The Evolution of Computers with Martin Casado and Steven Sinofsky

Summary

• Martin Casado’s core thesis: AI has moved the industry from engineering-bound to capital-bound, changing the priors investors use about capital, innovation, competition and defensibility. “Right now, if I give 20 people $1 billion, they can actually use it usefully.” Sinofsky adds the historical rhyme: computing was capital-bound for its first 30–40 years, then engineering-bound, and is now capital-bound again — “you have to hop back 40 years.”
• Both caution that headline math breakthroughs do not yet establish economic value. Casado’s test is economic utility: the summed postdoc salaries of people who have worked on these problems “is probably not very much,” so solving them shows that models are good at axiomatic systems, not that they unblock markets — “for me, it’s still in the domain of: it’s really good at playing a game. This is the best StarCraft player ever.” Claims that “if it can solve all math you can predict anything” are “a huge logical leap.”
• AI’s distribution and capital advantages have put startups on competitive footing with incumbents. Casado: “AI solves the distribution problem—it solves the demand problem” — demand for tokens and GPUs is effectively unlimited, so top-of-funnel growth becomes a spending decision — while mega-raises put challengers on competitive footing with Microsoft and other large companies. That’s why we’re seeing meteoric growth from Cursor, Anthropic and OpenAI.
• Incumbent failure remains largely cultural, while AI has weakened traditional startup disadvantages. Sinofsky, who fought ARM disruption from inside Microsoft (“I pulled out the first Surface… ‘it’s an ARM chip’”), says big companies cannot easily change scorecards, field sales, go-to-market, compensation or organizational structures. He says Google has abundant data and intelligence but its models are being trounced by OpenAI and Anthropic; Casado points to cultural factors and notes that large companies appear capital-constrained, including Google’s bond deal and unnamed rumors of token rationing that starved internal products.
• Sinofsky says we understand the mechanics but not the capability of enormous artifacts; Casado flags what he got wrong. Sinofsky describes current models as data-bound and largely in-distribution, says transfer learning is probably absent and that we are probably not in a fast takeoff, but admits nobody can reason about a digital artifact built with $5 billion, let alone a $100 billion training run. Casado says he had dismissed recursive self-improvement but had not realized how long scaling laws might continue to absorb capital. Sinofsky says a $20 billion artifact might perhaps cure cancer; Casado says the concentration of resources could also be dangerous.
• The tradeable reframe: previously infinite problems can become finite when capital can be applied. Casado: “I want to exhaustively explore every protein combination—we can just turn that into a money problem.” Corollary for VC: “too much capital chasing too few deals” is zero-sum thinking; more private capital can grow the market through capital-consuming technical waves and by letting companies stay private longer.
• Casado’s philosophical warning: this may not be just another abstraction layer, because it can mean abdicating logic itself. Every prior layer mapped down deterministically; expert systems and Prolog still had humans define the end state. Now “you kind of pray to the model god in the right words” and receive an answer that happens to be useful. That may force veterans to rebuild their assumptions about model-versus-app value capture, guarantees, productivity and defensibility.
• Sinofsky’s counterweight: tool panics recur, and the app wave is the prize. Graphing calculators, the Osborne computer being banned from Harvard Law exams, and Cornell refusing AI in freshman writing echo the same pattern: “people react to change more than they react to the baseline.” With capital replacing decade-long recruiting, domain experts may finally build software for “the world that’s unserved by software—which is literally all of it.” Erik summarizes the shift: “No-code is finally here.”

Deep dive

1. The math moment divides the world — and the mathematicians are the excited ones

• Erik opens with Jared Kaplan’s tweet from a few days earlier about telling Claude to try to solve the Riemann hypothesis and “try harder.” Sinofsky’s read: the moment splits people into the very excited and the “it’s fake, it’s going to put people out of jobs” camp — and what confuses everyone is that “the group that’s most excited are mostly the mathematicians,” the people most directly impacted.
• Both flag their limits upfront — “you’re talking to two systems guys, two product guys… this is not one of them” — since the discussion is deliberately peanut-gallery reasoning from first principles.
• Casado complicates the excitement claim: he’s found it mixed — some mathematicians say it “solves 20% of my job—the 20% I didn’t like anyway,” while others are in “existential crisis.” His cynical hypothesis, which he partly retracts, is that if someone is depressed that a problem got solved, “maybe literally the entire utility of that problem was keeping somebody employed to solve it.”

2. Casado’s economic-utility test: is this solving anything the market actually wanted solved?

• The core argument: sum the postdoc salaries spent on these famous problems over the years and it is “probably not very much.” There was never a large economic incentive, so the fact that a problem was longstanding is not evidence that solving it unlocks market value. A postdoc “ruminating on a problem, getting paid $30,000 a year for 5 years” is very different from the market deciding that this is the one thing to unlock.
• His second observation: it is unsurprising that AI excels at “an almost purely axiomatic domain” requiring breadth across disparate fields. Reading the celebrated solutions, “the solution was pretty straightforward. It just borrowed from a bit of math that I didn’t know.” The meta-learning is that AI will solve problems that require being “too broad for most humans or most educational systems.”
• Sinofsky offers a counterweight: math is “very much a leading-edge indicator of what the market might be interested in.” His example is an AT&T algorithm for linear algebra whose payoff was calculating the United Airlines flight map “3 hours less time than we could last week.” Casado holds his line: maybe today’s problems are key unlocks, “I just haven’t seen that yet.” Startup pitches split between the claim that “the foundations of AGI and reasoning are going to be math” and the response that math “doesn’t tell you anything about reality.”

