From IDEs to AI Agents with Steve Yegge
Steve Yegge has spent 40 years writing code, but today he barely looks at it. Instead, he runs a «factory» of AI agents that write thousands of lines while he naps. He believes 70% of engineers are still stuck using autocomplete while a new class of developers coordinates swarms of autonomous workers, and that big tech companies are quietly dying because they can't absorb the productivity gains their own engineers are achieving. The question isn't whether AI will replace your job — it's whether you'll learn to capture the value of being 100× more productive, or let your employer take it all. Yegge has drawn an eight-level ladder from no AI to orchestrating parallel agents, and he's convinced most people are dangerously far behind.
Ключевые выводы
There are eight levels of AI adoption for engineers, from no AI to running multiple agents in parallel, and roughly 70% of developers are still stuck at levels one or two, using basic autocomplete or asking yes/no questions in their IDE.
AI creates a «vampiric burnout effect»: engineers can be 100× more productive but may only get three good hours of deep thinking per day, and companies that try to extract eight hours of that intensity will break their teams.
Innovation at large companies is dying because they have more people than work; small teams using AI orchestrators can now rival the output of Fortune 500 engineering organizations, and we're entering a land rush where 2–20 person startups will disrupt incumbents.
The new work-life balance question is value capture: if you're 100× more productive, who benefits? Working 8 hours and producing 100× output means the company captures all the value; working 10 minutes means you do. Neither extreme is sustainable, and we lack cultural norms to navigate this.
Code written by AI agents accumulates «heresies» — incorrect architectural ideas that take root invisibly and spread like weeds — and the only fix is to document them explicitly in prompts or wait for the next model drop to be smart enough to avoid them.
Вкратце
Software engineering is undergoing the same abstraction leap that graphics went through in the 1990s, and engineers who don't move up the AI adoption ladder — from autocomplete to agents to orchestration — will be left behind as small teams of 2–20 people start to rival the output of Fortune 500 companies.
The Eight Levels of AI Adoption
Yegge maps developer AI usage from zero to orchestrating parallel agents.
Level 1: No AI You write all code by hand, no assistance.
Level 2: Yes/No Questions You ask your IDE «Can I do this thing?» and carefully review every line it suggests.
Level 3: YOLO Mode Your trust is growing; you let the agent do more without constant oversight.
Level 4: Conversation Over Code You focus on talking to the agent, not reviewing diffs. The code is being squeezed out of view.
Level 5: Agent-First Workflow You work entirely in the agent interface and only look at code in your IDE later.
Level 6: Multiplexing Agents You're bored waiting for one agent, so you spin up multiple agents and switch between them as they finish tasks.
Level 7: Coordination Chaos You've made a mess — agents collide, you texted the wrong one, and now you're debugging a project inside a project.
Level 8: Orchestration You build tooling (like Gas Town) to coordinate agents, assign roles, and manage parallel workflows at scale.
The Vampiric Burnout Effect
AI makes you vastly more productive but drains your cognitive battery faster.
The Vampiric Burnout Effect
Yegge and his peers are finding themselves napping during the day despite being «100 times more productive». The easy work is automated, so engineers now spend all their time on hard, system-2 thinking. Companies set up to extract maximum value will push engineers until they break, but at max vibe-coding speed, you might only get three productive hours. The new work-life balance is figuring out how much of that 100× gain you capture versus how much your employer does.
Why Big Tech Innovation Is Dead
Large companies have more people than work and can't absorb AI gains.
Yegge argues that Google stopped innovating around 2008 and has only acquired technology since. The turning point was when Larry Page became CEO and said «put more wood behind fewer arrows» — suddenly there were more people than projects. Engineers started fighting over work, leading to land grabs, backstabbing, and empire building. Yegge's friend at Amazon once said they avoid Google's problems because «everyone is always slightly oversubscribed».
Now AI is creating a paradox: engineers are vastly more productive, but large organizations can't absorb the output. They hit bottlenecks in legal, compliance, product, and process. Meanwhile, small teams of 2–20 people using orchestrators like Gas Town can rival the output of Fortune 500 companies. Yegge believes we're watching big tech companies die quietly, and a «land rush» of AI-native startups will displace them. The future of software may not come from Amazon or Google — it will come from a college kid running agents in their closet.
What Is Gas Town?
The Bitter Lesson and the Curves
Don't try to be smarter than the AI; scale wins every time.
“The bitter lesson is don't try to be smarter than the AI. You think that you've got special knowledge, special domain knowledge to this problem, and we're going to teach it so that the AI will be smarter. What we found was bigger is smarter. Always. More data.”
Key Numbers from the Frontier
Model capability ceilings, training scale, and half-lives are all accelerating.
The New Rules of Building Software
Transparency, forking, and prototyping until you ship are replacing old workflows.
What Cannot Be Cloned
Human connection will be the moat when software becomes trivially replicable.
If any software can be trivially cloned by an AI, what's left to compete on? Yegge believes human connections will become the primary moat. As automation increases, people will paradoxically want humans to do things — to curate, to deliver, to touch. He also believes we'll see an explosion of personal bespoke software, where everyone has their own custom apps built by agents. The innovation won't come from Walmart or Microsoft; it will come from random individuals and 2–20 person startups.
Yegge predicts that by the end of 2025, most people will program by talking to a face on a screen — an AI persona like a fox or a mayor who spins off workers invisibly. This is because most people can't or won't read the walls of text that agents produce today. The UIs that make orchestration accessible to non-readers will unlock the next wave of software creation.
Prediction: Your Family Will Contribute More Code Than You
Non-developers will out-code developers by mid-2026, starting with Yegge's wife.
Prediction: Your Family Will Contribute More Code Than You
Yegge's bold prediction for 2027: his wife — a non-developer — will be the top contributor to their family video game by summer. He believes programming will become an activity for everyone, not just engineers, and that the explosion of user-generated software will require new ecosystems of agents to search, curate, and surface the good stuff. If you want to build a big business right now, he says, go build agents that know how to find experiences people will love in the coming flood of AI-generated content.
Люди
Глоссарий
Отказ от ответственности: Это ИИ-сводка видео с YouTube, подготовленная в образовательных и справочных целях. Она не является инвестиционной, финансовой или юридической консультацией. Всегда проверяйте информацию по первоисточникам перед принятием решений. TubeReads не связан с автором контента.