The AI Industrial Revolution: How Frontier Founders Build at Scale
Three frontier founders—building AI cloud infrastructure, supersonic aircraft, and brain-computer interfaces—reveal how AI transforms engineering, manufacturing, and knowledge work. The shift from shipping code to building factories, from junior engineers to 100x-1000x leverage, and from pre-approval regulation to enforcement-based systems reshapes what's possible for small teams.
The Engineer's New Job: Building Factories, Not Features
From Output Shipping to Factory Building
The role of engineers has shifted from directly shipping output to building systems that produce multiplicative outputs. Instead of measuring individual productivity, teams now measure whether an engineer creates infrastructure that enables others to produce outputs B through Z.
100x and 1000x Engineers Are Now the Norm
In intellectual and digital domains, the productivity multiplier is no longer 10x but 100x or 1000x. Historical examples like Satoshi Nakamoto (Bitcoin), Brendan Eich (JavaScript), and John Carmack demonstrate that extreme leverage has always existed in software; AI now makes it accessible to more builders.
Token Consumption Is Not the Right Metric
Measuring AI productivity by token consumption is like measuring software productivity by lines of code—a misleading proxy. The real metric is time saved and output quality, not raw token spend.
How AI Models Amplify Existing Skill
Models Reflect Back the Judgment of the User
Claude, ChatGPT, and GPT are roughly as good as the person using them in their domain. A capable developer gets powerful results; a junior developer gets junior-level output. The quality of feedback and reprompting is critical at this stage.
Models Now Plan and Present Tradeoffs
Recent models have evolved from simple next-token prediction to intuitive planning mode. They now return to the user with multiple routes, tradeoffs, and architectural considerations—behaving more like principal engineers than junior engineers.
Waste Tokens, Save Time
Rather than optimizing token usage, throw multiple models at the same problem and waste tokens to save time. Tokens are cheaper than human time, and models improve every generation, so brute-force approaches are economically rational.
The Feedback Loop Removes Frustration
With AI agents, you no longer get stuck on narrow debugging problems for indefinite periods. Agents quickly find the right approach, removing the intrinsic frustration that used to be part of learning to code.
Hardware Engineering Transformed by Software Thinking
Hardware Engineering Workflows Treated as Software
Traditional hardware engineering happens in siloed Excel spreadsheets with VBScript, no source control, and manual handoffs via email. By converting these workflows into software frameworks with automation and testing, iteration costs drop dramatically.
Two Engineers Can Now Design an Entire Jet Engine
At Boom Supersonic, software engineers create system architectures while hardware engineers 'vibe code' their pieces. Real-time feedback between aerodynamics and structures analysis—a task that took one engineer one day per blade—now happens instantly, enabling two people to design a complete engine.
AI Will Soon Generate CAD and PCB Layouts
Currently AI generates software; within the next year it will generate step files and PCB layouts, bringing the same productivity gains to mechanical and electrical engineering that software engineering is experiencing.
Open-Source Models and the Hardware Advantage
China's Open-Source Strategy Levels the Playing Field
China is investing heavily in open-source models because they have hardware superiority and complex supply chains. By generating software on demand, they eliminate the disadvantage against Silicon Valley. Open AI, Google, and Anthropic lag in open-source offerings, leaving the field to Chinese models.
Frontier Intelligence Dominates for Coding, But Specialized Models Excel Elsewhere
Frontier models (OpenAI, Anthropic) dominate for coding tasks. However, models like Gemini excel at industrial production tasks—support, browser automation, and other non-coding work. Chinese models fill niches but don't compete on frontier coding.
Intelligence Is an Unalloyed Good
When choosing between models, always pick the most intelligent one available. You often don't know which answer is correct, so the smarter model's judgment is more reliable. This creates monopoly/oligopoly pressure but is economically rational.
Vertical Integration and the Limits of Off-the-Shelf
When Components Don't Exist, Build Them Yourself
Science Corp owns a captive MEMS foundry because the packaging and assembly needed for their brain-computer interface doesn't exist off-the-shelf. Vertical integration is necessary when innovation requires integration beyond what vendors provide.