3. The four-color theorem as precedent: compute turns problem classes into tools

• Sinofsky’s history lesson, taught to him by John Hopcroft at Cornell: the four-color theorem says that any 2D planar map can use four colors without adjacent areas sharing a color. It was not proved with “a proof that looked like calculus.” Instead, the solution space was shown to be finite—roughly 200 pages of combinations—and then all of them were computed. The result was possible because of compute and useful for setting practical bounds.
• The lesson he draws for AI: “when you have a new level of abstraction that says this is a whole class of problems that can be solved, you can then build tools working at that level of abstraction.” People do not all have to start from the 2–3-tree representation.
• Sinofsky contrasts mathematics with history: mathematics has a “super-long historic arc of layering on abstractions,” while history has “basically no abstraction… just a bunch of facts” from which people develop explanatory models.

4. Will math ever predict physical reality? Casado says the domains may be disjoint

• Casado, drawing on his own work in large simulation codes for phenomena such as exploding stars and airplanes in a wind simulator, says those systems compute enormous differential equations but are based on empirical results. Sinofsky clarifies that the equations of state were empirical as well.
• Casado asks whether simulation is computationally irreducible, such that one must actually run it, in which case AI’s math prowess may not help. He allows that some separate algorithmics, modeling or logistics domain might benefit, but is unsure.
• His verdict on the claim that solving all math means predicting anything: “I think that’s a huge logical leap,” and it is not clear that the claim is true or supported by any indication. “Will this star explode?” and “Will this building stand up?” remain different questions from solving an axiomatic game.

5. The Curta, the Osborne, and the eternal ban reflex: people react to change, not the baseline

• Sinofsky’s props carry the argument: a basic math tool bought in a Beijing market, followed by a Curta, an Austrian round slide rule “like a coffee grinder.” The Curta contains 600 pieces of machined metal, was brought back from the war by his uncle, and would cost $50,000 to make now. It illustrates how a new tool can create a new level of solvable problems.
• When the TI-85 arrived, “all of the math teachers had this crisis”: students could plot equations and solve them on a calculator, so “our field is dead.” Sinofsky’s broader point is that people react to change more than to the baseline they started with; calculus did not feel like a threat to people who already treated it as foundational.
• The receipts pile up: two students brought an Apple II and an Osborne to Harvard Law exams, and the school banned the computers. Articles of the era sounded like “don’t use the graphing calculator,” alongside broader cultural warnings such as “don’t listen to rap music” and “don’t play Dungeons and Dragons.” Three years ago Sinofsky tried to get Cornell to use AI in freshman writing and “they just stopped talking to me”; in fall 1983, he had needed a dean’s permission to use his own computer for freshman English papers.
• The deeper economic thread is that computing repeatedly emerged from specific utility: calculating tides, then ballistics and other war-related problems; Bletchley Park’s codebreaking; and machines that could perform such calculations thousands of times faster than a human. Those applications were culturally lauded, and parents encouraged children to study the mathematics behind them. Casado’s standing question is whether today’s model-math advances are similarly blocking real economic value: “I don’t know the answer to that.”

6. Casado’s real worry: this abstraction is different, because we’re abdicating logic

• Casado’s push, stated as doubt rather than conviction: “I don’t think in the history of computer science… have we ever abdicated actual reasoning or logic.” Compute, network and storage were resources; the human defined the problem. Even with third-party libraries, “correctness and logic for the program is under the programmer’s control.” Now “you’re actually abdicating logic to a third party. You’re like, tell me the answer”—and you may not even be sure what the question is.
• Sinofsky counters with the 1980s expert-systems era: Stanford projects combined AI with medical diagnosis and chemotherapy, and he worked on organic synthesis with a Harvard team. He says that period was an early example of turning over decision-making. Sinofsky also notes, “I’ve written a lot of Prolog.”
• Casado, who understands that tradition, says it “just didn’t work.” Even in Prolog, the programmer provides the end state and the system finds a route to it. The progression he sketches is imperative programming (you write the recipe) → declarative programming (SQL, Datalog and makefiles specify the end state) → a new system where “you don’t really know what the end state is,” so “you kind of pray to the model god in the right words” and get an answer that turns out to be useful.
• His conclusion is hedged but consequential: this “may really be the next abstraction,” a more human-level abstraction that “doesn’t map directly.” Veterans’ 40–50 years of systems intuitions—value to model versus app, what capital can do, what guarantees are possible and how productivity changes—may need to be rebuilt from fundamentals.

