By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Full Episode: The AI Industrial Revolution

Full episode with with 20 minutes of new material at the end. With Guillermo Rauch (Vercel), Blake Scholl (Boom Supersonic), and Max Hodak (Science).

Max HodakguestNaval RavikanthostBlake SchollguestGuillermo RauchguestNaval Ravikanthost

Jun 1, 20261h 10mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

Meet the frontier founders and the shift to “software factories”
The host introduces Naval and three founders—Guillermo Rauch, Blake Scholl, and Max Hodak—framing the conversation around what they’re learning from building, not just what they’re building. Naval sets up the core theme: engineering work is shifting from shipping outputs to building “factories” that generate multiplicative outputs.
- •The host frames the episode as extracting principles (“alpha”) from how frontier founders build.
- •Naval describes the transition from judging engineers by direct output to judging them by the “factory” they create.
- •Naval argues the world is moving beyond the old debate over “10X engineers” to 100X or 1,000X leverage.
- •Max notes he used to be criticized for arguing large differences in engineer productivity.
- •The group sets an open, jump-in style for the discussion.
ROI myths: token leaderboards, prompting skill, and taste-driven feedback loops
The conversation turns to how people mis-measure AI productivity—especially by counting tokens. Max and Naval describe how model output depends heavily on user judgment and re-prompting, while Blake compares token counting to the old “lines of code” fallacy.
- •Naval questions token leaderboards as a confusing proxy for ROI.
- •Blake compares token consumption metrics to measuring productivity by lines of code.
- •Max says models tend to be “about as good as you are” in a domain and depend on feedback quality.
- •Naval describes a new kind of help: telling someone how to re-prompt effectively.
- •Max expects prompting nuance to matter less as models get smarter, but says it matters a lot today.
“Waste tokens, save time”: brute force iteration and verifiable domains
Naval explains his working style: ignore prompt “tricks,” throw multiple models at problems, and spend tokens to save human time. He argues that in verifiable domains, models will keep resolving problems, even if early outputs aren’t production-grade.
- •Naval says he ignored prompt frameworks, betting models will improve faster than he can master tactics.
- •He describes using multiple models repeatedly and “wasting tokens” to reduce time-to-solution.
- •Naval argues model cost is still cheaper than human time, so the right metric is time and final output.
- •He notes even if code quality is initially low, more tokens can be used later to rewrite and harden it.
- •Naval distinguishes verifiable/solved problems from unsolved frontier problems that require closer collaboration.
Models “push back”: planning, trade-offs, and the human still completing the model
Guillermo describes a qualitative shift where models now respond with options and trade-offs rather than running away with a prompt. The group discusses remaining gaps: models may propose architectures, but humans still supply judgment, constraints, and access to resources.
- •Guillermo says models increasingly return plans with multiple routes and trade-offs, resembling more senior peers.
- •He notes models still make bad predictions and sometimes “bullshit,” but are more intellectually collaborative.
- •Guillermo and Max discuss that experienced architects may extract more leverage than juniors because of better judgment.
- •Max emphasizes career progression toward architectural choices and technology selection where human taste matters.
- •Guillermo wonders when the model will start giving humans instructions for tasks only humans can currently do (keys, capital, permissions).
Is pure software dead? Building blocks, shared infrastructure, and “token cache” reuse
Naval asks whether classic software engineering remains an investable moat if models can code from natural language. Guillermo argues for a “building block economy,” where reusable infrastructure and standardized components become even more valuable as agents assemble systems.
- •Naval questions whether pure software engineering becomes obsolete when models “speak English.”
- •Guillermo references the idea that agents need strong reusable building blocks rather than reinventing everything.
- •He argues shared standards (e.g., common infrastructure) have large-scale cooperation value.
- •Guillermo introduces the metaphor that existing components act like a “token cache” that agents should reuse.
- •Naval reframes this as libraries/dependencies, but oriented for agents rather than just humans.
You don’t get stuck anymore: vibe coding replaces the old debugging wall
Max and Naval describe how agents change the emotional and practical experience of building software. The key shift is eliminating the long stalls from obscure debugging, making building feel more like a tight feedback-loop game where real tools emerge.
- •Max says he’s built substantial software he uses daily without writing code by hand.
- •Guillermo likens effective leadership to “vibe coding through people,” now applied to agents.
- •Naval says he returned to coding after decades because agents remove the burden of learning new stacks and wiring infrastructure.