Regulatory Documentation Can Be Automated
AI dramatically reduces regulatory burden by automating compliance documentation. A 200-page lightning-strike test plan that took months can now be generated in minutes using RAG. This removes change aversion and enables rapid iteration.
Instrumentation Enables AI-Driven Optimization
By instrumenting foundries and manufacturing processes, companies create feedback loops that AI can optimize. As models improve, these optimizations compound in cell engineering, material science, and production efficiency.
Regulation as a Test Suite, Not a Barrier
Regulations Can Be Reframed as Guard Rails
Rather than viewing regulation as friction, frame it as a test suite. Agents can be tasked with complying with all regulations as exit criteria, turning regulatory requirements into automated guard rails that prevent shipping slop.
Pre-Approval vs. Enforcement-Based Models
Current system: guilty until proven innocent (pre-approval). Better model: enforcement-based, where you build and regulators catch problems. This mirrors how driving works—you don't submit a plan to the fire department before leaving your house.
The Asymmetric Incentive Problem in Regulation
Regulators face asymmetric incentives: approving a bad thing ends their career; blocking a good thing goes unnoticed. This creates structural slowdown. AI-driven compliance could reduce this friction by automating the verification layer.
Healthcare's Reimbursement Model Blocks Innovation
Unlike phones and laptops (where lower costs drive higher volume and total spending increases), healthcare has a fixed reimbursement bucket. Spending more on healthcare is seen as catastrophic, not progress, blocking the feedback loop that drives technological growth.
China's Lower Approval Costs Enable Market Competition
China brings medical devices to market at 1/10th the cost of the US because approval is faster and cheaper. This allows them to sell at $10,000 instead of $100,000, creating a private market where US healthcare has none.
Software Engineering: Dead or Evolved?
Pure Software May Be Obsolete
If models speak English and understand fuzzy human intent, the moat for pure software companies erodes. The advantage shifts to hardware founders (who can now build software easily) and to those training and fine-tuning models.
Infrastructure and Building Blocks Are the New Moat
Agents need reusable, right-sized building blocks (like message queues, databases, APIs) rather than reinventing infrastructure from first principles. Infrastructure software and standardized components become the valuable layer.
Spreadsheets Are Cooked
Spreadsheets succeeded because no one could build custom software. Now that AI can generate software on demand, Python models and custom tools replace spreadsheets for simulation and analysis.
Lawyers and Paralegals Got a Promotion
Junior lawyers and paralegals handling basic legal tasks (NDAs, agreements, research) have been automated away. The upside: they can now focus on higher-level legal thinking, or they've been promoted to senior-level work.
Autonomous Systems and the Verification Layer
Humans Become Verifiers, Not Executors
As AI generates code, documentation, and decisions, humans shift from executing work to verifying it. The new skill is signing off on correctness without reading every line—using test harnesses, simulations, and type checkers as confidence signals.
Autonomous Anomaly Detection and Remediation
Agents can detect infrastructure anomalies, investigate root causes, file incidents, and begin remediation—all without human intervention. This automates the on-call SRE job while keeping humans in the loop for final decisions.
Autonomous Security Research at Scale
Running 10,000 concurrent agents for security research can accomplish months of red-teaming work in days for ~$14,000 in tokens. This enables proactive security investment that would be impossible with human teams.
Agents Can't Yet Execute Outside Sandboxes
Agents still need humans to provide API keys, capital, and access to external systems. This is a temporary limitation; soon APIs will be text-based and agents can pay with crypto for what they need.
The Autonomous Company Experiment
Hackathon Results: Needle Movers Over Silly Projects
When a company stopped all work for a week and asked everyone (from receptionist to engineers) to build something using AI, the result was mostly needle-moving projects, not frivolous ones. The receptionist built an automation for package tracking that the company now uses.
The Barrier to Entry Is Imagination, Not Execution
Most people have ideas for things that would improve the world, but they don't pursue them because they assume execution is hard. When execution becomes trivial (via AI), the barrier becomes having the idea and the agency to iterate.
Vibe Coding Is More Addictive Than Traditional Programming
Building with AI feels like a video game with real-world output. The feedback loop is tight, and you don't get stuck. People who haven't coded in decades are now building software; people are choosing vibe coding over video games for entertainment.
