Simon Willison wrote recently about something most developers recognise immediately: a project he had wanted to build for eight years, finally shipped in three months with the help of AI tools. The post is personal and reflective, but the phenomenon it describes is structural, and worth examining more carefully than the productivity framing it usually gets.
Every developer maintains some version of the backlog. Not the professional one, tracked in Jira or Linear, but the private one. The ideas notebook, the README that never grew a codebase, the GitHub repo initialised in 2019 with a single commit that just says “initial.” These projects fail to launch not because the ideas are bad, but because the execution cost exceeds the available time and motivation at every moment they come up for consideration. The idea stays alive because you still care. The project stays unbuilt because you can’t justify the ramp-up.
Willison has been unusually candid about how AI coding tools have changed his output. He created the llm CLI tool, a command-line interface for interacting with language models locally and via API, and has built an ecosystem of plugins around it. He maintains Datasette, a mature Python tool for exploring and publishing SQLite databases that he has been developing since 2017. He publishes TIL notes almost daily. He is not someone short on output. So when someone with his track record describes a multi-year backlog item finally getting built, the interesting question is not “how did AI help” but “what specifically was the blocker.”
The Threshold Problem
Personal projects fail at a specific moment: the gap between having an idea and writing the first line of non-trivial code. That gap contains a lot. You need to research the right libraries. You need to decide on the architecture. You need to write boilerplate that has nothing to do with the interesting part. You need to hold the entire problem in your head while doing all of this, which is cognitively expensive. Most ideas expire at this threshold.
AI coding assistants change the cost structure of this gap substantially. Tools like Claude, Copilot, and Cursor do not make you a better architect; they do reduce the friction of the parts that are not architecture. Scaffolding a new project, wiring up a database schema, writing the HTTP client you already know how to write but have to type anyway, generating the test fixtures. These tasks take time proportional to their complexity, not proportional to how interesting they are. AI compresses the uninteresting parts.
For an experienced developer, this compression is disproportionately valuable. You already know what you want. You can evaluate the output immediately. You are not learning from the AI; you are offloading execution of a known plan. This is a different use case than a beginner leaning on AI to understand concepts, and it explains why the productivity gains reported by senior engineers are often larger than those reported by junior ones.
Three Months, Not Three Days
The “three months” figure in Willison’s title deserves attention. AI-assisted development is often talked about in terms of speed, and a common implication is that everything gets faster. But three months is not fast. It is faster than eight years of not starting, but it is still a significant sustained effort.
This is the more honest account of what AI changes. It does not eliminate the work of building something substantial; it lowers the activation energy enough that the work actually begins. Once the work begins, it proceeds at a pace determined by the complexity of the problem, not the typing speed of the developer. Architectural decisions, debugging subtle state bugs, understanding why the third-party API behaves unexpectedly: these do not collapse to minutes with AI assistance. They remain hard.
What AI handles well is the long tail of implementation detail that falls between the interesting decisions. In a project of meaningful size, that long tail is enormous. Reducing it meaningfully changes the total time without making the project feel “easy.”
The Vibe Coding Misread
There has been a lot of discourse about “vibe coding” in the last year or two, the practice of describing what you want to an AI and accepting whatever it produces without deeply understanding it. Willison has been consistently critical of this framing when applied to production software. His own use of AI is different: he understands every line, he reviews the output, he pushes back when the AI goes in a wrong direction.
This distinction matters for interpreting the “eight years to three months” story. He did not describe a project to an AI and receive a finished product. He used AI to reduce the cost of executing a plan he already had, in a domain he already understood. The domain knowledge is his. The architectural judgment is his. The AI handled significant portions of the implementation.
For developers reading this and wondering whether AI tools will finally unlock their own backlogs, the honest framing is: it depends on how well you understand the problem space. If you have a clear vision and the bottleneck is time and implementation effort, the change can be dramatic. If the bottleneck is uncertainty about the right approach, AI mostly accelerates the wrong decisions.
What This Means for Solo Developers
The indie developer and solo project space is where this shift is most visible. Building a useful tool as a single developer used to require ruthless scoping: you could build the interesting core, but the surrounding infrastructure (auth, billing integration, documentation, admin tooling) would either take years or not get built at all.
Willison’s broader ecosystem around Datasette illustrates both sides. Datasette itself is a well-scoped core product that has grown steadily over years. But around it he has built dozens of plugins, utilities, and integrations, many of which would be described in a corporate context as “out of scope.” The throughput on that peripheral work has clearly increased.
This is where the backlog problem gets genuinely interesting. The eight-year project presumably sat in the backlog not because the core was too hard, but because the full implementation, with all its surrounding requirements, exceeded what felt achievable. AI did not make the core easier; it made the surrounding requirements tractable.
The Compounding Effect
There is a compounding dynamic here that gets underreported. When you finally build the thing you wanted to build, you learn things you could not have learned without building it. Those learnings feed the next project. The private backlog does not just shrink; it generates new items as you develop clearer intuitions about what is possible.
Willison is a good example of this compounding. His rate of interesting project output has increased alongside his use of AI tools, and the projects build on each other in ways that require having shipped the earlier ones. This is not just productivity; it is a change in the shape of what a single developer can know and build over time.
The eight-year project was not just a project. It was a prerequisite for whatever comes next. That is the part of the story that gets lost when the conversation collapses into “AI makes developers faster.”
A Grounded Assessment
I have been using AI coding assistants heavily for Discord bot development over the last year and the shift in what gets built is real. Projects that would have lived as notes have become actual code. The quality is not uniformly good, the AI makes wrong choices regularly, and the review burden is genuine. But the shipping rate is higher, and more importantly, the activation rate is higher. More things get started.
The backlog does not vanish. New items accumulate faster than old ones clear. But the character of what sits in it has changed: fewer half-formed ideas, more specific technical challenges that require human judgment to resolve. That is a different kind of backlog, and a more productive one to maintain.
Willison’s eight years is the honest version of a story that usually gets told with shinier numbers. Three months is long. The project was hard. AI helped in specific, bounded ways. That is a more useful account than most of what gets written about this.