Steve Yegge: Big Tech Is Quietly Dying

Why Steve Yegge Thinks the Big Companies Are Already Dead

Steve Yegge — 40-year software veteran, former Google/Amazon engineer, now building the open-source agent orchestrator Gas Town — sat down with Gergely Orosz on The Pragmatic Engineer podcast for a sweeping conversation about what AI is doing to the software industry. His central claim: big tech companies are dying, they just don’t know it yet.

On big companies being dead already: “We’re all looking at the big companies going, ‘When are you going to give us something?’ And the answer is we’re looking at the big dead companies. We just don’t know they’re dead yet.” Yegge’s argument isn’t that these companies lack productive engineers — it’s that the organizations themselves can’t absorb the output. Engineers hit downstream bottlenecks (design, legal, compliance, release processes) and quit. Meanwhile, 2-20 person startups are shipping faster than entire divisions.

On the 8 levels of AI adoption: Yegge maps engineers on a spectrum from Level 0 (no AI at all) to the highest level where multiple parallel agents run simultaneously on your codebase. He estimates roughly 70% of engineers are still at the Copilot level or below — using basic autocomplete but not agentic workflows. The gap between those engineers and the ones running 80 parallel agents is, in his words, isomorphic to managing a team of 80 engineers.

On “vampire burnout”: Even the most productive AI-augmented engineers hit a wall. “You can be 100x productive but you only have about 3 good hours a day.” The bottleneck isn’t the AI — it’s human decision-making capacity. You’re reviewing, directing, and course-correcting constantly. Three hours of intense agent supervision is exhausting.

On “heresy” — the new tech debt: Yegge introduces a concept for vibe-coded codebases: a “heresy” is a wrong architectural idea that takes root among your agents and keeps coming back. “You try to get them all out, but there will be one reference to it in some doc somewhere that an agent picks up on and goes, ‘Oh, that makes sense.’ And it returns and rebuilds the heresy and it starts to spread again.” The fix: document heresies explicitly in your prompts and add tooling to prevent agents from rebuilding them.

On token burn as the key metric: For startups using AI agents, Yegge argues the single most important proxy metric is token burn — how much compute you’re consuming. Startups no longer ask how many employees they need; they ask how much compute they can reserve. “Make your token burn as high as your investors will let you go,” he advises, because that spend represents practice, iteration, and learning.

On the value capture problem: If you become 100x more productive, who benefits? If you work 8 hours and produce 100x the output, the company captured all that value. If you work 10 minutes and produce the same output as before, you captured all of it. Neither extreme is sustainable. Yegge acknowledges we don’t have cultural norms for this yet — it’s going to be messy.

On the bitter lesson and future models: Referencing Richard Sutton’s famous essay, Yegge argues you should never try to be smarter than the AI — bigger models always win. He estimates at least “two more cycles” of capability improvement remain, meaning models at least 16x smarter than today. That, he says, “is going to cause all of knowledge work to be subsumed.”

On personal software and forking: Everyone will want bespoke software. Forking — once a “declaration of war” in open source — becomes an everyday act when AI makes maintenance trivial. Programming itself becomes for everyone: Yegge predicts his non-developer wife will be the top contributor to their video game by summer 2027.

8 Key Takeaways from Steve Yegge on AI Agents

• Big tech is a dead man walking — Organizations can’t absorb the productivity their AI-augmented engineers produce, creating bottlenecks that drive talent away
• 70% of engineers are still below the agentic threshold — Using Copilot-level tools while a small minority runs 80 parallel agents
• Vampire burnout is real — 100x productivity but only 3 good hours/day of decision-making capacity
• “Heresy” is the new tech debt — Wrong ideas that spread among agents and keep rebuilding themselves in your codebase
• Token burn is the metric — Startups measure compute consumption, not headcount
• The bitter lesson applies — Don’t try to outsmart the AI with hand-crafted heuristics; scale always wins
• Value capture is unsolved — 100x productivity gains need new cultural norms for who benefits
• Programming is for everyone — Non-developers building serious software by 2027, forking becomes trivial and routine

What This Means for Organizations Using AI Agents

Yegge’s most actionable insight is that organizational structure — not model capability — is now the binding constraint. The engineers are productive. The models are capable. But companies built for the old world can’t move fast enough to use what AI gives them. The startups that win will be the ones designed from scratch for agentic workflows: small teams, transparent processes, high token burn, and no downstream bottlenecks. If your company still requires PRs, design reviews, and multi-sprint planning cycles for features an agent can ship in hours, you’re the big dead company that doesn’t know it yet.