Simon Willison hasn’t taken a proper break in three years. The co-creator of the Django web framework — one of the most widely deployed tools in modern software development — recently admitted something that surprised even his most devoted followers: he’s exhausted. Not the normal kind of tired that comes from shipping code on deadline. Something deeper. Something structural.
“I am mass-subscribing to feeds and trying to keep up with everything because I genuinely cannot afford to miss things,” Willison told Business Insider in a candid interview. The problem isn’t that AI is moving fast. It’s that it never decelerates. Every week brings new models, new capabilities, new frameworks that render last month’s best practices obsolete. For engineers like Willison, who has built his reputation on understanding and explaining these tools through his popular blog, the pace has become something close to unsustainable.
This isn’t a story about burnout in the conventional sense. It’s about an entire professional class — AI engineers, researchers, developer advocates, and technical leaders — caught in a cycle of perpetual obsolescence. The tools they master today may be irrelevant in six months. The architectures they design this quarter could be superseded by the next foundation model release. And the knowledge they accumulate doesn’t compound the way it used to. It depreciates.
Willison has been remarkably transparent about this. He’s described his daily routine as a relentless effort to monitor new releases, test new models, and document his findings — all while building his own AI-powered tools like his open-source project LLM and the Datasette platform. He’s not complaining, exactly. He’s diagnosing. And his diagnosis resonates with thousands of practitioners who’ve been quietly feeling the same thing but lacked the vocabulary — or the professional security — to say it out loud.
The numbers tell part of the story. OpenAI, Google DeepMind, Anthropic, Meta, and a growing roster of Chinese labs are shipping major model updates on overlapping timelines. In 2024 alone, the industry saw the release of GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama 3.1, and dozens of smaller but consequential open-weight models. The cadence has only intensified in 2025 and into 2026. Each release doesn’t just add features — it reshapes what’s possible, which means engineers must reassess their assumptions constantly.
“Every time a new model drops, I have to re-evaluate what I thought I knew,” Willison said, as reported by Business Insider. That re-evaluation isn’t trivial. It means running benchmarks, testing edge cases, updating documentation, and sometimes rewriting tools from scratch. Multiply that across a dozen major releases per year, and you begin to understand why even the most energetic engineers are hitting walls.
The exhaustion Willison describes has a specific character. It’s not the fatigue of doing hard work. It’s the fatigue of never finishing. Traditional software engineering has natural stopping points — you ship a release, you stabilize, you document, you move on. AI development in 2026 has no such rhythm. The ground shifts continuously. There’s no plateau.
And it’s not just about keeping up with models. The tooling layer is equally volatile. Frameworks for building AI agents, retrieval-augmented generation pipelines, fine-tuning workflows, and evaluation harnesses are proliferating at a rate that defies any single person’s ability to track them all. Willison, who has arguably done more than anyone to catalog and explain these tools to a broad audience, has essentially acknowledged that even he can’t keep pace.
This matters beyond individual well-being. It matters for the industry.
When the most experienced engineers are overwhelmed, the knowledge gap between experts and everyone else widens. Junior developers, who might normally learn from senior practitioners, find that their mentors are too busy re-learning to teach. Companies building on AI foundations discover that their technical leads are making architectural decisions based on information that was current three months ago — an eternity in this field. The result is a kind of institutional fragility, where organizations are building on shifting sand and hoping the next model release doesn’t invalidate their core assumptions.
Willison’s case is particularly instructive because he isn’t employed by a major AI lab. He’s an independent developer and writer, which means he has more freedom than most to set his own pace. That he still feels crushed by the velocity says something profound about the structural demands the field places on its practitioners. If Willison can’t manage it sustainably, what chance does a mid-career engineer at a Fortune 500 company have — someone juggling AI integration with legacy systems, compliance requirements, and quarterly OKRs?
The broader tech industry has faced waves of rapid change before. The mobile revolution of 2007–2012. The cloud migration era. The DevOps transformation. But veterans of those transitions point to a key difference: those shifts had discernible phases. You could learn iOS development in 2009 and your skills would remain relevant for years. You could master AWS in 2014 and build on that knowledge incrementally. AI in 2026 doesn’t work that way. The underlying capabilities of the models themselves keep expanding, which means the entire application layer must be continuously reconsidered.
