After Finishing a Week's Work in One Hour, Simon Willison Advises Developers: go weird
After compressing development time from a week to one hour, Simon Willison's new advice to all developers is: go weird. The creator of Datasette is building tools for local journalists, with the goal that someone wins a Pulitzer Prize using his software. The interview also includes his latest views on AI evaluation, prompt injection, and "why AI has made me busier."
Simon Willison has one piece of advice for anyone building software: go weird.
It doesn't mean writing bad code. It means trying ideas that seem useless. His reasoning is practical: AI has driven the cost of building things down to almost nothing, so trial and error costs almost nothing. When something goes from taking a week to build to taking an hour, weird ideas are no longer a luxury.
Willison is the creator of Datasette and has long explored the intersection of AI and software development. Turing Post recently did an interview with him, covering how AI is changing the way he works, how he evaluates new models, and what he wants to do next.
"Things that used to take a week to build now take an hour."
This isn't just about a productivity gain. It has completely changed his judgment about what is worth doing. Conventional projects are no longer appealing; he now finds himself looking forward to odd ideas he wouldn't have touched before. He has a rule: every project he starts has to be written down. At the time of the interview, he was probably five write-ups behind.
The most surprising part of this interview is what he is working on. He wants to build tools for local journalists to help them find stories in data. Local news outlets usually don't have engineering resources, so a lot of valuable data just sits there, unmined.
He said: "My goal is for someone to win a Pulitzer Prize with my software—even if the software is only 3% of their work. I want to contribute a small piece to a Pulitzer-winning story, ideally at a newspaper that can't afford to hire software developers."
For someone who builds developer tools, setting his goal as "helping local journalists win a Pulitzer Prize" is more concrete than most grand narratives about AI.
On model evaluation, his approach is also very practical: build evaluations around your actual work. Don't run a bunch of benchmarks unrelated to your work; just test the models directly on the tasks at hand. New models come out every week, and it's impossible to examine every one closely.
He also mentioned a paradox: AI has dramatically increased his efficiency, yet he has never been busier. Easy decisions get absorbed by machines, leaving him with only the hard problems.
The full interview also covers prompt injection, AI agents escaping sandboxes, and how local models can protect journalists' sources. His concerns about prompt injection are also being re-examined.
Video:
Full conversation on YouTube: https://www.youtube.com/watch?v=9C8JZTmpbE4
发布时间: 2026-11-22 07:33