possibility. To say that kind of essentialist about a language possibility is actually incorrect to how we discover truth and how we operationalize truth. You still have a robust theory of truth, which is not what the deconstructionists do. The robust theory of truth is partially grounded in this notion of language games and a biological form of life. Then you go into this deeply with saying, 'How is mathematics a language game?' as a classic language of truth. That's part of where you get what philosophers refer to as Kripkenstein, Saul Kripke's excellent lens on reading of what Wittgenstein was about. You apply all that to large language models. Language is this play out of this language game. Large language models are playing out this language game in various ways. But part of what is revealed is we don't just go, 'Truth is what is expressed in language.' Truth is a dynamic process. Human discourse, whether it's thesis, antithesis, synthesis or other things, is this human discourse coming out of this dialogic period, this truth discovery, this reasoning, whether it's induction, abduction, or deduction. These reasoning processes get us to what we think are theories of truth that are always to some degree works in progress.
That's really fascinating. I want to try to summarize that in case it was a little bit difficult to follow. There's a point in there that I think I missed something, so you tell me what I missed. I think one of the things I heard that was really interesting is when you think about how we built AI, which is predicting the next token, that's a very sort of late Wittgenstein-compatible idea, or pragmatic-compatible idea, where it's really about the relationship between different words in a sentence and we're not finding anything out about the world. There were other AI approaches in the 80s or 70s where it was literally 'let's list out every single object in the world,' and those didn't really work. That would be something along the lines of a more essential approach to AI. The one that works is a more pragmatic, more late Wittgensteinian one. But what's quite interesting is now that we have that pragmatic base that we've bootstrapped, we're in this process of trying to make it more grounded in reality, more reduced down to being able to talk about the essential ground truth. I think what's really interesting about Wittgenstein is he's sort of famous for saying, 'The limits of my language are the limits of my world.' I don't remember if that's late or early. But I think what you're saying is that Wittgenstein doesn't think that there's nothing outside of language, but he does think that the way we talk about the world, or the way that we use language, is part of this sort of social discourse where we're all kind of going back and forth to co-invent language and structures and language games together. You kind of see that happening with language models where when you do something like RLHF, that's sort of us playing with a language model, playing a language game to be like, 'No, you don't talk like that.' Is that generally what you're getting at?
Yes. Everything you said, but then the additional thing which later Wittgenstein was really trying to explore in various ways, because he wasn't trying to do a completely just social construction of truth. I'm actually a fan of you have to be a Wittgenstein scholar to understand how both early and late Wittgenstein are actually part of the same project. Late Wittgenstein wasn't 'early Wittgenstein was an idiot and now I've religiously converted to this different point of view.' There is a particular thing which is how do you get to the notion of understanding truth? Truth is that dynamic of discovery through language. It has to have some explicit external conditions. It isn't 'my truth, your truth.' There is only to some degree 'our truth' or 'the truth' in various ways. How do you get to that? In early Wittgenstein, the truth condition was it cashes out into a state of possibilities and actualities in this logical space of possibilities, which includes physical space as part of but broader than that. Later Wittgenstein said, 'Actually, this modeling of logical possibility is not the way this works. We're not grounding it that way. The way we're grounding it is in the notion of how we play language games, make moves in language.' The way that's grounded is to some degree sharing a certain biological form of life by which we recognize that's a valid move in the language game, this is not a valid move. This is what's interesting when it gets to large language models because you go, 'Large language models, are they the same biological form of life as us or are they different? How does that play out?' I think Wittgenstein would have found that question utterly fascinating and would have gone very deep on it trying to figure that out. By the way, the answer might be some and some, not 100% yes or 100% no. The argument in favor is large language models are trained on the corpus of human knowledge and language, and they're doing language patterns on that. Some might even argue that some of their patterns are very similar to the patterns of human learning and brains. Others would argue that it's not. But then you'd say it's also not a biological entity and it learns very differently than human beings learn. So maybe its language game, which looks like it's the human language game, is actually different in significant ways. Therefore, the truth functions are actually very different. In a sense, what we're trying to do when we are modifying and making progress with how we build these LLMs is to make them much more reliable on a truth basis. We love the creativity and the generativity, but we want it to have a better truth sense for a huge amount of the really useful cases in terms of amplifying humanity. The paradoxes in current GPT are when you can tease it out with very simple questions around prime numbers. You go, 'You got that answer wrong.' 'Oh yeah, I got it wrong, here's the answer.' 'Well, that answer is wrong too.' 'Oh, I got that one wrong too, here's the answer.' A human being would understand, 'I'm just getting these things wrong,' as opposed to, 'Oh, I'm sorry, you're right, I got it wrong, and here's another wrong answer.' We're trying to get that truth sense into it because we do have some notion of what's characteristic. Mathematics gets us in very pure definitions of certain kinds of language games. It's one of the reasons why centuries ago people thought math was maybe the language of the universe or the language of God, because there is the one where the purest truths, like 2 plus 2 equals 4, is kind of embedded. We're still working that out as we play with how we create these language tools. It's part of the reason why I think this question is really interesting because you can actually model it to some of the actual technological physics that we're trying to create when we're doing the next version. How do we get these things into good reasoning machines, not just good generativity machines? They have some reasoning from their generativity, but part of the classic showing where they break is showing where their reasoning stops working in ways that we value and aspire to as human beings in our best selves.