Some companies are responding by creating dedicated roles for what might be called “AI currency” — staying current. These are engineers whose primary job isn’t building products but rather monitoring the frontier and translating new developments into actionable guidance for their teams. It’s an acknowledgment that keeping up has become a full-time job in itself. But it’s also an expensive solution, and one that only well-funded organizations can afford.
Others are taking a different approach: intentionally slowing down. A growing number of engineering leaders are advocating for what they call “selective ignorance” — deliberately choosing not to track every new release and instead focusing on stable, well-understood tools. The logic is straightforward. Not every new model matters for every use case. Not every framework needs to be adopted. Sometimes the best response to a firehose of innovation is to step to the side.
Willison himself has gestured toward this kind of discipline, even as he struggles to practice it. He’s noted that much of what gets announced with fanfare turns out to be incremental, and that the truly consequential developments are rarer than the hype cycle suggests. But identifying signal from noise in real time is itself an exhausting task. You have to read everything to know what you can safely ignore.
There’s an irony here that’s hard to miss. AI tools are supposed to make knowledge workers more productive. And in many specific applications, they do. Willison has written extensively about how he uses large language models to accelerate his own coding, writing, and research. But the meta-problem — the work of understanding AI itself — has become a productivity drain that may offset many of those gains for the people closest to the technology.
The psychological toll is real. Engineers across the industry report anxiety about falling behind, imposter syndrome amplified by the constant arrival of new capabilities they haven’t yet mastered, and a gnawing sense that their hard-won expertise has a shorter shelf life than ever. Online forums and developer communities are filled with threads from practitioners expressing variations of the same sentiment: I can’t keep up, and I don’t know if I’m supposed to.
So where does this go?
One possibility is that the pace of meaningful capability improvements will eventually slow as models approach certain practical limits. The jump from GPT-3 to GPT-4 was enormous. The jumps since then have been real but arguably less dramatic in terms of what they enable for most applications. If this pattern holds, the industry may be approaching a period where the tooling and application layers can stabilize somewhat, giving engineers time to consolidate their knowledge.
Another possibility is that AI itself solves the problem. As coding assistants and AI-powered development tools improve, they may absorb some of the cognitive load of staying current — automatically suggesting updated approaches, flagging deprecated patterns, and translating new model capabilities into concrete implementation guidance. Willison has built tools that move in this direction. Whether they’ll be sufficient to offset the treadmill effect remains to be seen.
A third possibility, and perhaps the most likely in the near term, is that the industry simply accepts a higher baseline level of churn and adjusts expectations accordingly. Software engineering has always involved continuous learning. What’s different now is the speed and the stakes. But humans are adaptable, and professional norms evolve. The engineers who thrive in this environment may be those who develop a tolerance for incompleteness — who accept that they’ll never know everything and focus instead on knowing enough.
Willison, for his part, isn’t giving up. He continues to publish prolifically, test new tools, and share his findings with a large and grateful audience. But his honesty about the cost has opened a conversation the industry needed to have. The people building the future of AI are tired. Not because the work isn’t exciting — it is. Not because they lack talent or motivation — they have both in abundance. They’re tired because the field they chose has turned into a race with no finish line, and even the fastest runners eventually need to breathe.
The question for companies, for the open-source community, and for the AI labs driving this pace is whether they’re willing to acknowledge that human bandwidth is a finite resource — and to design their release cadences, their documentation, and their developer relations strategies accordingly. The technology won’t slow down because people are tired. But the organizations that recognize exhaustion as a systemic risk, not just a personal failing, will be the ones that retain their best talent and build the most durable products.
Simon Willison didn’t set out to become a symbol of AI fatigue. He just told the truth. And the truth, it turns out, is something a lot of engineers were desperate to hear.
The Treadmill That Never Stops: Why AI’s Fastest Engineers Are Running on Empty first appeared on Web and IT News.