In a recent episode of the a16z podcast, general partners Martin Casado and Erik Torenberg were joined by board partner Steven Sinofsky for a wide-ranging discussion that questioned the very foundations of the tech industry. The trio delved into what AI's breakthroughs in mathematics truly signify, whether the underlying logic of innovation is being fundamentally rewritten, and how this reshapes the competitive battle between nimble startups and entrenched tech giants.
The central thesis of their conversation is that the basic assumptions which have underpinned the computer industry for the past 75 years are being systematically re-examined by AI. The most critical of these assumptions—that innovation is an engineering problem, not a capital problem—has been invalidated. AI is rewriting the economic logic of innovation at its core, shifting the primary constraint from engineering talent to available capital. A team of 20 people can now effectively deploy one billion dollars, a reality that directly impacts startup versus incumbent dynamics, venture capital strategy, and our very perception of AI's capabilities.
From the Abacus to AI: A 75-Year Leap in Abstraction
Sinofsky illustrated his point by referencing a 1953 IBM brochure, which featured a cover of atoms orbiting a human head and opened with the line: "It has taken man millions of years to recognize the value of the wheel." The brochure dedicated an entire page to explaining what a "digital computer" was, breaking it down into input, storage, computation, control, and output. According to Sinofsky, this five-part framework has been our fundamental understanding of computers for three-quarters of a century.
From the abacus to the slide rule, from the difference engine to the personal computer, and from graphing calculators to cloud computing, each technological leap is essentially an "upward shift in the abstraction layer." Lower-level problems become encapsulated, allowing humanity to solve new, higher-level challenges. When the TI-85 graphing calculator appeared, math teachers collectively panicked, fearing their profession was obsolete. Yet, as Sinofsky noted, they never complained about the invention of calculus because it was already their starting point. People's reaction to change is always more intense than their reaction to their current baseline, a dynamic that perfectly mirrors the anxiety surrounding AI's mathematical breakthroughs.
AI Solving Math: A Breakthrough or a "Game Player"?
The conversation was sparked by recent attempts to have the AI model Claude tackle the Riemann Hypothesis, a famous unsolved problem in mathematics. Casado offered a sobering counterpoint, questioning the economic value of such an achievement. He pointed out that when you sum the salaries of all postdoctoral researchers who have ever worked on these problems, the total is surprisingly small. This suggests the market has never prioritized solving them, making it unclear if such an achievement unlocks any significant economic value. He likened AI's mathematical prowess to a StarCraft world champion—impressive, but difficult to directly connect to real-world economic productivity.
Sinofsky offered an alternative perspective, suggesting that breakthroughs in mathematics might be about "generating new abstract tools." He drew a parallel to the proof of the Four Color Theorem, which wasn't an elegant mathematical derivation but a brute-force enumeration of all possible cases by a computer. This trade of cleverness for computational power demonstrates that once a new abstraction layer exists, everyone can build tools on top of it rather than starting from scratch.
The Biggest Paradigm Shift: From Engineering Bottleneck to Capital Bottleneck
The core insight of the discussion came from a thought experiment proposed by Casado. Twenty years ago, giving a ten-person startup one billion dollars would have been pointless; they could only spend it on servers and wouldn't know what to do with the rest. Ten years ago, the money could be used to hire engineers, but the mythical man-month was a real constraint—adding more people to a project often made it slower. Today, however, a 20-person team can effectively and efficiently utilize one billion dollars. Casado's conclusion was stark: "We have transformed this industry from an engineering-constrained problem to a capital-constrained problem. That is fundamentally different. We have never been in this state before."
Sinofsky added that this is not the industry's first period of capital constraint. The first 30 to 40 years of computing were also capital-dominated, where the first step in any project was simply obtaining a computer. The industry then moved into an engineering-constrained era, and now it has cycled back to capital constraint.
Why Startups Like Cursor, Anthropic, and OpenAI Aren't Being Crushed by Incumbents
Conventional wisdom suggests that giants like Microsoft, Google, and Meta, with their vast resources, data, and distribution channels, should crush all competition. Yet, the reality is that companies like Cursor, Anthropic, and OpenAI are growing at "meteoric speed." Casado attributes this to two main factors. First, AI has solved the distribution problem. In the past, it was difficult to measure the effectiveness of marketing budgets. Now, with an almost infinite demand for compute and GPUs, companies can directly determine how much to invest in driving growth. Second, startups can now raise enough capital to compete with the giants on a level playing field.
Sinofsky added a candid observation about the incumbents' mindset: Microsoft is more worried about what Amazon and Google are doing than about any startup. Startups don't directly attack the giants, and the giants don't even notice them. He revealed a telling detail: some large companies are so focused on prioritizing compute for enterprise customers that their own internal product teams are in a state of "AI famine," while their competitors face no such scarcity. Drawing on his own experience at Microsoft, Sinofsky recalled presenting the first Surface tablet to Intel executives, whose interest immediately cooled when they learned it used an ARM chip. They viewed ARM as a "printer chip" and dismissed it, anchored as they were to the culture of Moore's Law, just as Google's culture is anchored to hyperscale computing rather than edge AI. He summarized that disruption is a cultural constant within large companies; you cannot change their scorecards, sales systems, compensation structures, historical baggage, or customer commitments. These are, in his words, "the laws of physics."
