This episode brings together three frontier founders—Guillermo Rauch (Vercel), Blake Scholl (Boom Supersonic), and Max Hodak (Science)—alongside Naval and Nivi, to discuss how AI is transforming the act of building itself. The central theme: AI is shifting the role of the engineer, scientist, and founder from doing the work to building the factory that does the work, creating massive leverage but also raising hard questions about regulation, creativity, and the future of human value.
AI Software Factories
The engineer’s job is changing from shipping output to building the factory that ships output. Guillermo describes the shift: previously, companies measured how good person A was at producing output B. Now the measure is whether you can produce the factory that generates outputs B through Z multiplicatively. This makes the idea of 100x or 1,000x engineers uncontroversial.
Naval notes this has always been true in idea domains—Satoshi, Notch, Brendan Eich, John Carmack were always 1,000x programmers. AI leverage just makes it visible and accessible.
Token consumption is the new lines of code—a misleading metric. What matters is the quality of the reprompting and the judgment the user brings. Max observes that models reflect the capability of the user: experienced developers get dramatically more out of them than junior developers.
“Waste tokens, save time.” Naval’s approach: he ignores prompt engineering tricks, brute-forces problems by throwing multiple models at them repeatedly, and treats token cost as negligible compared to his own time. Models are still cheaper than humans and will keep improving.
Models Instructing Humans
Models have graduated from junior to principal engineers. Guillermo notes a recent shift: models now proactively present trade-offs and architectural options without being asked to plan. They come back and say, “Here are three routes and their trade-offs”—behaving like a peer rather than a tool.
The feedback loop still matters enormously. Max points out that architectural decisions—choosing between Postgres and ClickHouse, ZeroMQ versus another queue—are where human judgment shapes the output. The models suggest, but experienced engineers override with taste.
The inversion is beginning. Guillermo observes that models are starting to instruct humans: “Go get me this API key,” “Raise this amount of capital.” This will accelerate as every SaaS tool gets a CLI/API and crypto enables autonomous payments.
Is Pure Software Dead?
Pure software may no longer be a defensible moat. Naval raises the question: if models can generate software on demand, what’s the founder’s advantage? Hardware becomes more valuable because it’s harder to replicate. Classic software engineering may be commoditized.
Guillermo pushes back with the “building block economy.” Mitchell Hashimoto’s argument: agents need powerful, reusable infrastructure building blocks (libraries, databases, queue systems). You don’t want an agent reinventing Postgres from scratch every time. The value shifts to the infrastructure and building blocks that agents depend on—which is what Vercel is building.
Naval reframes this: these are like libraries and dependencies, but for models. Anything already created acts like a token cache—the model forks from existing work rather than regenerating everything.
“You don’t get stuck anymore.” Max’s key observation: the biggest change is that agents eliminate the indefinite debugging black holes that used to make programming intrinsically frustrating. This lowers the barrier dramatically.
Vibe Coding Hardware
Blake describes how Boom Supersonic applies AI to hardware engineering. Traditional hardware engineering runs on siloed Excel spreadsheets with VBScript—no source control, no automated testing, manual handoffs between aerodynamicists and structures engineers. Boom built software frameworks to automate these flows, but couldn’t afford enough software engineers.
The new model: software engineers create the architectures and tools; hardware engineers vibe-code their domain-specific pieces. Example: designing a turbine blade used to take one engineer one day for one blade (with ~1,000 blades in a jet engine). Now two engineers can design an entire jet engine because they can change geometry and see real-time structures and aerodynamics results.
Max predicts the next frontier: AI currently generates software, but within 2026 it will generate STEP files and PCB layouts—mechanical and electrical engineering artifacts. This will be transformative for hardware companies.
Enterprise software is being replaced from within. Guillermo notes that startups building hardware collaboration tools are obsolete because companies just build exactly what they need internally.
Open Source Compounds China’s Advantage
China’s push for open-source models is strategically motivated by hardware dominance. Naval argues: China has complex supply chains and component ecosystems but historically worse software. If they can generate software on demand via open-source models, they erase Silicon Valley’s software advantage and let their hardware superiority dominate.
The Chinese government funds ecosystem-wide efforts in network-effect businesses. Open-source AI helps their entire hardware ecosystem—every “crappy little software” component on Amazon gets better fast.
Guillermo adds: falling behind in software generation means falling behind in everything, because every piece of the hardware pipeline requires software generation.
You Always Want the Smartest Model
Naval argues for always using the most intelligent model available. Intelligence is an unalloyed good—you always want the best judgment, especially when you’re pouring capital, code, and people behind a decision. When you don’t know which answer is correct, you default to the smarter model.