Here's something exhausting: the export-import dance. Let me know if this sounds familiar. You design something in Figma, you export it, you paste it somewhere else, and you pray that nothing breaks, and usually something does. It's almost 2026, which means that it's definitely time to stop doing this. Framer already built the fastest way to publish beautiful, production-ready websites, and now it's redefining how we design for the web. With the recent launch of Design Pages, a free canvas-based tool, Framer is more than a site builder. It's a true all-in-one design platform. From social assets to campaign visuals to vectors and icons, all the way to a live site, Framer is where ideas go live from start to finish. Framer's design tool is different from old school website builders you might see advertising on other podcasts. It offers vector editing, 3D transforms, gradients, animations, all for free. It has unlimited projects, unlimited pages, and unlimited collaborators. But what really changed how I think about Framer is that there's no handoff. What you design is the website. No developer interpretation, no 'can you make it match the mockup' conversation. You design it, you publish it, and it's live. Are you ready to design, iterate, and publish in one tool? Start creating for free at framer.com/design and use the code DAN for a month of Framer Pro. That's framer.com/design, promo code DAN. Rules and restrictions may apply. And now back to the episode.
That's really fascinating. You said a lot there. I really want to get into the reasoning thing in a second, but I want to go back to the way that you talked about late Wittgenstein versus early Wittgenstein because I haven't really heard it said that way. The usual thing people say is he just disagreed with everything when he was older. What I hear you saying now is in both cases, he's saying some of the same things, but the real difference is how he cashes out what it means for something to be true. In his first period, he's talking about truth in terms of a logical space of possibilities that can be broken down into these little atomic facts, and you can kind of build up truth from there, mapping those possibilities into actualities. In later Wittgenstein, it's all about these language games, the social relationships, the use of that word or phrase in the context of people. One of the things I really wanted to ask you about is that first version of Wittgenstein, where it's that logical space of possibilities. What that reminds me of is embeddings. Embeddings are one of the key underlying technologies that gave rise to AI. In traditional NLP, they allow you to represent words or tokens in a high-dimensional space. The language model innovation is that it's not just words, it's words in their particular context. Each word in a particular context has its own part of the space. In a language model, the word 'king' if it's tokenized that way, there's a king in chess, there's an actual king, there's a King of England, there's King Lear. They're all kind of kings, but they're in different spaces. Language models are able to represent all of those different meanings. That reminds me a lot of atomic facts or early Wittgenstein's early work. I'm just kind of curious because I think you said that language models, because of the next token prediction, are sort of late Wittgensteinian, but I wonder how you factor in the fact that embeddings work and are a core part of this.
Well, and actually this is part of the fact that late Wittgenstein is not 'early Wittgenstein was an idiot.' Because yes, I do think that the kind of the early Wittgenstein's logical space of possibilities is a very interesting model. Embeddings are a way of representing that logical space. But the key insight of late Wittgenstein is that the meaning is not just in the mapping of the logical space, but in the use, in the language game. So embeddings give us the space, but the language model's next token prediction is the game. They are both part of the same project, just like early and late Wittgenstein are part of the same project. The embeddings provide the logical space of possibilities, and the next token prediction is the language game that plays out in that space. So it's not one or the other, it's both working together.
The notion of, call it as it were, a probabilistic bet for what are the set of different tokens that apply are kind of there. Now, the reason why I would slant more as current practice late Wittgenstein than early Wittgenstein is because early Wittgenstein thought that once you had the grasp on the logic of it, you then almost by speaking correctly couldn't make truth mistakes because the logic was embedded in it. And even though the token embeddings are part of a very broad symbolic, quasi-symbolic network, the reason it's quasi-symbolic is because it's still activations and so forth and isn't purely the reasoning around a token of king or 15 different tokens of king or 23 different partial tokens of king, as much as there's conceptual spaces in that tokenization as mapped from a very large use of language. But part of language isn't just the historical language, but is the replication of it. Like if you say, 'This is the king of podcasts,' right? Or 'This is the king of microphones.'
Yes. Yes. But just as instances. That's part of why later Wittgenstein went to, well, it's how we're playing these language games and how we're reapplying them. And when we say, for example, on this podcast, 'This could become the king of podcasts,' we all have a sense of what we're doing. It's like, well, what would be the cases where that would be true and what would be the cases that be false? And what prediction is that making? And how is that a useful thing? I'm sure someone said 'king of podcast' before, but I've never heard it before, right? And it's a different tokenization, especially as it gets developed and elaborated in discussion. And then if you suddenly had another terabyte of information about discussions of kings and kingdoms and all the rest, all of a sudden that token space that it's learning from would change, and the generalizations off it would change. And that's part of the reason I would say it's more later Wittgenstein, even though not completely disconnected from those embeddings early. And it's one of the reasons why later Wittgenstein is not 'truth is just what language says.' No, there are ways in which it's embedded in the world by how we navigate as biological beings, and that's part of how the world comes and impacts it, and therefore it's not just language by itself, free-floating like the Cartesian consciousness, but it's embedded in some ways. And part of what he was trying to do is figure out, from a philosophy standpoint, how do we understand those embeddings and how do we drive our truth discourse in language based upon that biological embedding.
That makes sense. So I think what I hear you saying is, despite the fact that embeddings are in this sort of mapping words into this high-dimensional space which seems like mapping words into atomic facts or logical possibility space, the way that space is constructed and what makes something go into one part of the space or another is more late Wittgensteinian because it's very much about how it's used in practice and whether it's useful for humans in the world, rather than about some deep underlying logical ordering where if you've created that ordering, you can't say anything wrong because you're only using words from that space. Is that kind of on target?
Yes, exactly. And part of it is we know that there are truths where the coherent use of language still is a falsity. And so part of what we're trying to figure out is how do we get more of those truths and truth-telling and reasoning, because reasoning is about finding truth, into how these LLMs work.