The Unknown of a $20 Billion Model
The conversation concluded by confronting a deep, unresolved unknown. Casado admitted to a previous error in judgment. He had correctly predicted that recursive self-improvement and rapid takeoff were unlikely, but he hadn't anticipated the ability to infinitely pour capital into scaling laws that continue to hold. When you take $20 billion and invest it into a single model, he confessed, no one can truly comprehend what that means. With that much compute and data, the capability ceiling is unknown. Sinofsky agreed, likening it to exponential growth, which no one can intuitively model. He gave a concrete example: enumerating all protein combinations, once an "infinite problem," can now be turned into a "capital problem." This ability to convert seemingly infinite problems into finite ones by throwing capital at them is, as Sinofsky put it, a "very strange thing." It is this very uncertainty that has led Casado to stop trying to predict AI's limits: he has concluded that he cannot forecast the boundaries of something created with $20 billion.
How AI Changes the Economics of Innovation
The podcast, recorded on August 26, 2026, and running 1 hour, 2 minutes, and 29 seconds, featured Martin Casado and Erik Torenberg, general partners, and Steven Sinofsky, board partner, all from a16z. They explored how AI's latest breakthroughs in mathematics are reshaping the technological landscape and questioning the basic assumptions that have supported computer science for decades.
The debate centered on whether AI's progress in mathematics represents a true leap in reasoning or simply a new tool for solving problems at a higher level of abstraction. By tracing the history from the Four Color Theorem and early computers to graphing calculators and today's models, they examined how new technologies have repeatedly changed the question of which problems require human attention, asking: what is different this time?
The conversation then pivoted to one of the most significant shifts in AI: the transition from problems once limited by engineering talent to those now conquerable with capital and compute. They discussed the implications for startups versus incumbents, venture capital, and the coming wave of AI applications, and why investing billions into increasingly capable models might force a complete rethink of what these systems can ultimately achieve.
Martin Casado 00:00 Today, if I give 20 people a billion dollars, they can genuinely put it to good use. We have, in a sense, transformed this industry from an engineering-constrained problem to a capital-constrained problem, which is fundamentally different. Mathematics, to a large extent, is a leading indicator of potential market interest. Some people walk in and say the foundation of AGI and reasoning will be mathematics. But for me, that doesn't say anything about reality—it still remains in the category of "playing a game very well." Startups are not going to directly challenge the existing giants.
Steven Sinofsky 00:28 And the incumbents aren't really paying attention to startups either. Microsoft is more worried about what Amazon and Google are doing than about any startup. Everyone from a big Silicon Valley company thinks "we're going to crush these little companies," only to find they never get crushed. I think that's exactly why we're seeing Cursor, Anthropic, and OpenAI rise so fast, despite scarce capital and difficult fundraising.
Erik Torenberg 00:54 First, thank you both for taking the time. A few days ago, Jared Sumner tweeted that he had Claude attempt to solve the Riemann Hypothesis, asking it to "try harder." I don't know if there was any real progress, but it's part of a larger topic—there seems to be some real achievement happening, and we need to understand what it means for mathematics.
Steven Sinofsky 01:17 ...Real achievements are being produced, and how should we understand them? What do they mean for mathematics? I'll start with my view. I'm not a mathematician at all, but I think this is an important moment because it divides the world into two groups: those who are very excited and think "wow, these problems are solved, even if I can't understand them"—in fact, very few people can—and those who think "this is fake, it will cost people jobs, and no one knows where these fields are heading." Interestingly, the most excited group is primarily the mathematicians themselves. This confuses many people—because following the logic that "AI will cause unemployment, make people stupid, and mark the beginning of a dumb age," you'd expect those most impacted to be the most upset, yet they are the most excited.
Martin Casado 02:28 This itself shines a spotlight on our current moment. I must state upfront that we're both systems people and product people. This isn't our true specialty, so I can only offer an outsider's perspective. First, I think economic utility is a critical measure when evaluating AI. These mathematical problems have consumed immense effort, but if you sum the salaries of all the postdocs who have worked on these problems over the years, it's not a huge number. So my question is: these capabilities are fine, but the problems have remained unsolved for a long time, and that doesn't necessarily mean solving them unlocks massive economic value—perhaps it's simply that no one has ever been willing to pay the high economic cost to drive their solution. It's not that these problems aren't hard; we just lack the validation that a solution would unleash a flood of economic value. Second, it doesn't surprise me that AI excels in an almost purely axiomatic field. This field requires mastering a vast body of diverse knowledge and piecing together answers from a highly dispersed space. I've read a lot about these mathematical solutions, and many people's reaction is: "The solution is actually quite direct; it just borrows from a branch of math I wasn't familiar with." On a meta-level, the takeaway is that there exists a class of problems requiring a breadth of knowledge beyond what most people or most educational systems can cover, and AI can solve them. It's clearly very good at axiomatic systems, but I don't think that's strong evidence it can solve problems the market truly can't—because no real market has ever formed around them. That's the question we need to answer next.
