The Evolution of Computers & Abdication of Reasoning

Key Takeaway

Treat AI as a probabilistic collaborator, not an authority. Today, choose one task where an AI can accelerate research, drafting, pattern-finding, or prototyping, then retain ownership of the problem definition, source material, success criteria, and final verification. The episode’s central caution is that AI can produce useful answers even when the user has abdicated the underlying reasoning. Use it to expand your range, but write down what a correct answer must accomplish before you prompt it.

Episode Overview

The discussion traces computing from calculators, symbolic math, and deterministic software abstractions to today’s stochastic AI models. The speakers debate whether mathematical progress signals real-world utility, whether AI represents an entirely new layer of abstraction, and how capital-intensive model building is reshaping startup competition. They argue that AI may make previously engineering-limited problems capital-limited while raising new questions about human reasoning, verification, and concentrated power.

Full Transcript

Transcript of The Evolution of Computers & Abdication of Reasoning from A16Z. Auto-generated from episode audio; may contain minor errors.

Now, if I give 20 people a billion dollars, they can really put it to good use. It's as if we've shifted the industry from engineering-limited problems to capital-limited problems. These are fundamentally very different things. Mathematics is a leading indicator of what the market may be interested in. Why? Some people come along and say that the foundation of AGI and reasoning will be mathematics, but that says nothing about reality. To me, it's still within the realm of it being very good at playing the game. Startups don't target market leaders directly, and market leaders simply don't pay attention.

Microsoft is much more concerned about what Amazon and Google are doing than any of the startups. Everyone who works at a big company in Silicon Valley always thinks, "Oh my God, we're just going to crush all these little companies." And then you realize that they are never crushed. And I think that's why we're seeing such a rapid rise of Cursor, Anthropic, and OpenAI. Although capital is scarce, it's hard to get, and there are all these other things when you get it. First of all, thank you both for taking the time to come here.

This is great. Jared Sir tweeted something a few days ago about how he was asking Claude to try to solve the Riemann hypothesis and try harder. And I do n't know if there's actually been any progress, but it's part of a larger discussion that there seems to be some progress being made. How do we make sense of this in terms of what's actually happening, and what it means for mathematics? I'll give the floor to Steve. Oh, well, I'm not a mathematician at all , but I think this is an important point because it kind of divides the world into two groups: those who are really, really excited that, " Oh my god, these things are being solved." It doesn't matter whether you understand them.

In fact, no one understands numbers. The number of people who understand what this is is very small. And then there are people who say, "Oh, this is a fake, it will put people out of work, and no one will know where these areas are going." And the interesting thing here is that the group of people who are most excited about this are mostly mathematicians , and it's them who are really puzzling everyone. Because if you're the kind of person who says, "This is going to put people out of work, we're all going to get dumber," and that this is, you know, the beginning of idiocracy, because computers are doing all our work.

You might be surprised to learn that the people who are most affected by this level of AI are the most enthusiastic . Yes. Yes. And I think that in itself sheds light on the moment we are in right now . You know, you're talking to two systems guys, two product guys. You will hear that we will have the same reservations. I feel like there are things we both actually get along really well with. This is not one of them. So, as they say, I have two comments for my part.

So one of them is this: I think economic utility is a very important metric when it comes to AI, right? So I tried to think: many hours were spent trying to solve some math problem, right? But if you add up all the postdoc salaries of all the people who have worked on these problems for years, it probably isn't that much. And so part of me says: it's great that these opportunities exist, but I'm not sure that the fact that they've existed for a long time is a great indicator, since there hasn't been a huge economic incentive to address them.

That doesn't mean it's not hard or anything. I just don't think we have any evidence that this opens up any kind of economic value stream . And the second thing is, it does n't surprise me at all that AI does very well in an almost purely axiomatic field, which, you know, requires knowledge of many different things and a combination of solutions from very different fields. Because often when I read—I , like everyone else, obsessively read all this —they say, "Oh, I came up with a solution." It sounds like, "Yeah, the solution was pretty simple." It just borrowed some math that I didn't know.

So I think if there's any meta-conclusion here, it's that there 's a set of problems that are probably too broad for most people or most education systems to address, and it's going to address them. He's obviously very good at solving axiomatic systems , but I don't think that gives a strong indication of whether he's solving things that the market couldn't solve because there wasn't really a market around them. And so I think these are the best questions we should answer. This is very exciting. It seems quite reasonable and understandable.

Not sure what the long-term effects will be. I still think there's something interesting about mathematics being a leading indicator of what the market might be interested in. Why? Well, for example, I remember when I was studying, there was some big story about someone at AT&T inventing a new algorithm, a new program for linear algebra, a new way of solving linear equations, which is extremely important in the world of AI right now. But the main achievement was that we can now calculate the United Airlines flight map 3 hours faster than last week, right?

But let's dig deeper: I just don't understand whether the problems being solved are the obstacles that hinder existing economically useful tasks, right? And if they were, I don't understand why they would n't have been solved sooner. For example, a postdoc who spends five years pondering a problem, receiving 30 thousand a year. This is not at all the same as the market deciding that this is exactly what needs to be unlocked. Maybe they are there. Perhaps these problems, which are currently being solved, are key to opening some large economically productive case.

I just haven't seen it yet. So for me, that's the next thing I'm looking for. I don't even know what 12-dimensional space is or what it means. So I completely agree with you that I don't even understand what problems exist in 12- dimensional space. Are you very thin? Are you very tiny? I'm really confused by this. And maybe I'm wrong, but to me it's still in the realm of him just playing the game really well. Like, he's the best Starcraft player in history, which is cool and very powerful, but I'm having a hard time connecting that to the fact that a) maybe the reason we didn't have them before is simply a lack of economic need, and b) how does that actually translate into reality?

So, listen, there's a huge spectrum of these things. We are constantly being offered ideas. Some people come in and say, you know, the foundation for AI and reasoning is going to be mathematics. And once you do that, you can answer any question, because the universe is based on certain, you know, fundamental mathematical principles. And once you understand that, you will understand everything. And then, you know, there are other people who frankly walk in the door and just say, listen, this is great. Um, but that doesn't tell you anything about reality.