Guillermo’s data from AI Gateway confirms this: frontier intelligence dominates usage. Open models are used, but the top is heavily dominated by the most capable models. Gemini excels at industrial production tasks (support, browser automation) at the right cost-performance combination. But for pushing the frontier in coding, only 2-3 models matter, and Chinese models aren’t among them.
Max notes Science does use some Chinese models (Qwen, DeepSeek) internally, including a large internal fine-tune, but their preference is always to buy the best available.
Software Still Needs Hands
Max on vertical integration at Science: they own a captive MEMS foundry because the components they need don’t exist off-the-shelf. The closer products get to being “a single block of covalently bonded matter,” the better they perform—but this requires building fabrication capability yourself.
AI’s biggest internal impact at Science has been regulatory. Generating documentation, tracing ISO standards compliance—work that took a regulatory team several months now takes the AI minutes. This is an underappreciated application.
The boundary remains physical. Software still needs hands. Science has instrumented their foundry so that as models get better, improvements in cell engineering and material science should show up immediately.
Humans Are Becoming Verifiers
The human role is shifting to verification. Naval observes: junior engineers have been promoted to senior engineers by agents; similarly, paralegals become senior lawyers. The function moves to verifying the stack and standing behind it.
Guillermo on the PR problem: the mountain of AI-generated slop ending up in pull requests is the new challenge. The goal isn’t reading every line—it’s being able to say “I understand the consequences of this PR” because you wrote the test harness, simulations, and proofs that give you confidence. Someone still gets paged if production goes down.
Naval: humans are becoming verifiers. We train models with good verification data, and now we need human verifiers to stand behind outputs.
The Regulatory Frontier
Regulatory compliance is being crushed by AI. Blake describes certifying an airplane for lightning strike resistance: 200 pages of documentation taking months classically. With RAG, this takes minutes. The second-order effect is that changing airplane specifications now takes minutes instead of months, making engineers willing to iterate. The third-order effect: you replace mediocre compliance engineers with a small number of creative ones.
Max pushes back on Silicon Valley’s anti-regulation consensus. Many regulations reflect genuine progress (clean cities, swimmable rivers). The problem is that compliance is impossibly difficult for humans. If AI makes compliance frictionless, that’s a win, not a loss.
Naval predicts a Red Queen’s race: regulators will deploy agents, companies will deploy agents, and it becomes agent-on-agent. But government agencies will be slow adopters and may get DDoSed with documents, potentially extending approval times.
Why There’s No Innovation in Healthcare
Naval: the FDA approval process is a nightmare that strangles innovation. The two biggest tech advances of the last decade—AI and crypto—are in the math domain because it’s the last unregulated domain. Peter Thiel’s lament about no physical innovation is directly caused by regulatory barriers.
Blake: the regulatory model should shift from pre-approval to enforcement-based. Currently, building physical infrastructure requires submitting plans and waiting months for approval—guilty until proven innocent. This is insane compared to how we drive cars.
Max offers nuance: some regulation does make us safer, but the incentive structure is deeply asymmetric. If the FDA approves a bad drug, careers end. If they block a good drug, nobody notices. This creates systematic slowdown.
The problem is deeper than the FDA—it reflects voter attitudes. Americans are risk-averse about human-subjects research. Even the Right to Try Act and Single Patient IND pathway (which the FDA approves 99%+ of) don’t work because the IP owner won’t give up clinical-grade drug, and the FDA draws adverse inferences globally from any bad outcome.
Max’s proposed fix: prohibit the FDA from drawing adverse inferences across different users of the same technology (e.g., a capsid). This would accelerate innovation with light regulatory touch.
China’s CFDA is innovating. They have an approved, getting-paid implantable BCI. Their system allows trying things in humans and on market, with much lower costs. China is bringing down the cost of bringing things to market so they can sell for $10,000 instead of $100,000.
Healthcare Is a Communist Society Inside Capitalism
Naval’s core critique: there is no private market in healthcare. It functions like a small communist society inside capitalism—you don’t pay at point of service, you submit receipts to insurers/government. This eliminates price signals, competition, and feedback loops.
Naval’s proposed reform: make the first 20% of annual income your healthcare deductible. If you’re broke, it’s zero. If you’re rich, it’s millions. The rest is covered by government/insurance. This would create a private market for many procedures.
Evidence this works: dental (veneers, braces), plastic surgery, and LASIK are all advancing rapidly because they’re private-pay. People vote with their money, and the feedback loop drives improvement.
Currently, wealthy people can’t even spend voluntarily into the system efficiently—out-of-pocket prices are often 10x insurance rates because the system isn’t designed for it.
Max’s “Sid’s story” illustrates N-of-1 medicine. GitLab founder Sid, diagnosed with rare cancer, lived far past his prognosis by creating his own personalized treatment plan. Six or seven companies and twenty to thirty drugs in his escalation ladder came from his case. This kind of N-of-1 medicine is a rich source of research—but it requires enormous agency from the patient at their weakest moment. AI should democratize this.