Steven Sinofsky 04:34 That's exciting and sounds reasonable, but the long-term impact is unclear. I do think math is, to a large extent, a leading indicator of potential market interest, which is interesting. I remember in school, someone at AT&T invented a new linear algebra algorithm—a new way to solve linear equations, which is still crucial in AI today. And his "killer app" at the time was merely: we can now calculate United Airlines' flight graph in three hours less than before.
Martin Casado 05:15 But the question is whether these solved math problems are actually the "bottlenecks" for tasks of major economic significance. If they were, they would have been solved already. There's a huge gap between a postdoc earning $30,000 a year thinking about the same problem for five years and a market deeming a problem "the key to unlocking a major economic productivity use case." Maybe these solved problems are truly the key, but I haven't seen such a case yet. For me, that's the question I most look forward to having answered. I don't even know what 12-dimensional space is or what it means. To me, this still remains in the category of "playing a game very well"—like the strongest StarCraft player ever. It's cool and strong, but I find it hard to connect to real-world applications. Perhaps these problems weren't solved before, first because there was no economic demand, and second because how these answers actually translate into practice remains unclear. We get all kinds of project pitches; some people walk in and say: "The foundation of AGI and reasoning will be mathematics. Once achieved, it can answer all questions because the universe is built on fundamental mathematical principles—understand math, understand everything." Others walk in and say: "This is all well and good, but it can't tell you anything about reality."
Steven Sinofsky 06:48 So there's a lot of work to be done here, and it's not just about "making math better." Part of the reason mathematicians are so excited is that their way of working is unique. History, for example, has almost no abstraction layers—just a pile of facts, and then scholars develop models like "mechanical diagrams" to explain wars or famines. Math, on the other hand, has an incredibly long lineage of layering abstraction upon abstraction. Mathematicians are so excited precisely because a large body of mathematical work has suddenly jumped to a new level of abstraction. Of course, I've found the situation is mixed—some are very excited, others are having an existential crisis. The excited ones basically say: "This solves 20% of my work, and it happens to be the 20% I didn't like, freeing me to explore truly important new areas." I've been thinking about whether this depends on the type of problem being solved. If AI cured cancer, cancer researchers certainly wouldn't say "I'm so depressed." But if someone has dedicated their life to a particular math problem and AI solves it, they might indeed feel a loss—perhaps the problem's entire meaning was just to give someone something to do, or to write a few papers on "where I went wrong." In other words, behind the solution, maybe there's nothing. But maybe that's too cynical; I'm going too far.
Steven Sinofsky 08:45 A bit cynical, but not entirely. Let me offer another angle. Looking back at the history of computer science, I once took a course that I've since noticed is no longer required in university catalogs—discrete mathematics and computational complexity theory, which was required for a long time. Donald Knuth's "Concrete Mathematics"? I didn't take it; I went to a state university. That's a Cornell thing. My course was taught by John Hopcroft, a titan of algorithms—ironically, he's a Stanford PhD and invented important data structures like 2-3 trees. He taught us these things, and questions like P versus NP. I also remember the Four Color Theorem—proved by computer, yes, that's exactly the point I'm making. You've hit on the key conclusion. For those unfamiliar: we specifically studied this theorem in university, and it reduces to this proposition: for any two-dimensional planar map, only four colors are needed so that adjacent regions have no same color—a fifth color is never necessary. That's how we learned it back then; everyone could recite it, and it left a deep impression.
Steven Sinofsky 11:06 Its significance lies in this: theorists believed that if you could solve a class of problems in polynomial time (rather than exponential time), you could solve all other similar problems faster, like the Traveling Salesman Problem. This was critical at the time because computational resources were extremely limited. AT&T had a network with 6,000 switching nodes; to compute the optimal routing plan, simulations would have run for two years. The proof of the Four Color Theorem is fascinating—no one ultimately produced an elegant proof like calculus. Instead, they proved that the number of potential solutions was finite, then used a computer to enumerate all cases, proving four colors were indeed sufficient. It was a "roundabout" proof, but only achievable with computation. And it had strong practical applications, setting important boundaries in topology. This is a great lesson for me: when you have a new level of abstraction that can claim "this is a class of problems that can be solved," you can build tools at that level without starting from scratch every time.