And so, you know, I think there's still a lot of work to do, and it's not just about knowing math better. Yes. I think what's interesting is that partly mathematicians are very excited about this because of the way they work. Let's say if you work in history, there is almost no abstraction in history. There's just a bunch of facts, and then people develop models that you can think of almost like force diagrams that explain wars, famines, or something like that. Whereas mathematicians and mathematics have this super long historical arc of layering abstractions on top of each other, and we'll get to, don't worry, we'll get to OSI in a minute.

But this idea of why they're so excited is that a bunch of math suddenly becomes a new level of abstraction. Is it true that they are so ...I found it ambiguous. I found that some were very excited, and some were in an existential crisis . Those who are excited , they're basically saying: listen, this solves 20% of my work , this is the 20% that I did n't like anyway. So it allows me to explore new horizons, which is very important or something. And I always asked myself: does it depend on the type of problem being solved?

I just can't imagine that if AI came along and cured cancer, someone working on cancer would say, "Oh, I'm so existentially depressed." This is amazing. But let me... on the other hand, like, "Oh, we solved this math problem, oh , I'm so depressed that I solved this problem"—maybe the whole point of that problem was to keep someone busy solving it. Or I was just writing articles backwards, as another attempt, and that's where I went wrong. So let me put it this way. There is nothing on the other side of the solution, and so we are depressed because now this useless pursuit is gone.

Give me , give me, let me stop. No, I mean that's too cynical. I love mathematics. I think we caught you being a little cynical, but the truth is, let's look at this from the perspective of the history of computer science. Because I had to take this course after studying the program catalogs of many educational institutions. You don't need to study it anymore. It was something like discrete mathematics. Yes. Yes. And, you know, algorithmic complexity theory, which was a required subject for a very long time, and now...

Remember " Concrete Mathematics " by Donald Knuth? I don't, well, you're a Stanford guy. I don't, but my state university didn't have that , it's a joke to us Cornell graduates. But, listen, my course was taught by one of the luminaries in the field of algorithms, ironically, a PhD from Stanford, John Hopcroft. Of course. Who, for practitioners, invented 2-3-trees and a bunch of other things as his dissertation at Stanford. This is a legend . But John was our professor for all this nonsense, and we taught all this stuff about P equals NP, and I remember how this four- color theorem...

Proved by computers. No, but that's exactly what I'm leading up to . You just covered up the main point. Yes, but for those who don't know: we had to take an entire university course that was all about this problem. And I wonder why: theorists postulated that if you could solve this problem in polynomial time, rather than exponential time, you could solve many others much faster, such as the traveling salesman problem, which was important because our computers were resource-constrained. So, if you worked at AT&T and had a node with 6,000 switches, how would you find the optimal route?

You would say, "We don't have enough capacity." That's like two years of continuous simulation to solve. Yes. Yes. And it turned out that one of the interesting moments was the proof of the four- color theorem. Yes. But they did it... by the way, the four-color theorem states that any 2D map can be colored. You can only use four colors. So that no two neighboring regions have the same color. Yes. That's right. And, um, and you only need four colors. You will never need five colors. And we studied it— just so you kids know—that's literally how we studied it.

And we could all repeat it like this. This is a very strange print on this problem. And what happened was that no one ever got to what could be considered a proof in the style of calculus. Instead, they proved that the number of potential solutions is finite. Have you seen this proof? Yes. Yes. 200 pages of combinations, but they proved that there were a finite number of them, then they just calculated all of them and said, "Look, there are only four colors." So this is a kind of " indirect" proof, but it was only possible thanks to calculations.

And as you said, it was actually very, very useful for practical applications. Yes. Yes. And certainly, as a topology specialist, for example, setting hard boundaries and things like that. So actually I see and I think that was a really good lesson for me of how when you have a new level of abstraction that says this is a whole class of problems that can be solved. Yes. You can build tools working at this level of abstraction, and nobody needs to start from scratch, like: okay, what is the representation of 2-3 trees in what we do.

So listen, it's hard not to get philosophical when you're talking about AI. So, I'll philosophize, and you can tell me to shut up, but I just can't...you understand. So this math thing seems a little different to me because it raises the following question: can math ever be a representation of a physical phenomenon? Has anyone ever taken a set of equations and actually predicted something physical? And I don't know the answer to that. I've worked with these big simulation codes, and they're actually trying to calculate physical phenomena, like a star exploding or what would happen to an airplane in a flight simulator or wind.

But all of them , although they calculate these large differential equations, were based on empirical results. Yes. literally the equation of state for what, they were modeling, they were just saying that we can measure the temperature in these places like this , all of this was based on empirical equations of state, and I always wondered, for example, is the simulation computationally unintegrated, and therefore should it really be run? In that case, it's unclear to me to what extent AI helps here . I know people are trying to solve this problem with AI, but I'm not sure if these mathematical answers have any impact on things like this, do they?

Yes. So maybe there's some particular area of algorithms, as you say, where they have an impact, or let's say modeling or logistics. But when it comes to, for example, "will this star explode?" "will this building stand?" As a simulation itself, I think these things are quite far apart . And I read a lot of discussions about mathematical solutions where the claim is that if you can solve all of mathematics, then you can predict anything , and I think that's a huge logical leap that, to me, is not at all obvious .

Yes. Or there is no indication at all that it is true. like that. So the way I think I could, um , talk about this is, you know, again, it's not my level of math itself for people working with systems, I'm not bad, but I have to, I'm a tool person by nature, so I understand this part, which is that maybe AI won't be the next tool for solving mathematical problems on a large scale, but it could lead to the development of a new kind, a new level of model, and so I brought the props to demonstrate that.

So, of course, this is an original tool for mathematics. And so, before anything like this came along, it's, you know, one of these real ones, like in the market in Beijing. Well, I know, that's what they tell tourists in French . But I'm very proud of it because I got it for about seven cents. But, um, you know, it became a level of abstraction, and suddenly you had this basic mathematical thing, and then you just jump over a big period of time. I brought this because it's just incredibly cool.