The Autonomous Company
Guillermo describes autonomous infrastructure at Vercel. Anomaly detection fires automatically; agents investigate, create incidents, and begin remediation. Engineers are served solutions on a silver platter. They haven’t given agents the keys to change production yet.
Autonomous security research: Vercel open-sourced deepsec, which runs 10,000 concurrent agents against their entire monorepo. It found several quarters’ worth of security research progress in a couple of days for $14,000 in tokens.
Naval’s bug-reporting daemon: a simple system that compiles TestFlight bug reports, proactively analyzes and fixes them in the background, and ships a new build for testing. He jokes about an app literally built by its users—a “Homer Simpson car” with every feature.
Blake ran a company-wide hackathon: everyone from the receptionist to engineers spent a week building whatever they thought most important using AI. The receptionist built a shipping-and-receiving automation that the company actually uses. The result was mostly needle-movers, not silly projects.
The lesson: everyone has ideas about what could exist, but first-order ideas often seem stupid until you can actually build and iterate on them.
“Your job is to train the agent.” Guillermo describes the cultural shift: people are beginning to understand their role isn’t to do the work but to train the agent that does the work. Vercel is building a feature to extract skills from inputs/outputs so people can share and pool learned capabilities.
The Next Lord of the Rings
The returns are shifting from intelligence to agency. Max argues it was 70% intelligence / 30% agency; now it’s flipping to 70% agency / 30% intelligence. The humans best fit for the future are the ones who can open Claude and think “what should I build?” instead of watching YouTube.
Naval counters: it’s 99% intelligence and 1% agency, because the agents exercise the agency. You tell the agent “I’m making smart decisions; go implement stuff.” The agent can even suggest what to build next by looking at logs.
Vibe coding is replacing other entertainment. Naval stopped playing first-person shooters and now spends that time vibe coding—it’s more entertaining and produces real things. But most people still see coding as a black box. The percentage of coders may have gone up 100x (from 0.01% to ~1%), but 99% of people still never will.
Max’s bet with Andrej Karpathy: by 2030, you’ll be able to dump in a book and get a movie out. He’s specifically waiting to feed the last three Expanse books (conditioned on the TV series) and generate the final three seasons. We’ll get dozens of Lord of the Rings-level works from fans making their own takes.
Can AI Have New Ideas?
The deepest question: can an LLM go out of distribution and have a genuinely new idea? Max poses this directly. In reinforcement learning, you can sample from a distribution and get randomness that explores new territory. But can a transformer produce something truly absent from its training set?
Naval’s view: training sets are so vast that almost everything is in them somewhere. But ideas outside language—physics, emotion, evolution, feeling—may still be beyond AI’s reach. Humans can cut through infinite possibility space and eliminate huge swaths so their creativity makes sense within the larger scheme. This is a unique human capability.
Art requires intent and out-of-distribution behavior. Naval defines art as conveying emotion—something you felt, transmitted to another person. A computer is almost incapable of this because intent matters. A beautiful photo taken by a person versus AI-generated pixel-for-pixel identically: the human version has more meaning because someone felt something and wanted you to feel it.
Max’s definition of art: meaningful out-of-distribution behavior—something surprising that changes your future trajectory through the universe. This can include military maneuvers or any domain.
The bar for surprise keeps rising. OpenAI “destroyed” Studio Ghibli for everyone—once you’ve seen Ghibli-style everywhere, it’s no longer surprising and the art value is gone. Humans generate surprise completely out of the data distribution, and they do it with intent.
Naval’s Gödel analogy: an AI trained to be perfect at mathematics operates within a formal system. Kurt Gödel stepped outside the system with the incompleteness theorem. That kind of meta-level breakthrough—stepping outside to break the system—is something Naval doesn’t think AI can achieve.
The current era is human + AI. Naval bets this lasts longer than people think. Human without AI is uncompetitive; pure AI isn’t there yet; human plus AI is where all the value is being created.
A Very Large Number of Small Teams
AI increases productivity, which means hiring more people, not fewer—but in smaller teams. Basic economics: higher productivity means more wealth, which means more demand. The number of people required per task drops, but the number of different things you can attempt explodes.
Guillermo’s hypothesis: a very large number of very small teams. Instead of one company doing one jet engine with a thousand people, you get a hundred companies each doing different jet engines with ten people each. Explosion of entrepreneurship.
Generalists are having a field day. AI provides base-level intelligence and cuts through jargon. You no longer need twenty years of specialized learning to contribute. What’s left is creativity, taste, and enough agency to get started. The people hurt most are those whose expertise was purely memorized jargon and scaffolding—AI cuts right through that.
The single best thing you can do: get really good with AI tools and always know the edges of what they can and can’t do. That’s a moving target.