Martin Casado 12:09 Speaking of AI, I can't help but raise a somewhat philosophical question: can mathematics represent physical phenomena? Has anyone ever actually predicted a physical fact with a bunch of equations? I've worked in large-scale simulation codes that try to simulate physics—stellar explosions, aircraft in wind tunnels—but the differential equations these simulations rely on are all built on experimental data; they're fundamentally empirical equations. This makes me wonder: is simulation computationally incompressible? That is, do you have to actually "run" the simulation to get the result? If so, I'm not sure how much AI can help. Maybe in algorithmic domains like logistics or modeling, but for practical simulation problems like "will this star explode" or "can this building stand," I think the connection to mathematical breakthroughs is quite distant. I've read a lot of the discussion about these mathematical solutions, and there's a claim that "if AI can solve all of math, it can predict everything," and I think that's a huge logical leap.
Steven Sinofsky 14:03 There's certainly no indication this is an obvious truth. I look at this from a tool perspective: AI might not be the next tool for solving math problems, but it could drive the emergence of a new type and level of model. I brought some things to illustrate this point. This is an abacus—the most primitive math tool. Before it, everything was different; with it, you had a new abstraction level. Then fast-forward through history—I also brought a Curta calculator, an Austrian circular slide rule that works like a coffee grinder, doing addition and subtraction through rotation, with 600 precision parts inside; if it were remanufactured today, it would cost about $50,000. After this, higher-level problems were solved. Are you saying new models are like new calculators or graphing calculators? I remember when the TI-85 came out, all math teachers went into crisis—"We used to have you draw X-Y equation graphs on paper; now they can do it with a calculator. Our field is over."
Martin Casado 16:26 Yes, the teachers completely collapsed.
Steven Sinofsky 16:29 But the interesting thing is, this is exactly why AI is so important. Those teachers wouldn't complain about the invention of calculus, because calculus was already their starting point, their foundation. People's resistance to change is far greater than their resistance to the status quo of their starting point. Back then, the TI-35 was one of the first calculators in schools, quite limited in function; many university courses had "no calculators allowed." My entire student career was on the edge between "allowed" and "not allowed," experiencing the shift of the graphing calculator era firsthand.
Martin Casado 17:14 I had blue exam booklets back then, having to write out every step to prove I hadn't just typed the problem into a calculator. I was the generation that had graphing calculators, and honestly, most of us wrote games on the back of them and didn't care about their math functions.
Steven Sinofsky 17:38 But if you look at this in reverse, you realize people went through the long journey of algebra, linear algebra, calculus, then Fourier transforms and fluid dynamics, all originally born from practical needs. All computers essentially originate from the difference engine, which was simply for calculating integrals.
Martin Casado 18:07 More accurately, it was for calculating integrals to shoot cannons and missiles at each other more precisely. Though actually, calculating tides was its very first origin, with huge economic value. These computational architectures were later conscripted into wartime service for calculating ballistic logarithms—all of it came from this. The difference engine was 5,000 times faster than a human and didn't make mistakes; that was very important, but its use was very specific. The interesting question is whether these models are clearly good at a certain type of math, and if so, which type—has some "atomic-level" application been unlocked?
Steven Sinofsky 19:00 I don't know the answer either. But I think it's worth deep investigation. The work of calculating ballistic tables for war ultimately unlocked the space race, jet engines, and factory automation. And at the time, people were thrilled; every parent was telling their kids "go learn this, go compete in the Westinghouse competition, go win the GE math contest."
Steven Sinofsky 19:38 The Cold War was the cultural backdrop, but there was also a general optimism about the future. I found a 1953 IBM promotional brochure with a cover of a man with atoms spinning around his head, titled "The Future of Computing." It was written in 1953, when computers were the whole world—this was before the 7040 and the 370. The brochure opens: "It has taken man millions of years to invent and recognize the value of the wheel." People back then devoured this information hungrily. The part I most want to highlight is a page titled: "The Organization of a Digital Computer." That's input, storage, computation, control, and output—for 75 years, that's how we've understood computers, and every course in school basically revolved around these five parts. Incidentally, people always forget the networking part. By the late 90s, networking was no longer a required course for computer science majors because the problem had been solved. For me, a similar watershed was the transistor—the last time a CS student had to know what a transistor was. Honestly, I don't remember it too well now.
Steven Sinofsky 22:36 These abstraction layers spawned their own specialized fields: someone spends 20 years on storage, witnessing the evolution from magnetic tubes, magnetic drums, rotating disks to tape; someone in output witnesses the shift from teletype to line terminals, then black-and-white, color, and vector displays. All these fields developed in parallel, and computer science departments emerged from math departments, initially due to the needs of missile calculation.
Martin Casado 23:28 It all eventually converged into those of us who do systems. Let me push this thread one step further. I love this "rising abstraction layer" framework—each layer has its own problems. But I still think the concept of "economic demand" is critical. Bletchley Park was for breaking wartime codes; ENIAC was for nuclear research and war efforts. We needed to compute integrals, which was done by hand, so these machines were hailed as "saving humanity," and physicists cheered. Now these math problems have been solved, but I don't know what's on the other side. That's the problem. Maybe there is something, but you have to demonstrate it.