That's right. So, everyone knows what a slide rule is . But no one knows how to use them. Yes. It's called a " kurta," it's an Austrian, well, it's actually a circular slide rule . Yes. And it's like a coffee grinder or a pepper mill, you have all these ways to set the numbers on the side, and then you turn one way to add, the other way to subtract. Oho. And this thing is... Wait, is this used for large-number arithmetic or something like logarithms? No, just arithmetic.

Of course. Well, I think so, but of course it depends on how you use it. But, but... this is probably the middle of the 20th century, I think. And my uncle brought this back from the war. And, you know, what's incredible is that there are about 600 machined metal parts. Wow. Inside this. Today, making one like this would cost about $50,000. Now, do you know how to use it? I actually knew, but I wo n't try to do it. I'm actually, preparing for this so I don't look like a complete fool...

Just look what I have. I really took the trouble to learn how to use it, even though it had been sitting on my shelf for years. But what's interesting is that, you know, suddenly a whole new level of problems is being solved. And... wait, so you mean the new model is a new calculator or a new graphing calculator, I really remember, for example, remember the TI-85? Oh, of course. Yes. Yes. I remember when he came out. They are like this: " All math teachers had a crisis." I'm like, " You know, we used to give you a piece of paper.

We'd do xy equations. Now they can do it on a calculator and they can solve equations, and our industry is dead." But what's interesting is that this is exactly why it's so important for artificial intelligence today. Those people didn't complain when calculus came along because calculus was basic to them . And the thing is, there's this perception that people respond to change more than they do to the starting point they started from. And that's why there's so much anxiety...I literally lived, I had a TI-35, those were the first calculators in schools.

The only higher math they did was the percent key and the factorial, which we didn't even know what it was, and you could calculate, say, 59 factorial, and that was the maximum that could be displayed. You know, I went to college, and calculators were banned in class. All my life, I've actually been on the border between what 's permitted and what's forbidden for everyone. I mean, I saw the advent of graphing calculators. You literally had a blue notebook. Yes. Just to show the entire work process, to prove that you're not just entering data into a graph.

You see, I skipped the graphics stage, even though most of us in the back office were writing video games and didn't care at all about their ability to calculate math. But. Certainly. Absolutely. And if you rewind all of this, you realize that after all these stages you go from algebra, linear algebra, mathematical analysis, and then you get to Fourier transforms and hydrodynamics, and all of this, as you mentioned earlier, was based on need. I mean, a lot of this is math. Well, all computers come from difference machines that simply tried to calculate integrals.

But, of course, we need to clearly understand: we calculated integrals so that we could shoot missiles and cannons at each other. That's it, yes. I'm not condemning this, I'm just stating a fact. Well, yes, I don't want to be pedantic, but this is one of my favorite parts of the story. In fact, it all started with tides, which also had enormous economic value. You're trying to calculate the tides , and here old architectures appeared, which were then, of course, involved in military efforts to calculate algorithms.

That's where I got it from; It was actually very interesting. It was about 5,000 times faster than a human when it came to, you know, these kinds of calculations, of course. And it made no mistakes , which was to some extent . But this was very specific mathematics aimed at a specific economic benefit. And the interesting question for me is that these models obviously do a good job of doing a certain type of math. Is this something that has somehow slowed down a certain type of economic development?

Yes. And I don't know the answer to this question. Oh yes. I, I, you know, I think it's really interesting to continue this topic, because I think the coolest thing is that doing those basic calculations for war, creating tables for missiles and stuff, then essentially gave rise to the space race, jet engines, factory automation and all that. And you know, people were rooting for it, and for me that's the most interesting cultural aspect. That is, not only did they cheer, each parent looked at their children and said: " Go study this at school." Win the Westinghouse contest.

"Win the GE Math Contest." And was it because of the Cold War? Or because Well, of course, the Cold War was a big part of the culture, no doubt, but it was just a general aspiration for the future. I found this incredibly cool brochure from IBM, it's from 1953. So in 1953. Are you just keeping this at home? I just happened to come across it. This is the one I just received. I do n't even believe that such a thing exists. But this is a brochure about the future being calculated.

Wait, I want to see. But first you need to look. There's even something about nuclear energy, well, all that stuff. The future of computing looks like a guy with atoms spinning around his head. Oh , wait, we're rolling in with the camera like Carol Merrell. So, the most interesting thing is that it's from 1953. Then there was ENIAC, and at that point that was it. It was a computer of that time. This was before the 704th and 370th models appeared. This is a brochure from IBM that explains what a computer can be.

Not even what is, but what could be. It says: "It took millions of years to invent the wheel and recognize its usefulness." This is an introductory sentence, and people then just swallowed this information. But here's the part I want to get to. It's about computers and their two families. Yes. And , of course, there is a mention of a slide rule , which explains the story. And what this really leads to is that we can do this for text as well. Yes. The idea is, imagine who was reading this in 1953 , if you had to explain hexadecimal, decimal, binary systems and compare them to Roman numerals.

This is amazing. And all because no one, no one knew. That's the name of this thing. It's simply called "IBM Light of the Future" with a sort of spotlight shaped like a rocket test tube. And that's incredible. There are like oscilloscope waves in the background . This is the most incredible thing. There's this dictionary in the back. Imagine someone explaining a computer for the first time, and the dictionary is, you know, arithmetic device, binary bit, bit, well, let's say, cathode ray tube, electrostatic storage tube, and you know, but the thing is, I opened this because there's one cool page that's really important.

How are digital computers organized? And this is what brings us to the issue of abstraction for us and AI. This is how we've imagined organizing computers for 75 years. Yes. Input, storage, arithmetic, control, and output. Yes. And that's all we learned in school. You basically took courses in each of these areas. Last night we discussed the abstractions that will remain in computer science. And you added networks, which are a kind of control. Yes. By the way, everyone forgets about networks. Of course, it was well, because most people stop caring about networks, stop worrying about them, as soon as the packet leaves the computer, I would say, you know, the late 90s—that was the end of the mandatory networking course.