- •Max highlights the old pattern of hitting a random blocker and losing indefinite time; agents largely remove that stuckness.
- •The group describes vibe coding as addictive and more rewarding than previous leisure activities for some people.
Hardware workflows become software: real-time turbine blade and engine design iteration
Blake explains how hardware engineering often lives in fragile spreadsheets and manual handoffs, and how Boom turned workflows into software to reduce iteration cost. With AI assistance, software engineers define architectures while hardware engineers “vibe code” domain-specific pieces, enabling dramatic speedups like real-time turbine blade analysis.
- •Blake describes hardware engineering’s spreadsheet-driven, siloed workflows lacking source control and testing.
- •Boom built software frameworks to make hardware engineering flows repeatable and automatable.
- •AI enables a new division of labor: software engineers provide system architecture; hardware engineers implement via vibe coding.
- •Blake gives the turbine blade example: real-time structures and aerodynamics feedback from geometry changes.
- •He claims this productivity shift can allow very small teams to design extremely complex systems like jet engines.
Open source models and hardware ecosystems: why model access matters in supply chains
Naval argues AI-generated software is a major boon for hardware-heavy ecosystems, and speculates this is one reason China pushes open source models. Guillermo asks who actually uses Chinese models, leading to a discussion about frontier intelligence versus cost/performance at scale.
- •Naval says AI helps hardware and gadget companies that historically shipped poor software due to lack of talent.
- •He argues open source model emphasis benefits hardware-centric ecosystems and supply chains.
- •Naval suggests China’s open source push also relates to catching up via distillation and pooled effort.
- •Guillermo asks whether the others use Chinese models or know users of them.
- •The discussion sets up a later debate: cheap “good enough” versus always using the smartest model.
Always want the smartest model vs. industrial scale trade-offs in model choice
Naval argues intelligence is an “unalloyed good” because model errors are hard to detect and stakes are high. Guillermo describes observed usage patterns: frontier models dominate, but some models win for cost/performance at scale for non-coding industrial tasks.
- •Naval rejects the idea of using a cheaper model for most tasks and reserving top models only for hard ones.
- •He says when models make mistakes you often can’t tell, so you prefer the model you believe is smarter.
- •Naval notes this preference could lead to monopoly/oligopoly outcomes in AI.
- •Guillermo cites gateway traffic showing heavy dominance by frontier intelligence, with exceptions driven by cost/performance.
- •He says some models are strong for scaled tasks like support and browser automation, while frontier coding needs top-tier coding models.
Vertical integration, hands-on reality, and AI’s immediate wins in compliance work
Max explains why his company vertically integrates: key components and packaging don’t exist off the shelf if you want tight performance. He notes AI’s biggest current impact inside his company is regulatory and documentation work, while physical manufacturing still demands “hands.”
- •Max says the preference is to buy commoditized parts, but deeper integration requires making what doesn’t exist.
- •He argues tighter integration improves power, size, performance, and longevity, but requires owning key processes.
- •Max says they bought a captive MEMS foundry because required packaging/assembly couldn’t be sourced otherwise.
- •He highlights AI’s large impact on regulatory interactions: generating documentation and mapping applicable standards.
- •Max emphasizes limits: software can be smarter than humans, but physical work still needs hands and instrumentation to benefit from better models.
Humans as verifiers: from legal work to code review, guardrails, and long-term maintenance
The group connects AI-driven automation in law and engineering to a broader role shift: people increasingly verify rather than produce first drafts. Guillermo argues that with AI-generated “slop,” engineering must rely on tests, simulators, and evaluators to confidently sign off, especially over long time horizons.
- •Naval says basic legal tasks have moved away from lawyers, similar to junior engineering work moving to agents.
- •Guillermo draws a parallel to lawyers as trusted authorities: in engineering, someone must still sign off on consequences.
- •He argues teams may not read every line, but must build evaluators (tests, simulations, proofs, type checks) to ensure safety.
- •Guillermo stresses the neglected problem of software quality over “1,000 days”: security, performance, production maintenance.
- •Naval summarizes the shift: humans increasingly become verifiers who stand behind a stack and support it when it breaks.
Regulatory iteration and the coming agent-on-agent “Red Queen” race
Blake explains how AI slashes the cost of regulatory documentation, reducing change aversion and enabling faster iteration in heavily regulated domains like aviation. Max adds that many regulations embody real progress, but humans can’t handle the compliance friction; Naval counters with a warning about agencies being flooded as everyone deploys agents.