Steven Sinofsky 24:30 We've been in similar situations before. Remember the AlphaGo moment? Before that, there was Deep Blue. I did a podcast with Frank trying to get people to understand why AlphaGo was significant. I think we can reasonably ask: what problems do these things solve, and on those problems there's indeed a lot of value and utility that pushes things forward. When that happens, people tend to get excited and dive in. For other solved problems, people have less positive reactions, and I think that's because "solving the problem" has itself become the end, not the means.
Martin Casado 25:17 If you're genuinely upset that something was solved, maybe it wasn't worth doing in the first place. Like drawing a sand mandala—perhaps it brings inner peace, but it doesn't drive the economy forward.
Steven Sinofsky 25:42 I completely agree that the wave of applications is what truly matters. The internet went through the same process. The internet in 1995 and 1996 was very exciting, but most people just said blankly, "I don't know what this does for me." There's a book now called "The Exile of Jobs" which I think is a must-read if you're listening to this podcast. It recounts that Jobs' NeXT workstation was actually the machine Tim Berners-Lee used to write the HTTP protocol. Interesting, but at the time, no one knew what the machine was for. He said "it's for looking up other researchers' phone numbers and sharing papers," and people were still baffled. The book also mentions a company called Cyber Pizza—an early food delivery platform that delivered only pizza and didn't even survive to 2000. Its product launch demo was done on a NeXT platform, with real pizzas kept in the back in case the system crashed. And the audience's reaction was: "Cool, but have you heard of the telephone?"
Martin Casado 27:34 Exactly—not everything comes with clear instructions for use. Not all platforms have obvious purposes at first, but eventually, truly useful things grow on that platform.
Steven Sinofsky 27:55 The word processor appeared in 1982. People used it on Apple II and CPM systems, and the reaction was "I don't understand why I'd use this instead of typing." It wasn't until people actually used it that they realized the typewriter approach just didn't work anymore. Someone at Harvard Law brought in the first "portable computer"—an Osborne, 25 pounds, no battery, only plug-in power, bigger than the legal carry-on suitcase, and it was the computer I used in college. In my senior year, it was banned from the exam room at Harvard Law. Previously, students could bring typewriters to exams, but when two people brought computers, the school banned them. Every single reason in those articles—don't use graphing calculators, don't listen to rap music, don't play Dungeons & Dragons—is exactly the same as today's arguments against AI.
Steven Sinofsky 29:42 Three years ago, I proposed to Cornell University to introduce AI into freshman writing courses, and they immediately stopped talking to me. But ironically, when I was a freshman, I was the only person in my 90-person dormitory building with a computer, and I had to get permission from the dean to use it to write my English essays—that was in the fall of 1983. That's exactly where we are now, just like back then.
Martin Casado 30:17 Now nobody goes to university without a computer. Let me raise a point that occasionally gives me pause: looking back at the history of computer science, we've never actually given up reasoning or logic itself—it's always just been a resource, be it compute, network, or storage. Humans define the high-level goal and then use these resources to compute the answer; the logical setup was always provided by humans. Now it feels different—you're actually handing over logical reasoning to a third party. You say "tell me the answer," and you're not even sure what the problem is yourself. Maybe it's just a higher level of abstraction, but it does feel somewhat different. When you move up abstraction layers, you still have a deterministic system where humans have defined everything about the problem at a higher level. Now it feels different. The graphing calculator felt like cheating to me back then too.
Steven Sinofsky 31:41 The exam question was "draw a graph," and the graphing calculator could just do it, which is why everyone's reactions are so similar. Speaking of computers and mathematicians, when I was a freshman, something new appeared—Maxima, MIT's symbolic mathematics software package, which could directly integrate equations on a computer. Later came Mathematica. Maxima started in MIT's labs in the sixties and seventies; by my freshman engineering class, we used its version on the IBM PC, called MuMath. We'd get our calculus homework, go to the engineering library to borrow a PC disk, and just type the answers in to get them done.
Martin Casado 32:44 That was cheating. Let me dig a little deeper. I tend to agree with you, but I occasionally have doubts: I can't recall a time when writing a program meant truly giving up logic. Using cloud databases, storage, networks—but the correctness and logic of the program were still under the programmer's control. Even with third-party libraries, I know the input, know the output, and it's me making the choices. Now it feels like we've entered a realm where we truly hand logic to a third party: you say "tell me the answer" and then wait. Maybe it's just a higher abstraction layer, but it does feel somewhat different.
Steven Sinofsky 33:33 That's the heart of this debate. For example: during the AI winter, in the 80s, one of Stanford's most important directions was integrating AI with medical schools, giving rise to medical diagnosis projects and work in chemotherapy and organic synthesis. Those were the earliest attempts to have computers involved in decision-making; the entire 80s theme was expert systems. Expert systems were our first taste of this debate. But they never truly worked, and now they do—we're making it work by giving up logic. Even with Prolog, you're still programming, still operating at the algorithmic level; you still provide the goal state, and the computer finds the path. Now, you barely know what the goal state is; you just pray to the model in the right way, and it produces an answer that happens to be useful.