Because, well, it seemed like it was decided. For example, for me it was a transistor. I remember the last time computer science students needed to know what a transistor was, and believe me, I don't really understand what it is now, it's just a triangle. But the interesting thing is that these abstractions led to...okay, now we have these abstractions. These were actually industries that dealt with each of them, for example, you spent 20 years of your career in data storage, watching the transition from tubes to drums, spinning disks, tapes, and so on.

And you know, if you've been involved in output , you've seen the evolution from teletype to line type, to black and white terminal, to color, to vector, and so on . And all these industries developed in parallel. Any computer science department that grew out of a math department through rocketry eventually became departments made up of these things, and then all of that came together and gave birth to us, the systems group. Of course. Yes. Yes. Yes. So let me elaborate on that aspect a little bit.

Of course. Of course. Because I, listen , I definitely like this approach, and we move to abstraction, and behind every abstraction there are still a number of problems. It's just a higher level of abstraction. But I still think the concept of economic sense is very important. Oh yes. Yes. That's right. For example, Bletchley Park was engaged in breaking codes for war, and this effort gave rise to the innovation that resulted in victory in World War II. Or, we were trying to do nuclear development, not just research, but innovation in the context of the war effort, we needed to calculate integrals, and we were doing it manually, and so at that point.

This was perceived as the salvation of humanity. Everyone was incredibly enthusiastic. All physicists fell in love with computers and started using them. And for me, the problem with the current way of solving math problems is that I don't know what's on the other side. Oh yes. No, but the other side of it is, I really think we've had this in the past. And that's why, you know. Well, we had a moment with AlphaGo. It was the same. We were recording a podcast, not in this room, but.

And before AlphaGo, remember when chess was defeated at chess. Yeah, we had that IBM chess thing, and Frank Chen and I were recording a podcast about AlphaGo, and we had to explain to people why it was a good idea, and. And I think it's quite appropriate to ask: there are things that these technologies solve, and there's a lot of benefit and value in that, and that will move everything forward. And when that happens, people are usually happy and supportive of it, but there are things that are being resolved , and I think people are looking at them.

not so positive, and I would assume that it's because it's as if the very solution to the problem has become an end in itself, rather than a real result. But maybe we should all take a step back and say: if you're truly upset about something you decided, maybe you shouldn't have worked on it in the first place . Is n't that right? Sure, you know, it's like creating a sand mandala for inner peace or something like that, but it doesn't move the economy forward. Is n't that right?

Well, we're both systems people, but I'm really an applications person. I know you're not...I'm not a system. Yes, we are both systems engineers, but as a system... I am, of course, absolutely sure that the main wave is applications and, of course, the Internet. The same problem was in 1995-96 with the Internet: it was very exciting, but most people just sat there and said, "I don't understand what's in it for me." Listen, there's a great book out now called "Steve Jobs in Exile," which I consider a must- read if you're listening to this podcast.

Kane wrote this book with Catmull from Pixar and Dan Levine, who was a very close friend of Steve and also worked at Microsoft. They ...this book is just fantastic because it explains and embodies this whole idea of ​​creating things that people actually need and that solve their problems. But it clearly indicates that Next was the same machine that Tim Berners-Lee used to write the HTTP protocol, right? So he's actually... the Next machine... and he used the Next machine. It's interesting... and it's super interesting because no one knew what this machine was for or what it did.

But then he built it, and still no one knew what the machine was for or what it did, because he said, "Well, it's to look up other researchers' phone numbers and share articles." And I'm like, "Hmm," and I think there was a great example of a company out of Seattle called Cyber Pizza, which was a dot-com thing that didn't even seem to make it to 2000. But the idea was that it was essentially Uber or DoorDash for pizza, just for pizza. You placed an order, and they found the nearest pizzeria and sent you a pizza.

This was a demo show at the Next presentation on stage. They did it, and they even had real pizzas backstage in case nothing worked. And I have to say, for NextStep or OpenStep. But the idea was that it showed what could be done with it. And the reaction was literally, "Wow, that's really cool, but have you heard of the phone?" Yes. Yes. Your argument is that we don't know how to use everything. So I'm a little bit focused on the fact that on the other side there was a solution that people were looking for.

You say there are a lot of platforms being built where it's not obvious, but it's clear that they're... well, spreadsheets were like... I'll show you probably one of my last visual aids for today. But word processors came along, and this was in 1982, and people were using them on Apple II computers and on this new type of computer called CP/M, which is the ancestor of DOS. And people said, "I don't understand why just print." And once people started using computers, the idea of typing on a typewriter just didn't work anymore.

And so, some people in law school... Well, show me. Yes, I'll show you. I just love to build intrigue. But these people at Harvard Law School brought in the first, um, laptop. So, this is the first laptop. Aren't they called " portable"? Well, no. They...it was just called, it was literally called Osborne, and he was the only one. So, as an assumption, Eric, you're a child. How long did the battery last in this? Ahem. Not for long. There was no battery. This giant suitcase weighs 25 pounds, and it doesn't have a battery.

It just plugged into the outlet. But that was a trick question, because every time I turned mine off—I still had mine from college. People ask, “Well, how long does the battery last?” "—and it's literally the size of a sewing machine." It is larger than any permitted carry-on baggage. And this was my computer in college. But in his senior year of high school, he was banned from Harvard Law School. Someone came to take exams. At Harvard, it was common to bring a typewriter to exams so that the professor could read the text.

And two students brought computers. One brought an Apple II, the other an Osborne, and the school banned them. Oho. They just said no, for all the reasons you could read about, I have articles from Time and the New York Times. Every article you read is like: don't use a graphing calculator, don't listen to rap music or don't listen to...I don't know, jazz, or don't play Dungeons and Dragons, or whatever the argument is that's going on right now. Three years ago, I tried to convince Cornell to use AI in their freshman courses, and they just stopped talking to me .

Oho. And that's the irony of it. In my freshman year, when I got this computer, I was, of course, the only one in my 90-person dorm with a computer, and I had to ask the dean for permission to use it to write freshman English papers. It was the fall of 1983. And , in fact, this is exactly the situation we are in now with all of this. And this...this whole thing...this...you could also think of it as a level of abstraction, because, for example, nobody goes to college now without a computer , like...can I, can I, can I...

just...no, I agree with you, but let me. True. True. From time to time I think: maybe things are a little different. So here's the argument. Ahem. I don't think I can remember in the history of computer science a time when we've ever given up on actual reasoning or logic. It's always been a resource, right? It was like a computing network and storage, and it's something that you provide, and then the human inputs the basic data, and it uses the computing network and storage to calculate the answer.