7. From engineering-bound to capital-bound: the billion-dollar thought experiment

• Casado’s signature framing: 20 years ago, hand a 10-person startup $1 billion and it would not know what to do with the money. The exchange points to buying computers and hiring engineers; the team would overwhelm its ability to use the capital because “the mythical man-month is very real.”
• “Right now, if I give 20 people $1 billion, they can actually use it usefully.” The industry has moved from engineering-bound to capital-bound, which Casado calls fundamentally different and unlike prior periods.
• Sinofsky adds that engineering does not scale: his career involved recruiting teams, giving them money and waiting while the engine developed. Computing was capital-bound for its first 30–40 years—step one was “we have to get one”—then became engineering-bound, and is now capital-bound again.
• The Lean Startup versus “fat startup” debate—Eric Ries versus Ben Horowitz—is newly relevant. Engineering complexity was historically the natural limiter on deploying capital; Patrick Collison’s question to Sam Altman about colossal raises anticipated the change. “We now have a discipline for taking a lot of money with small teams and using it productively. That’s a very, very big change.”
• Casado’s VC-industry corollary: the decade-long complaint that there is “too much capital chasing too few deals” is “zero-sum thinking.” More private capital can grow the market through capital-consuming technical waves such as AI, and because companies can stay private longer, more value can accrue on the private side.

8. Incumbents never crush the startups — and this time the startups have the capital too

• Sinofsky’s lived lesson from the big-company side: “you always think, oh my God, we’re just going to crush all of these little companies. And then you realize they never get crushed.” Startups do not aim directly at incumbents, while incumbents focus on one another; “Microsoft is worried way more about what Amazon and Google are doing than anyone in the startup space.”
• He repeats the disruption lesson that Clay framed as something that should be treated as a fact in the physics department rather than merely a theory in business school.
• Casado’s two structural changes are that AI “solves the distribution problem—it solves the demand problem,” because demand for tokens and GPUs is so large that top-of-funnel growth becomes a budget decision, and that these companies can raise enough money to reach competitive footing with the Microsofts and other large incumbents. Hence the meteoric growth of Cursor, Anthropic and OpenAI.
• Unlike cloud, where “nobody thought they could put AWS out of business” and startups built in a smaller corner while accepting the platform oligopoly, these AI companies are taking on incumbents directly.
• Sinofsky says Google has all the data and intelligence but that its models are being trounced by OpenAI and Anthropic. Martin attributes this to the cultural element; Sinofsky agrees culture is likely important and adds that freeing enough capital may also be difficult. Casado notes that large companies appear capital-constrained, citing Google’s bond deal and rumors about an unnamed company rationing tokens toward enterprise customers while internal products were AI-starved.
• Sinofsky’s ARM war story makes the cultural point concrete: he presented the first ARM-based Surface to Intel leadership, and “the fact I even brought one into the building was very, very tough.” Intel treated ARM as a printer chip rather than a threat, just as large companies can treat new AI directions as outside their established operating model.
• The app-wave upside is that capital can replace years of recruiting. The domain expert—such as a commercial-real-estate practitioner or the doctor who spent years writing a DOS program to schedule appointments around equipment and appointment length—can now build software for “the world that’s unserved by software… which is literally all of it.” Erik calls the shift “No-code is finally here.”

9. Nobody can predict what a $20 billion artifact can do — and Casado admits what he got wrong

• Sinofsky’s current mental model is: “we know exactly how these things work. You put a bunch of data in them. They’re stuck to that data.” He describes them as capable of in-distribution work along the data manifold, says transfer learning is probably absent, and says that RLVR on one thing does not necessarily transfer to another. He thinks many people agree that we are not in a fast takeoff.
• But nobody can reason confidently about a digital artifact built with $5 billion, let alone one built with $10 billion or $20 billion. “In the history of humanity we’ve never created a single digital artifact” with that much data and that many FLOPs. Sinofsky says he cannot predict what such an artifact is capable of; “will that be able to cure cancer? Maybe.”
• Casado admits he had dismissed the Bostrom-style idea of recursive self-improvement and fast takeoff because that was not what he saw happening. What he got wrong was not realizing that capital could continue to pour into training while scaling laws held. A $100 billion training run enabled by this “meta-economic machinery” might be aimed at curing cancer, but could also be used to create a weapon. Casado says the risk discussion should shift from fear of fast takeoff toward the implications of concentrating that much resource, which “you could reasonably argue is very dangerous.”
• Sinofsky’s biomedical example comes from the research doctor in his household, who uses an NVIDIA DGX Spark for brain and surgical-brain work. If 10,000 relevant papers are in the model, AI can find patterns that previously required individual experience, opening research directions and possible solutions. But “this is not magic for discovering drugs”: the bottleneck has historically been clinical trials, efficacy and safety in human patients, not merely generating candidate compounds. Since the 1980s, candidates have been developed faster than they could be tested.
• The closing synthesis is Casado’s conditional example: “I want to exhaustively explore every protein combination—we can just turn that into a money problem.” The broader claim is that capital can make previously infinite problems finite, “just making it a capital problem and not an engineering problem”—a very different set of laws of physics.