Martin Casado 34:55 That's precisely what's unsettling. And what's even more interesting is that our discomfort is largely due to the real-world context we're in. For instance, many people now oppose building data centers, but two years ago, state governors were competing to attract them; ten years ago it was car factories, now it's data centers.
Martin Casado 35:33 Context is crucial to these discussions; you can't separate the two. I want to return to that level of the question: my entire career has been moving up the technology stack, but each layer has always corresponded to the next, mapping down in a basically deterministic way. This time, it feels like a different kind of new layer—perhaps this really is the next abstraction layer, one closer to human cognition, that no longer directly maps to the layers below. From transistor logic to compute, from hardware to operating systems, from applications to platforms, you've been climbing this path. But now may be a moment when we need to rethink fundamental assumptions. The biggest difference is that we're moving from computational programming to imperative programming, and now to this—arbitrary, stochastic, statistical methods.
Martin Casado 37:05 Imperative programming: you know all the steps, write the recipe, and it follows it. Declarative programming: you know the goal state, like Prolog or SQL, you describe the result, and the computer finds the path to the goal, though you can't pre-constrain the compute cost. This new thing: you don't even know the goal state clearly; you just "pray" to the model in the right language, and it produces an answer that happens to be useful.
Steven Sinofsky 37:49 That's an objective fact, and it's worth considering whether this itself is the next abstraction layer in our understanding of computation. Perhaps computing and natural phenomena will deeply intersect here, because the answers come from human-generated output—language—which is indeed different from our previous patterns.
Martin Casado 38:00 If we truly need to re-examine some fundamental assumptions, what would they be? I think people like Steven and I, through 40 or 50 years of observation, have built deep intuitions about how systems work and how they affect industries. But there are many questions where I'm genuinely uncertain: does value flow to the model layer or the application layer? How much capital can this absorb? What classes of problems can be solved, and which can't? What guarantees can be provided? What's the impact on productivity? We have intuitions about many things, and for me the biggest question is: to what extent do we need to reshape these assumptions? To give a concrete example, I've said this many times: 20 years ago, if you were a ten-person startup and I gave you a billion dollars, what would you do with it? You'd probably buy servers, hire people, and not know how to spend that much money. Ten years ago, give you a billion dollars, you'd hire engineers—code needs to be written. But in terms of engineering, the mythical man-month effect is real—adding more people doesn't proportionally speed things up. Today, if I give 20 people a billion dollars, they can genuinely put it to work and create value. We've already, to some extent, transformed this industry from an engineering-constrained problem to a capital-constrained problem, which is fundamentally different. It's never been like this before. This is a "law of physics"-level change—our old intuition that "all problems are engineering problems" is starting to shift.
Steven Sinofsky 40:28 That's a great framework; it forces you to build new mental models. It's also worth noting that computing was capital-constrained for its first 30 to 40 years—if you wanted to do something, the first step was "we need to get a computer." You were capital-limited. Then it became engineering-limited. Now it's back to capital-limited—it's come full circle. There's a scene in "Mad Men" where a computer appears in an ad agency and people run around trying to figure out what it can do. They don't figure it out, but having the machine made them look knowledgeable.
Erik Torenberg 41:11 Five years ago, Patrick Collison interviewed Sam Altman on a podcast, and Patrick noted that we've been in the lean startup era, but OpenAI and a few projects you've been involved with raised massive funding from the start. Is that underestimated?
Martin Casado 41:37 That's exactly what you're talking about. Before AI, there was always a debate between Eric Ries' lean startup and Ben Horowitz's fast-fundraising route. But there was a real constraint: engineering complexity. Now Patrick is right—we've developed the ability to efficiently absorb large amounts of capital with small teams. This is a very significant change that we haven't fully internalized. It's also why there's so much optimism now.
Steven Sinofsky 42:20 Capital scarcity and difficult fundraising are real, but once in place, building things purely with people was also hard—scaling was difficult; a team of nine isn't much faster than one person. I spent a lot of time helping early teams raise funds, then helping them hire, and then waiting two years.
Martin Casado 42:42 This also has profound implications for venture capital. Over the past decade, many people said capital was in surplus, which is a strange thing to say. Someone in VC saying "too much capital, too few good projects"—as a venture capitalist, don't you believe in positive-sum outcomes? Looking at the data, the more capital flows into private markets, the bigger the market gets, for two reasons: one, technology waves like AI can genuinely absorb massive capital; two, the more abundant private market capital is, the longer companies stay private, and more value accumulates in the private stage. So capital flowing into private markets is expanding the overall market size, not fighting over a fixed pie.