But all that initial training that we provide -- maybe it's not about the internet, but right now it seems to me that you're actually giving up on thinking, like, " give me the answer, " without even being sure what the actual question is. That's right . And again, I think you could say that Google was something similar too. But it was still a social phenomenon, and not quite the same. So it feels a bit different from just moving to abstractions. Because when you move to abstractions, you still usually have a deterministic system, which is a higher level of abstraction, where the human determines everything related to the task statement.

It feels a little different. Well, it definitely feels different. Here's what else I think: graphing calculators felt different to me. I thought graphing calculators were a scam. And because you know, the exam task was to build a graph. And here's what's happening now: the opportunities are matching the exam task. Now, going back to what we were talking about— computers and mathematicians. In my freshman year, there was a new product, a new thing— Maxima, which was a symbolic mathematical calculation package from MIT, and you could literally type, for example, an integral into the computer.

I remember when I first saw Mathematica, I thought: this is some kind of magic. Maxima—it's, you know, ah, machine-assisted , what was that? Machine- supported computing, symbolic mathematics , I think, and this was a lab at MIT, started in the late '60s, early '70s, and it started to spread. So in my freshman engineering class, we had a version that ran on an IBM PC. It was called MUMath, and when we got a math homework assignment , we would go to the engineering library, get a diskette with the program, and just type in the answers so you don't think it was cheating .

Let me digress a bit, because I tend to agree, but sometimes I have moments of doubt. So, ah , I don't remember writing programs where you actually give up logic, like, if I'm writing a program, I'll, for example, use a cloud database, I'll use storage. I will use the network. You know, whatever you want, but the correctness and logic of the program remain under the control of the programmer. Maybe I'll use a third-party library. But then again, it's me who chooses the library. I know the input data.

I know the source data. And it seems to me that we are entering an area where you are actually translating the logic to a third party. It's as if you're saying: give me an answer. So maybe it's just a higher level of abstraction. It seems a little different to me. No, actually, that's what the discussion is about. I, like, I fully support this discussion. Here's an example of this, an example from Stanford. So, during the "AI winter " that was in the '80s , Stanford had one of the biggest "AI winters," and we have a podcast about it from 15 years ago .

One of the biggest trends at Stanford has been the combination of new AI with the medical school. And so there were many projects, like medical diagnostics, chemotherapy, and so on. I worked on one that was doing organic synthesis with a team at Harvard , and they were all kind of first attempts, " let me hand over the decision-making." Actually, this whole '80s era in computers was the dawn of what used to be called expert systems. I remember and so expert systems were the first time we got a taste of this discussion.

I remember well. I just your classes were about this it just didn't work. Your classes were mostly about how you had a lot of lessons about this. a bunch of expert systems. I had to build expert systems, right? I have written a lot in Prolog . That's right. So I understand that very well. I just thought it never really worked and OF COURSE. So the big difference is that it used to work, but it works now, and we're giving up logic, using these but it's interesting because you compare and contrast and even in the case of Prolog, you're like you're programming it for them.

It's algorithmic. you're still specifying the end state, and it's just a way to reach it, whereas here you're almost asking what the end state should be. So it feels a little different and it feels special, so I agree, I love having this discussion because I think a lot of it comes down to the anxiety and the worry that you feel when you think about it. It's really because of the context that we're in. And, you know, think about the fact that we have all this stuff going on where people don't want to build data centers, but two years ago people were kind of fighting each other...

please, governors were competing to build data centers, or, you know, 10 years ago, like, "Build a car factory in our state," one that smokes and requires hard labor. And that's why context is really important to these discussions; you cannot separate them from it. yeah, but I just want to get back to that, yeah. And I don't mean, I just, I just think like, um, my whole career has been about moving up the stack, but there's always been a computer stack . Yes. You could always map it to the next level in a fairly deterministic way.

There are higher levels of computational abstractions. This is the first time it feels like another layer of the stack. Perhaps this is indeed the next abstraction, which is more like a human level of abstraction, which is not directly reflected , and therefore actually different. So it could, I think, you know, whatever it is, starting with transistor logic, then to compute, then to hardware, then to OS, then to applications, and then to platforms, you moved up the stack like that . It may be that we're at a level where we have to rethink the basics , because to me it feels very different than just the next level.

The big difference is—and we can argue, debate, or call it whatever we want—that we're actually moving from computation to imperative programming. Yes, that's where we were and where everyone else is now. Then there was a brief period where we were in a mode where the data actually drove the program, and that was the first realization: all these findings and stuff. And now we are at a point where it is all arbitrary, random, and statistical. So, I think of it this way: imperative programming is when you know all the steps, you write a recipe, and it executes those steps.

Okay, and then there's declarative programming, where you know the final state. It's like Prolog for humans. Either Datalog, or SQL, where you just know the final state. But then the computer does all the work to get to that state, and you can't limit how long it runs, so it's like makefiles: you tell it what the final result should look like, and it does it. And this is something completely new, where it's almost like... you don't know exactly what the end state is going to be, and you just, you know, pray to the " Model God" with the right words, and he gives you an answer that turns out to be useful .

Yes. Yes. And but I look, and it's stochastic. This is a factual statement. But it's also interesting to think about it in the future: is this the next level of abstraction in our understanding of computing? Yes. Maybe it's like a computational baby, where the computer and natural phenomena overlap a lot , because the response is formed based on a human product—a language that's a little different from... If we need to rethink some fundamental assumptions, what might that look like? Well, I just think that people like Stephen and I have developed a deep intuition about how systems work and how they impact the industry, having watched it for 40-50 years, and I just do n't know, like, is the value going to go to the model or to the application?