Skill Extraction and Agent Training
Companies can now extract skills from employee work (inputs and outputs) and use those to train agents. This creates a flywheel where agents learn from human expertise and humans are freed to focus on creative work.
The Future of Work: Agency Over Intelligence
Returns Shift from Intelligence to Agency
Historically, returns were 70% intelligence and 30% agency. In the future, this inverts: 70% agency, 30% intelligence. The ability to act, iterate, and make decisions matters more than raw intelligence.
Coding Population Grew 10x in One Year
The percentage of people coding has likely increased 10-fold in the past year due to AI. However, 99% of the population still doesn't code, and many don't realize how much easier it's become.
Smaller Teams, More Companies
The number of people needed to accomplish a given task drops dramatically. Instead of one large company with 1,000 engineers, you get 100 small companies with 10 engineers each. Entrepreneurship explodes.
Generalists Thrive; Credentials Lose Value
Expertise and credentials matter less when AI provides domain knowledge and jargon translation. Generalists who can think across domains, have good taste, and can iterate rapidly outperform narrow specialists.
Productivity Gains Mean More Hiring, Not Less
Basic economics: when productivity increases, wages rise and more people are hired, not fewer. The question is whether people adapt to new roles (creative, agentic) or resist change.
Creativity, Art, and the Role of Humans
Art Requires Surprise and Intent
Art is meaningful out-of-distribution behavior—something that surprises you and changes your trajectory. Computers lack intent, so they can't create art in the human sense, though they can generate impressive outputs.
Humans Generate Surprise; AI Generates Distribution
Humans can step outside the training distribution and generate truly novel ideas (like Gödel's incompleteness theorem). AI generates outputs within the learned distribution, no matter how impressive.
Attribution and Verification Matter
A photo taken by a human has more meaning than an AI-generated identical photo because of human intent and attribution. Startups are building hardware-based verification to prove humans took photos.
AI Will Drown Us in Slop
As AI generation becomes trivial, the internet will fill with low-quality, derivative content. The bar for surprise and meaning will rise, making truly novel human creativity more valuable.
Humans Plus AI Is the Winning Formula
Pure AI isn't there yet for true creativity. The era of human-plus-AI will last longer than people think. Humans provide intent, taste, and out-of-distribution thinking; AI provides execution and leverage.
Video and Media Generation Lags Code Generation
While AI code generation is mature, AI video and media generation still lack taste and judgment. Generating a movie from a book or creating original stories at Studio Ghibli quality remains aspirational.
The Immediate Imperative: Get Good with AI
AI Skill Is the New Baseline
The single best thing anyone can do right now is get comfortable with frontier AI models, understand their boundaries, and stay aware of what they can and can't do. This is a moving target as models improve.
Hiring Criteria Shift to AI Fluency
Companies are hiring for people who are really good with agents and AI, quick to adapt, and creative. Juniors and super-seniors are equally valued if they have these skills.
Automate Everything Repetitive
If you're doing the same thing twice, automate it. The goal is to work in your maximum zone of creativity and interest, with AI handling all repetitive work.
Notable quotes
We used to believe there's 10x engineers. Now clearly there's 100x or a thousandx engineers. — Gumo
Just waste tokens, save time. Don't look at the tokens either as inputs or outputs. — Gumo
The bar is going to be raised massively. It's going to take more and more to surprise you. — Naval
Action items
- Get comfortable with frontier AI models (Claude, ChatGPT, GPT) and understand their current capabilities and limitations.
- Identify repetitive tasks in your workflow and automate them using AI agents.
- Experiment with 'vibe coding'—building software by directing AI rather than writing code by hand.
- If you manage engineers, shift hiring criteria from credentials to AI fluency and adaptability.
- Frame regulatory requirements as test suites and guard rails rather than barriers.
- Extract and document skills from your work to train agents that can replicate your expertise.
- Consider vertical integration for components that don't exist off-the-shelf and are critical to your product.
- Run autonomous security research and anomaly detection in your infrastructure.
- Experiment with small, autonomous teams to accomplish tasks that previously required large groups.