Steve Yegge Says We Are All Doomed and I Am Not Sure He Is Wrong

Steve Yegge on AI Agents and the Future of Software Engineering

TLDR: Steve Yegge, a 40-year software engineering veteran who worked at Amazon, Google, and Grab, sat down with Gergely Orosz on The Pragmatic Engineer to deliver a scorching take on where AI is taking our industry. He argues we are on a steep exponential curve that shows no signs of flattening, big companies are effectively dead, and roughly half of all engineers at large organizations will be laid off. But he also insists that AI is an augmentation tool, not a replacement, and that building software has never been more fun.

Summary:

Let me paint you a picture. You have a guy who spent four decades writing code at some of the most demanding engineering organizations on the planet. He built compilers, debuggers, and worked on systems at Amazon and Google scale. And now he is telling you, with a straight face, that the days of coding by hand are over. That is Steve Yegge, and whether you agree with him or not, you should probably pay attention to what he is saying.

The conversation starts with the S-curve argument. Yegge admits he was initially skeptical of large language models, but his skepticism evaporated the moment he tried Claude Code. His exact quote was, "oh, I get it. We're all doomed." He points to the shrinking half-life between model releases, from four months down to two months, and argues that every bug and mistake gets fed back as training data. The models keep getting better, and there is no indication the curve is flattening anytime soon. Now, I want to push back on this a bit, because every technology has an S-curve and the people riding the steep part always think it will go on forever. But Yegge is not just some hype merchant. He spent a year reading AI papers to understand the fundamentals before making these claims, and he has Dr. Erik Meijer, one of the most important compiler people in the world, backing up the same conclusion.

Then comes the really uncomfortable part: the "50% dial." Yegge argues that every large company now has a mental dial they are turning, deciding what percentage of engineers to let go so they can fund AI tooling for the rest. His estimate is that roughly half of all engineers at big companies will be cut, which would dwarf the pandemic-era layoffs. The reasoning is blunt: if you want the remaining engineers to be maximally productive with AI, you need to pay for their token usage, and that money has to come from somewhere. He points to Amazon laying off 16,000 people and blaming AI without even having an AI strategy. That is the part that should genuinely worry people, because it suggests the layoffs are happening regardless of whether companies have a coherent plan for what comes next.

Yegge also introduced an eight-level spectrum of AI adoption that is both amusing and terrifying. Level one is no AI at all. Level two is having a coding agent in your IDE. By level five, you have abandoned the IDE entirely and just talk to the agent. At level six, you are running multiple agents simultaneously because you get bored waiting for one to finish. By level eight, you are building your own orchestrator to coordinate ten or more agents. His point is that engineers stuck at levels one through three are going to be left behind, and some of the best engineers he knows personally are sitting at level two, carefully reviewing every line of AI-generated code. He feels sorry for them because he thinks they are going to get fired.

One of the most interesting insights is what Yegge calls the "Dracula effect." Vibe coding at full speed is physically and mentally draining. He and his friends at startups find themselves napping during the day, getting tired and cranky. His argument is that companies should not expect more than three productive hours of AI-augmented work per day, because even at that rate, engineers are still producing a hundred times more output than before. This challenges the assumption that AI tools should simply multiply output across an eight-hour day. The reality is that the cognitive load of directing multiple AI agents, maintaining context, and making rapid decisions is exhausting in a way that traditional coding never was.

Finally, Yegge makes the bold claim that big companies are essentially already dead, they just do not know it yet. Innovation at large organizations has effectively stopped because even when they have hyper-productive engineers, the companies themselves cannot absorb the output. There are bottlenecks downstream, politics, process, approvals, and all the institutional friction that exists at scale. The real innovation, he argues, will come from small AI-augmented teams, much like how cloud computing shifted the balance of power away from enterprises that owned their own data centers. This is probably the most debatable point in the entire conversation, because big companies have survived every previous technological disruption by simply acquiring the small companies that innovate. But the speed of this particular shift might change the equation.

Key takeaways:

• The exponential improvement curve in AI models shows no signs of flattening, with the half-life between major model releases shrinking from four months to two months
• Roughly 50% of engineers at large companies may be laid off as organizations redirect salary budgets toward AI tooling and token costs
• Yegge defines eight levels of AI adoption, from no AI to building your own agent orchestrator, and warns that engineers stuck at low levels risk being left behind
• The "Dracula effect" means vibe coding is physically draining, and companies should expect only about three productive AI-augmented hours per day from engineers
• Big companies cannot absorb the output their most productive engineers create, leading to bottlenecks that may push real innovation to small, agile teams
• AI is an augmentation function, not a replacement function, but the demand for software will keep growing even as the way we build it changes fundamentally

Tradeoffs:

• Yegge's S-curve argument assumes continuous improvement, but every technology eventually hits diminishing returns. The question is whether that happens in six months or six years, and the difference matters enormously for career planning.
• The "50% dial" framing assumes companies will act rationally and invest the savings in AI tooling. In practice, many companies will simply cut costs and pocket the difference without making their remaining engineers more productive.
• The eight-level adoption framework creates a false sense that higher is always better. An engineer carefully reviewing AI output at level two might produce more reliable software than someone at level seven with ten uncoordinated agents creating merge conflicts.
• The claim that big companies are dead ignores their massive advantages in data, distribution, and capital. Startups building with AI still need customers, and big companies still own the customer relationships.

Steve Yegge on AI Agents and the Future of Software Engineering