- •Blake describes using retrieval-based systems to produce regulatory test documentation far faster, enabling specification changes.
- •He outlines second- and third-order effects: faster change, fewer low-value roles, more creative iteration by small teams.
- •Max says the issue isn’t only that regulation exists; friction and slow correspondence make compliance costly.
- •Naval predicts an agent-on-agent escalation where regulators respond with more generated paperwork and get “DDoSed” by submissions.
- •Blake argues even imperfect automated critique could beat today’s months-long permitting delays, using the fire department as an example of delay-driven pain.
No innovation in healthcare: incentives, “50-state experiments,” and benchmarking regulators
The conversation shifts to why healthcare innovation is so constrained, focusing on regulatory incentives and political realities. Naval proposes state-level experimentation zones; Max argues drug and device constraints include global inferences and IP-holder incentives, then describes alternative regulatory models and China’s pace.
- •Naval claims FDA processes and broader regulation suppress physical-world innovation and create contradictory constraints.
- •Max agrees incentives are asymmetric: regulators get punished for visible failures, not for blocking unseen successes, reflecting voter preferences.
- •Naval proposes a stronger “50-state experiment” with meaningful regulatory diversity, including opt-in zones for higher-risk innovation.
- •Max argues innovation zones don’t fully solve drug access because clinical-grade supply and regulatory adverse inferences are global.
- •Max describes Europe’s notified bodies as a different review structure and says China has an approved paid implantable BCI and a system that could outcompete if the U.S. doesn’t adapt.
Healthcare as a “communist society inside capitalism,” N-of-1 medicine, and AI’s promise for patient agency
Max explains why healthcare doesn’t behave like a technological growth industry: reimbursement and fixed spending pools prevent the feedback loop seen in consumer tech. Naval argues the lack of a private market breaks pricing and incentives; Max shares an N-of-1 cancer story, and Naval points to AI as a way to democratize individualized research and options.
- •Max contrasts consumer electronics growth with healthcare’s fixed spending bucket and reimbursement-driven dynamics.
- •He argues progress requires lowering costs to bring treatments to market so they can be bought more like normal products.
- •Naval says healthcare lacks a private market, likening it to reimbursed restaurants and calling it a “communist society inside capitalism.”
- •Naval proposes a thought experiment with an income-based deductible to force market feedback (and invites critique).
- •Max recounts Sid’s rare cancer experience as an example of building a large personalized treatment ladder; Naval adds that doing N-of-1 work requires patient agency when patients are weakest and suggests AI could help democratize that capability.
Autonomous infrastructure and security: anomaly agents, remediation loops, and 10,000-agent red teaming
Guillermo and Naval describe the most autonomous systems they’re running today: agents that investigate anomalies, draft incidents, and propose remediations, plus large-scale automated security research. The theme is moving from manual thresholds and human-only ops to agent-driven continuous improvement and verification.
- •Naval describes automating SRE patterns that historically relied on hand-tuned alarms and thresholds.
- •Guillermo says anomaly alerts can trigger agent investigation, incident creation, and remediation workflows, stopping short of autonomous prod changes.
- •Guillermo describes autonomous optimization processes and autonomous security research as strong near-term wins.
- •He cites an internal/open-sourced security tool run with massive concurrency to surface substantial security findings quickly at token cost.
- •The group notes escalating cybersecurity pressure and the need for proactive, continuous agent-driven defense.
Train the agent, not just do the work: company-wide building weeks and the rise of tiny teams
Blake shares an experiment where everyone built AI-powered tools for a week, yielding unexpectedly serious, high-impact projects. The group generalizes: future work increasingly means training agents, capturing skills from workflows, and enabling more founders and smaller teams—while debating what remains uniquely human, especially creativity and taste.
- •Naval describes an app workflow where bug reports are analyzed and fixes are prepared automatically, with humans approving releases.
- •Blake describes stopping work for a week so everyone—across roles—built something with AI and demoed it; many outputs were “needle movers.”
- •He highlights a non-engineering role building a useful automation, suggesting rapid build/iterate reveals which ideas are actually valuable.
- •Guillermo argues the workforce trend is toward training agents rather than directly doing tasks, and imagines systems that extract “skills” from inputs/outputs.
- •Blake and Naval predict more entrepreneurship and a “very large number of small teams” as required team size drops; Naval closes by emphasizing creativity, taste, and getting good with tools as the key differentiator moving forward.

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.