Steven Sinofsky 43:36 Early-stage VCs, who should most believe in positive-sum outcomes, have paradoxically fallen into zero-sum thinking. But from another angle, all this is laying the foundation for what I think is the most exciting thing: we're truly on the eve of an application wave. You can deploy capital without needing to have built a decade-long engineering team to produce actual output. Now, all the sectors not yet served by software—which is nearly every sector—have opportunity. Medical records, appointment scheduling, legal services... everyone says AI will finally automate lawyers' work, and it's the one change in the world that everyone welcomes.
Steven Sinofsky 44:24 VCs used to say: commercial real estate is complex; if only someone who truly understood it would build a software company, but they don't understand software; find a technical co-founder and have them spend 20 years learning commercial real estate expertise—hard. Now, going from such an idea to a product becomes a capital problem—a new abstraction layer. I remember my first client visit in my career was to a doctor who, after majoring in early computer science, went to medical school and then wrote his own DOS program to manage clinic appointments. You'd think it's just a calendar, but it's not at all. They'd hear keywords on the phone and have to judge whether it was a 5-minute or 20-minute visit, whether X-rays were needed, whether an EKG was needed, while also scheduling blood draws—not just the 10 minutes you need, but the parallel scheduling of the entire visit flow. His software did this, but it took him years. That kind of thing can now be done by people who truly understand healthcare, without needing to write code.
Steven Sinofsky 45:55 And it's not just about being able to write it; everyone's abstraction level is rising. If you're developing for mobile now, the abstraction level is already very high—you don't need to write text controls or UI components yourself. Twenty years ago, the first step of a startup was building all that from scratch.
Erik Torenberg 46:18 Yes, this is crucial for what can be accomplished. I want to talk about other fundamental assumptions worth re-examining—like the relationship between incumbents and startups. We've talked a lot about the innovator's dilemma; what's the situation now? Incumbents have capital advantages, but at the same time, you'd think startups that incumbents should have crushed long ago are thriving.
Martin Casado 46:39 It's really remarkable. If you'd asked me six months ago "what advantages do incumbents have," the answer was: capital, cash flow, distribution channels. But AI has essentially solved the distribution problem—it directly solves the demand-side issue. And the amount of capital these startups can raise puts them on equal footing with Microsoft and Meta at the capital level. So on both these points, I think we've entered an entirely new competitive landscape. Distribution advantages are often underestimated; they used to be incredibly difficult to overcome. Previously, to get users to adopt your product, you had to spend massive marketing budgets without knowing the return. Now, demand for GPU and compute tokens is almost infinite; you can decide in a fully quantifiable way how much to invest to drive growth and the top of the funnel. What used to be extremely difficult for startups is now much easier. This is exactly why Cursor, Anthropic, and OpenAI are rising so fast—the change in capital access has put them in a competitive position equal to the giants.
Steven Sinofsky 48:12 Yes, my entire career was built on managing thousands of engineers to build things that couldn't be built anywhere else—that's what makes an operating system moat, an almost infinitely deep trench. Read "The Exile of Jobs" and you'll really feel how NeXT's code started from Carnegie Mellon's Mach kernel; it couldn't possibly be built from absolute scratch. On how to disrupt incumbents, there's a joke that circulates at Harvard Business School—when Clayton Christensen was still there: they teach "disruption" as a theory in business school, but it should really be taught as a law of nature in the physics department. I was there in 1998 when Christensen was writing that book. Everyone from big Silicon Valley companies thought they'd crush the little companies, and then you find they never get crushed. Did AWS really push anyone out of the market? No. Because startups don't directly attack the giants, and the giants only watch other giants. Microsoft is more worried about what Amazon and Google are doing and doesn't care about any startup. But what truly matters in disruption are cultural factors—constants within large companies that cannot be changed, "laws of physics": you can't change the performance review system, you can't change enterprise sales and marketing strategies, you can't change compensation structures and organizational design, you can't change historical baggage and existing customer relationships. When you serve 500,000 customers, there are too many things you simply cannot do. That's the essence of disruption. Now we're in a magical moment—not only are the cultural factors unchanged, but the startup ecosystem itself is highly mature, just like in the cloud computing era: you don't build your own data center, you don't call AT&T, you can start running on day one at night.
Martin Casado 50:59 With cloud computing, everyone accepted the oligopoly; nobody really thought about taking down AWS, but rather "can we survive in our own little corner?" The question was whether Microsoft or Google would just add your features for free—the answer was often "yes," but that was a later concern. Now, these startups are truly challenging the giants head-on. You made a great point: previously, in this industry, competitiveness came from extremely complex large-scale engineering projects—building chips, systems, distributed clusters; engineers of Jeff Dean's era were the real moat, an engineering barrier that no startup could replicate. Now, for these models, the barrier is truly just capital. Google has massive data and top talent, yet its models are being outperformed by OpenAI and Anthropic. The fundamental reason is culture—once you've lived inside a big company, you know how hard it is to drive change internally. GCP is a remarkable engineering achievement, but releasing that much capital for a particular direction within a large company is itself very difficult. From their earnings calls, you can see how much they struggle with capital allocation—some move debt off balance sheets, some reportedly throttle tokens for internal product teams to prioritize enterprise customers, while those products' competitors face no such constraints.