How much capital can be invested in these things? What classes of problems can be solved and which cannot? erm, what guarantees can you provide, er, how does this affect performance? There are many things that we have an intuitive understanding of , and for me the main question is: should we change these assumptions or not? And to what extent should we do this, since the laws of physics feel a little different? I will give you just one example. I mean, I've said this many times, I think it's very important: 20 years ago, if you were a startup of 10 people and I gave you a billion dollars, what would you do with it?

You would end up spending a lot of money on building, buying your own computers and so on , if you hired people and bought equipment, you would go broke. You just wouldn't know what to do with a billion dollars. O. Oh, I see what you're saying. Yes. Yes. Yes. 10 years ago, I'll give you a billion. You hire, you hire engineers, and you're done. Yes. Yes. Yes. Well, what to do? Well, that's right, writing code . You need, you know , you have, you know, this billion.

The important part of this is the billion. It's not that you have money, it's that it's a huge billion. That's a lot of money. If I gave you a billion dollars two years ago, because 10 million, you would buy a bunch of things from Hewlett-Packard and the money would run out. And exactly, exactly. This is a billion dollars. I mean, like in software: you hire people, and then it all comes down to scaling. The mythical man is quite real. Yes. Now, if I give 20 people a billion dollars, they can actually put it to good use.

It's very...it's like we've moved the industry from an engineering-limited problem to a capital problem, which is fundamentally very different. We have never been in this situation before. And so it's like a law of physics where our early intuition that all problems are engineering problems starts to change. So I think there's an open question that we, especially people like us, should be asking: To what extent should we reevaluate our ideas about this? And this is not just one level of abstraction, it actually changes the nature of capital relative to innovation, competition, security, etc.

I think that's a great, great way to think about it, because it forces you to think about a new model. It's also interesting that for the first 30 or 40 years, computing was dependent on capital. I mean, if you wanted to do something, this is such an important point, if you wanted to do something on a computer, the first step was to get it, and you couldn't. You depended on capital, then on engineering, and now we depend on capital again, which is just crazy. So, it's like you're going back 40 years.

Yes. There's a scene in "Mad Men" where a computer appears in an advertising agency and everyone runs around trying to figure out what it's for, just like they did with the copier earlier . But it was interesting because they didn't know what to do with it , but they were glad they had the capital to buy it, and it made it look like they knew what they were doing. Five years ago, Patrick Collison interviewed Sam Altman on a podcast, and Patrick said, "We were living in an era of frugal startups, but for your projects like OpenAI or energy projects, you raised enormous amounts of money." You raised huge amounts of money at the start , is this an underrated approach?

This just resonates with what you're saying. Yes. Yes. You know, that's interesting. Before the AI ​​era, there was a debate between Eric Ries and Ben Horowitz, right? Yes. Yes. One for the “frugal startup,” and Mark and Ben wrote about the “ art of the fat startup,” where they argued: raise money and act. But there has always been a natural limiter —engineering. Yes. Complexity is what was reality. So Patrick Collison is right: we now have the discipline of how to use a lot of money effectively with small teams.

This is a very big change. I don't think we've realized that yet. Which is also incredible, and that's why there's so much optimism right now. Because even though capital is scarce and hard to get, once you get it... as we know, building on people is also hard, scaling it, and nine people can't do anything faster, and I'm telling you, yeah, my 10- year work, you know. I was hired and I literally gave these early teams money, helped them with hiring, and then waited two years for the engineering to just scale.

This also has implications for venture capital, because for the last ten years people have been saying, "Hey, there's too much capital, too much capital." I think that's just a crazy view. There's this zero-sum game mindset in the venture world, which is quite funny to people who shouldn't be thinking like that, you know, and they'll talk about... Too much capital chasing too few deals, and it all looks like you're a venture capitalist, don't you believe in positive sum, huh? But if you look at the numbers, the more capital that flows into private markets, the bigger the market itself becomes .

And there are, there are, there are several reasons for this. One of them is the one we were talking about: tech waves that are really capable of absorbing capital, like AI. But there is another: if more capital is available in private markets , companies will stay private longer, so more value accumulates on the private side. So I think the capital that goes into private markets increases the total market size (TAM). This is not a limited market. And I think people who would be —it's so funny— early- stage venture capitalists who have to think in terms of positive-sum outcomes have to stop thinking in terms of zero-sum.

Well, one way to think about it—I'll go back to what I think your foundation is doing that's most exciting: we're really on the cusp of a wave of applications. And the fact that you can now apply capital without having been a recruiter for 10 years and have a result for the whole world that is not covered by software, and that's literally everything. Like everyone who complains about medical records, doctor's appointments, or— my favorite example—lawyers. No one was more excited that "Oh my god, we can finally automate lawyers with AI," which is the strangest thing in a world where everyone is against everything except having more lawyers.

And all of this means that someone who has industry experience, as we used to love in venture capital, is like, "Oh, you know, it turns out it's really, really hard to build commercial real estate." “Wouldn’t it be great if someone who knew commercial real estate started a software company but didn’t know how to build software. Well, they should find a co-founder who knows how to build software and teach him 20 years of experience in commercial reality. This is really difficult. But now the path away from such an idea is a question of capital, and this is a new level of abstraction.

I mean, I remember my very first client visit as a professional product developer: I visited a doctor who had graduated from medical school after studying early computer science. Oho. And he wrote something like a DOS program for scheduling the work of a doctor's office. This is amazing. You think it's just planning. This is a calendar with hours. But it turned out that it was me, 20-year-old me, listening to this guy explain, “No, you don’t understand. “You call the doctor and, you know, you talk to the dispatcher.

So they listen for keywords to decide whether it's 5 minutes or 20 minutes, do they need X-ray machines, do they need an EKG? And they actually schedule in parallel, like blood donations and everything else, not just the 10 minutes that you need with the doctor. And that's exactly what his software did. It took him years to create it himself. Yes. And and this is what you just said but this is exactly what can, how can now this problem be solved by a person who knows code, no-code is finally here.

Well, that may indeed be the case, and maybe you're not just creating this one-off code that's hard for you, but everyone else's level of abstraction is also increasing. So you don't need to design that part of the code. For example, if you're doing this for a phone, the phone's level of abstraction has increased, so you're not creating a text control; you are not creating interface controls. Whereas 20 years ago the first step in starting a company was to create all of these things, so it's important to realize how much this impacts what you can do.

