Steve Yegge is a 40-year software engineering veteran (Amazon, Google), known for his influential blog posts and brutal honesty about the industry. He recently built Gas Town, an open-source AI agent orchestrator, and co-authored the book Vibe Coding with Gene Kim. This conversation covers his framework for AI adoption, why big tech is quietly dying, the physical and psychological toll of AI-augmented work, and where software engineering is headed.
The eight levels of AI adoption for engineers
Steve defines a spectrum of AI adoption that most engineers are navigating:
Level 1: No AI at all.
Level 2: Using AI as a yes/no assistant inside the IDE (“can I do this thing?”).
Level 3: YOLO mode — telling the AI to just do it, trust is building.
Level 4: Squeezing code out of the agent, focusing more on the conversation than reviewing diffs.
Level 5: The agent does the work, the engineer reviews code in the IDE later — no longer coding directly in the IDE.
Level 6: Bored because the agent is busy, so you spin up another agent. You quickly reach an equilibrium where there’s always an agent waiting for you, and you’re multiplexing between them.
Level 7: Chaos — agents conflict, make messes, work on the wrong things. This is the motivation for orchestration.
Level 8: Orchestration — agents running agents, which is what Gas Town does.
Steve estimates 70% of engineers are still at levels 1 or 2, using tools like Copilot and occasionally asking questions, which he considers dangerously behind.
The single most important proxy metric for a company is token burn — it signals that people are experimenting, failing, and learning.
Gas Town: an agent orchestrator
What it is: Gas Town is an open-source orchestrator — agents running agents. You talk to a “mayor,” which dispatches “workers” (sub-agents called polecats and crew) to do tasks.
Two worker roles reflect a fundamental tension in AI workflows:
Polecats (minimize context): For well-specified, self-contained tasks. Short context windows reduce cost and keep cognition sharp.
Crew (maximize context): For hard design problems requiring deep conversation. Loaded with documentation and history.
Steve built Gas Town partly to move the Overton window — to make orchestration and swarms a credible conversation rather than dismissed as impossible.
It’s currently too hard for most people to use (command-line-driven, invisible internals). Steve predicts that by the end of 2026, most people will program by talking to a face on screen — a visual, conversational interface that hides the complexity.
Gas Town is migrating to Dolt, a Git-backed database, reflecting Steve’s interest in version-controlled data.
Steve doesn’t expect Gas Town to last more than 6 months before being supplanted, but it’s serving as a research vehicle for what comes next.
The vampiric effect of AI on developers
AI creates a vampiric burnout effect: it gets you excited, you work incredibly hard, and you capture enormous value — but it drains you fast.
Engineers at startups are finding themselves napping during the day from mental exhaustion, even though they’re vastly more productive.
The core problem: companies are set up to extract value until you break. If an engineer becomes 100 times more productive, the company captures all of that value if they work 8 hours. But the engineer’s battery drains much faster — Steve estimates you might only get 3 truly productive hours per day at max vibe-coding speed.
The new work-life balance question is: how much value do you capture for yourself vs. pass to your employer? There are no cultural norms for this yet.
Steve draws a parallel to the 996 culture (9am–9pm, 6 days/week) common in Southeast Asia — some AI startups are drifting toward this, while others capture all the value by working 10 minutes a day and telling no one.
Big tech is quietly dying
Steve’s provocative claim: innovation at large companies is now dead, and we’re only going to see innovation from small teams of 2–20 people.
His evidence: Google’s golden age ended around 2008 when Larry Page became CEO and said “put more wood behind fewer arrows.” Before that, there was more work than people; after, there were more people than work, leading to land grabs, territoriality, and politics.
Amazon laid off 16,000 people while blaming AI, despite having no clear AI strategy. Steve’s explanation: companies have a dial for what percentage of engineers they need to cut to afford AI tokens for the rest. The default setting is about 50%.
Big companies can’t absorb the output of hyper-productive engineers — downstream bottlenecks shut them down and they quit.
Anthropic is the counter-example: it operates as a “hive mind” with infinite work, everyone over-subscribed, and a culture of pure functional data structures (“yes, and” rather than “yes, but”). Claude Cowork was prototyped and launched in 10 days.
Steve’s advice to engineers at companies still on Copilot: “Get out.” Find 30 minutes a day to play with Claude Code or similar tools. Proof of work (visible, forkable, cloneable) is becoming more important than resumes.
Vibe coding and its debts
Vibe coding (a term from Steve’s book) means writing code through conversation with AI rather than by hand. The back cover of his book says “the days of coding by hand are over” — written over a year ago, before Opus 4.5.
FAFO (one of the book’s frameworks) stands for the dimensions of value from vibe coding, with the “O” being optionality — the ability to create many prototypes quickly and defer decisions until you know the right answer. Boris Cherney at Anthropic built 20 working prototypes in 2 days for a single feature (“slot machine programming”).
Vibe coding debt is real. Steve calls architectural errors that take root in AI-generated code “heresies” — wrong ideas about data flow or architecture that agents keep rebuilding even after you try to remove them. They’re invisible, hard to weed out, and can resurface from a single reference in documentation.
The ceiling for what AI can build before dissolving into chaos is rising — currently somewhere between 500,000 and 5 million lines of code. Monolithic codebases are the biggest obstacle; AI will never fit them in a context window, so companies need to break them up or rewrite.
The Bitter Lesson and the future of AI capability
The Bitter Lesson (Richard Sutton): don’t try to be smarter than the AI. Every time researchers tried to inject human domain knowledge, they were eventually overtaken by approaches that simply used more compute and data. Bigger is smarter.
Steve believes there are at least two more cycles left in AI scaling, meaning models will be at least 16 times smarter than today. This will subsume all knowledge work.
He acknowledges the curve is S-shaped and will eventually flatten when resources run out, but we’re not near that point yet.
Model release cadence is accelerating: from ~4 months between drops to ~2 months. Each new model fixes the bugs people complained about in the previous one.
What happens to the craft of software engineering
Steve’s 2012 blog post “Rich Programmer Food” argued that you can’t be a good programmer without understanding compilers. He now says that’s no longer true — the abstraction ladder has moved up, and what engineers need to know keeps changing.
Language design still matters for optimization and energy efficiency, but for everyday work, it won’t matter what language you pick. There may even be purpose-built languages designed by AIs for AIs.
Debugging is an open question — agents currently just use print statements. They may wake up to proper debuggers in 6 months, or we may not need debuggers anymore.
Developer workstations will shift to servers and mobile devices. Steve wants to do serious work from his phone. Local laptops are already stressed by running multiple agents.
Open source is being transformed: forking is becoming an everyday occurrence rather than a declaration of war. People remix projects with AI without asking permission. Cursor itself is a fork.
Steve went through a grief period (about a week of intense sadness, a few months total) where he mourned the skills that made him special — memorization, writing, computing. He came out the other side writing 10 times more code and having more fun.
Predictions for 2027
Steve’s wife (not a developer) will be the top contributor to their video game project, because programming is becoming accessible to everyone through AI.
His kids treat AI as normal — enhancing photos with squirrels on heads, creating things through imagination. This is the future.
We’ll need ecosystems of agents that search and curate the explosion of new software, just like search engines were needed when everyone could make webpages.
The biggest business opportunity right now: build agents that know how to find and surface the good stuff in the coming flood of software.
Steve is fundamentally optimistic: “It’s all going to work out.” The future is more fun than the present; we’re just holding on to the old.