Steven Sinofsky 53:23 I've learned to truly respect that cultural force. I fought hard against mobile platform disruption, which was essentially ARM disrupting Intel. Intel didn't care at all. I brought the first Surface to Intel and put it on the table, saying "this is our new computer." They were excited until I said it had an ARM chip inside. Their reaction? ARM? That's a chip used in printers. They said "we're also an ARM licensee, we know all about it," but they simply didn't believe ARM could do what I said—power efficiency, graphics, always-on connectivity—because Intel's culture is Moore's Law and hyperscale computing, not this. Just as Google's culture is hyperscale, not edge AI.
Steven Sinofsky 54:42 The capital inversion opportunity you're pointing to is very noteworthy: go raise money, then go after areas that seem to belong to the giants. And people no longer think you're crazy—look at the funding rounds happening. These companies have achieved considerable success as a result.
Erik Torenberg 55:05 One final question: Vishal came on the podcast earlier and argued that these AI systems are accomplished but relatively pessimistic when it comes to inventing new discoveries, especially scientific breakthroughs. I'm curious whether you think the progress in mathematics aligns with that judgment, or what your latest thoughts are on the limitations of current model architectures.
Martin Casado 55:40 My latest view is this: I think we completely understand how these systems work—you stuff a bunch of data in, it gets locked into that data, can only do in-distribution inference, and moves along that manifold in a Bayesian way. We can articulate these mechanisms, but the question is: what does this mean, and what problems can it solve? It's hard for humans to understand a digital artifact built with $5 billion. In human history, we've never created a single digital artifact using this much compute and this much data. So on one hand, we do understand its operating mechanisms; on the other hand, the sheer volume of data and compute it contains means perhaps everything you could want is already in there. The discussion has shifted from "how do they work" to "can they do out-of-distribution things—no; is there transfer learning—probably not; is the singularity coming—probably not; most agree we're not on a rapid takeoff path." But I don't think anyone truly knows the answer to: after continuously pouring in billions of dollars, where exactly is its capability boundary now? If you consider the "meta-economic machine" that allows Anthropic to raise massive funds and pour them all into models, I genuinely cannot predict what that means. It's the same question: can this thing cure cancer? Maybe. If you throw $20 billion at it, maybe it could effectively cure cancer. I've decided I cannot predict the capability limits of an artifact built with $20 billion.
Steven Sinofsky 57:42 For someone deeply focused on everything new, admitting you can't predict is a very important and good attitude. I wrote 58 memos about where the internet was going, and many were wrong. And this question is even different from those—if I invest $20 billion in training a model, and then the two of us casually use it for anything, I truly don't think we can understand what that means. That much compute, that much data, I don't know where its capability limits are. But speaking of biomedicine, my other half is a research physician who uses AI for neurosurgery-related work. What's impressive is that it can recognize patterns that only experience reveals, and if ten thousand relevant papers are in its model, it can discover patterns no single person could find alone. This is a relatively basic AI capability, but it's genuinely opening new research directions. However, regarding AI for drug discovery, I'll say this: the hard part of pharmaceuticals has never been discovering candidate drugs—since the 80s, we've been able to discover more candidate compounds than we can test. The real challenge has always been efficacy and safety validation, which requires testing in real humans, and that's extremely difficult.
Martin Casado 59:46 Let me be candid about a past misjudgment. On the general capability question of these systems, I was initially responding to Bostrom's "recursive self-improvement and rapid takeoff" theory—you create it, step back, and it takes over the world. I thought that was clearly not what was happening, and most people agree. But what I didn't anticipate was that we could so effectively keep pouring money in, and the scaling laws would continue to hold. I don't know what a $100 billion training run means—that money comes from that meta-economic machine, and the people behind it might want to cure cancer, might want to build weapons, who knows? Concentrating that much resource and using it effectively, I think that's very new, and we don't understand its implications. Directing that much capital in the wrong direction could be very dangerous. So I think this conversation needs to shift from "singularity explosion moment" to: what does it mean to concentrate so much resource—that's the direction that truly needs to evolve.
Steven Sinofsky 01:01:11 You're essentially discussing exponential growth, except here the exponential is denominated in dollars. And as we all know, no one builds good intuition for exponentials. But I'll say this: it's different from stacking complex engineering projects—you're not using massive funding to tackle dozens of problems simultaneously; you're building a machine. I remember hearing again and again at Intel meetings "we have 5GHz, we have this many transistors," and then no one knew what to do with it; everyone was just building the machine. This time, if you say "I want to enumerate all protein combinations," you just turn it into a money problem. That's a wonderful way to put it—we can take problems that were previously almost infinite and make them finite by injecting capital. Turning it from an engineering problem into a capital problem—that's an entirely different underlying law. Let's end on that. It's 2:30 now. This episode was fantastic. Thank you both. Thanks, thank you everyone.