I want to talk about any other fundamental assumptions that would be interesting to revisit, like what about market leaders versus startups, we talked a lot about the innovator's dilemma. Does that mean that now that these startups are market leaders, or they have the capital advantage, are they able to do more, but at the same time we see startups that you would think the market leaders would just destroy . The craziest thing is, if you had told me six months ago that you would ask this question, what advantages do market leaders have?

They have the same benefits as always. They have capital, cash flow, and, well, that sort of thing. Distribution, distribution, and you know what's crazy, AI solves the distribution problem. It simply solves the demand problem, and B, these companies are able to raise so much money that they are actually on par with giants like Microsoft. So I think we're in completely new territory when it comes to these new challenges for market leaders, for these two reasons. You know, I think the distribution aspect is often underestimated in how important it is.

In the past, if you had a company and you wanted people to use your product. It was difficult. You would hire marketers. You have no idea how much to invest and where, and you don't know what return on investment you will get . But the demand for tokens and GPUs is so unlimited. You can literally just decide how much money to invest to drive growth at the entrance of the funnel. And so things that used to be very, very difficult for startups are now much easier.

And I think that's why we're seeing such rapid growth of Cursor, Anthropic, and OpenAI, which provides access to capital and puts them in a favorable position. So very much. I do, and I think this is to your point about how difficult it is: listen, my whole life has been managing thousands of engineers to build things that couldn't be built anywhere else. That was a real barrier— to create an operating system. It was endless, and I think so too. You must have one leader. Is this the one?

Well, this, but this is really, read, read. This was the main thing. Yes, I know. No, I understand. No, but, but he's brilliant. But read the book about Steve Jobs's ouster, because there you really get a feel for what it's like to build from scratch. In fact, you know, of course, NeXT is famous for just taking the code from Mach at Carnegie Mellon and starting with it. We wouldn't be able to do it completely from scratch . But this, this whole idea of how important it is to think about the specifics of the domain and how you challenge others.

Because at Harvard Business School, they used to joke when Clay was with us that it was strange to teach subversion as a theory in the business school, when it should simply be a fact in the physics department. And I really like that I was there in '98 when he was writing the book, the article and everything else. That's when I was teaching. This is nice. Perfectly. And I really, I used to be, of course, there's a belief among everyone who comes from a big company to Silicon Valley, or when you come like me, the theory is always this: you think, "Oh my God, we're just going to crush all these small companies." You always think that way when you're in a big company, and then you realize they never get crushed.

Ben always emphasizes this, as does Mark in his, um... his film, it was AWS that actually forced the shutdown. Exactly, exactly, and because, you know, startups don't target market leaders directly , and they simply don't pay attention. Market leaders are only interested in what other leaders are doing . Microsoft is much more worried about what Amazon and Google are doing than about any of the startups. But the elements of disruption that matter are the cultural characteristics of a large company, and they are immutable. These are the laws of physics, and you simply cannot change them.

You cannot change performance metrics. You can't change direct sales, go-to-market strategy, compensation, organizational structures, legacy, and customers because, you know, the way you operate, if you serve, say, 500,000 customers... there's a bunch of things you just can't do, you're kind of stuck. And that's the real essence of disruption, and that's why we're at a magical moment right now, where it's not just about the culture staying the same, but also about the startup ecosystem. This is a reflection of what happened during the development of cloud technologies, when there were a whole bunch of necessary things.

Again, this is a layer of abstraction. You know, if you're a startup like you, you don't have to build a data center, create your own communication channels, call AT&T and do all that. Now you can get started in the early hours of your first dinner. But with the cloud, I actually think that... Yes, you put it very well . And I actually use it because it's great. It was like this with the cloud: no one believed they could displace AWS from the market. That's right. Of course.

You simply accepted this oligopoly and built on it . And the question is, will they destroy us in our little corner? The answer was either they would just add you for free, or for a price that could compete. I really think, you know, and that's always been the main question. Will Microsoft do this? Right. But now these companies are really challenging the market leaders. And you rightly pointed out —I hadn't even thought of it that way—that success used to be determined by complex engineering efforts: creating chips, systems.

Are you talking about the " New Machine Era" or something? Yes. Yes. Yes. "Solaris ", well... the solution is a solar car. Yes. A wonderful book. She talked about how difficult it was to build such systems, even operating systems or cloud architectures. Jeff Dean and his generation were building distributed clusters , and they were the first to learn how to do it. Once this was achieved, a huge advantage emerged. These were large-scale engineering projects beyond the capabilities of startups, and now such models only require capital.

So different laws of physics apply here: if you have capital, you can achieve a lot . Look: Google is Google, they have all the data, all the intelligence, but their models lose to OpenAI and Anthropic, and it goes on and on . Because it is a cultural element. I think people underestimate it on the outside until they themselves encounter the cultural barrier of trying to do something. I'm sure it's all about culture. This is not a typical engineering task, which they are very strong at, because they do a great job of executing, for example, GCP—it's just fantastic.

It's an engineering feat, but I also think it's hard for them to allocate that much capital to one of these companies. Well, for all the big companies—you can see from their income statements how they worry about this capital. Here's Google doing its bond deals to actually take this off the balance sheet in some tricky way. There were rumors — I don't remember which company — that they were limiting the use of tokens for internal products, preferring corporate clients, so internal products were lacking AI. And, of course, none of these products' competitors are lacking.

And I ... I just learned to really appreciate... listen, I fought and fought to keep us from being supplanted by mobile platforms, essentially ARM, and Intel just didn't care. You know, I came here, sat down at a table across from the entire Intel leadership, took out the first Surface and said, "Here's our new computer." They were very excited, and then asked: " What's inside?" I replied, "Well, it's an ARM chip." Ovva. And you know, the fact that I even brought such a thing into the building...it was very, very difficult.

They just never thought it was anything serious, they thought it was a chip for a printer. And they looked at me and said, "We...we are ARM experts." We knew everything, and I say: "But it's all about energy efficiency, graphics, constant connectivity and everything else." Their culture at Intel was to follow Moore's Law. And at Google, for example, they focus on hyperscaling. Yes. So if AI moves to devices... Yes. Yes. Of course. This is simply not their profile. Yes. Of course. As for Microsoft, they were squeezed.

They're being squeezed right now, and I think you've raised some really interesting points about the opportunities for... startups... with this capital inversion process. raise capital and get down to... Well, it's not just that, you're saying, "We won't question your attempts to raise capital, we won't look at you like you're crazy." And if you look at the investments that are happening now, these companies have achieved considerable success because of this. One last thing: when Vishal was on our podcast, he considered LLMs to be a great achievement, but was skeptical about their ability to make new discoveries, especially in the scientific field.

I'm curious if you think the advances in mathematics support this, and what are your latest thoughts on the limitations of the current model architecture versus whether we'll need something more? Ahem. So here's my kind of new take: I think we know exactly how these things work. You put a bunch of data in them , and they are tied to that data. They can only do what is within the distribution, and they can move along this manifold in a completely Bayesian way. So, we, okay, we can say these words, and then the question is, okay, but what are the consequences of that?

For example, what problems can this solve, right? I think it's just very difficult for a person to reason, to reason about a digital artifact, in this case a model created for 5 billion dollars. In all of human history, we have never created a single digital artifact that had so many flops and so much data. So, on the one hand, we know exactly how it works from a mechanical perspective. On the other hand, there's so much data out there and so much computing power, maybe it's all there already, and it can solve anything you want.

And so, you know, the conversation shifted from how do these things work? Do we know if it can do things outside of the distribution? No. Is there, you know, transfer learning? Probably not. For example, if it teaches one thing, it doesn't teach something else, is the singularity near? Probably not. I think everyone, most people, agree that we are not in a position to take off quickly. You know, we're limited by the fact that it works within the boundaries of the distribution. We did n't get close.

We all agree with this. But I don't think anyone knows, okay, but you're still putting $10 billion into this thing. What is it capable of now? And if you factor in this meta-economic mechanism, which means Anthropic's ability to raise a lot of money and then pour all that money into this thing to create this super-powered thing. I don't think any of us can predict what this means or where it will lead. And so this is a different conversation, but the question remains the same . For example, could it cure cancer?

Perhaps. But if you invest $20 billion in something, maybe it can effectively cure cancer. And this is, in my opinion, where the discourse has evolved and where it is now. And I, honestly, decided that I couldn't predict what an artifact that cost $20 billion to create was capable of. I think it's just so important. It is important for people who follow everything new deeply to recognize that they cannot predict. And I think that's great , because it turns out I wrote 58 notes about what the internet would become, and I was very wrong about a lot of them .

But I really think that about this, I... But even that's a little different. It's like I'm taking 20 billion dollars and putting it into a model, right? And then you and I look at this model and can do whatever we want with it. I do n't think we can grasp what that even means. That's so many flops and so much data. I don't know what it's capable of. I think that's true . And I think, but I'll say in terms of biomedicine in particular: look, the other half of my household is a research physician who uses AI.

We have a Spark at home, and it downloaded a bunch of things like Spark from Sun. No, no. Nvidia Spark. Yes. No, no, no. Not with a "C", but with a "K". Oh, yes. Oho. We were in the old days, I thought this wasn't Spark by Scott McNeely. No, no. Hmm, it seems like a relic and not. And um, and all this AI, it deals with brain issues and brain surgery. Everything is AI. And it's very interesting to watch because what it can really do is see patterns that you don't see, ones that only experience could tell you about.

But if there are 10,000 articles on a topic that are part of its model, then it simply finds patterns that no one else has seen . And this is a pretty basic AI capability at this point, but it actually opens up solutions or problems or research directions and so on. I will just say that this is not magic for drug discovery, because the hardest part of drug development has always been the non- candidates. It has always been efficiency and safety. Since the 80s, we have been able to develop more candidates than we have been able to test.

These are human patients, and it's very, very, very difficult. Can I just tell you something I did wrong? So, um, I really like the question you asked: "How has our thinking evolved about, you know, do these things have, you know, their capabilities in general?" " That's, um, I was responding to this idea of ​​Bostrom's recursive self-improvement and rapid takeoff: you build one of these things, you step aside, and it takes over the world, right? And I was kind of skeptical about it, because obviously that's not happening, and I think a lot of people would agree that's the case, right?

But here's where I was wrong. What I was wrong about was that I didn't know that we could effectively keep pouring money into this as if the laws of scaling were working. And I, I don't know, you know, I don't know what that means—well, let's say we're doing a hundred billion dollar training, we have this thing that you're putting a hundred billion dollars into, and that money comes from this meta-economic mechanism that could want to solve anything —maybe they want to cure cancer, but they could also want to create weapons, who knows?

And so this concentration of so many resources in a useful way, I think, is very new. I don't think we understand the consequences. I think it's fair to say that it's very dangerous if you somehow misuse that $100 billion . So I think that's where this discussion should develop. Less about abstract concerns, you know, and more about what it means to be able to concentrate resources, which was also...I mean, you're essentially talking about exponential growth, and it's just an exponent in dollars, and we all know that none of us are very good at modeling an exponent.

Yes, we could never do this as a complex engineering project. You didn't just solve one problem with a lot of money or just build this car. Well, we, we, you know, you're right. You're 100% right, and I totally agree, but I just remember sitting in meetings at Intel where they were saying, we have 5 gigahertz, we have so many gigahertz, so many transistors, and literally no one knows what we're going to do with all of this. No, no, you are building a mechanism. I'm saying that in this case, if you want to exhaustively explore every combination of proteins, right?

We can just turn this into a money problem. Yes . It turned out very strange, which is a great way of saying that we can take previously infinite problems , put in capital, and they become finite. This simply turns it into a capital problem, not an engineering problem. Yes. Yes. Which are just completely different laws of physics. Aha. Yes. We 're rounding up, because it's already half past two. This was a great episode. Thank you for Thank you.

Go deeper

• All A16Z episode guides