Quoting Paul Dix
Simon Willison’s Weblog Simon Willison
Opinion — commentary, not a factual news event.
Paul Dix says AI wrote 1M lines of code, then kept refining them for months. He says the surprising part is how far AI gets when you give it a way to check its work.
Based on reporting by Simon Willison’s Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Paul Dix just put a very big number on the table: AI wrote 1 million lines of code, then kept refining that code for months until it became a reliable piece of software running on millions of developer machines. That is not a toy demo. That is software that survived contact with reality.
The obvious counterpoint is that the job was easier because there was an oracle to compare against. In other words, the system had a reference point, so this was not pure invention from scratch. But Dix’s point is that this still undersells what happened. The hard part wasn’t only translation from one language to another. It was getting the AI to keep improving until the thing actually worked.
That is the real claim here. If you can build a verification system and give the model proper direction, AI can produce something highly complex and highly sophisticated, then keep polishing it over time instead of wandering off into nonsense. The quote is less about raw generation and more about iteration, correction, and persistence.
And that is why it lands. Plenty of people are still arguing about whether AI can “write code.” Dix is talking about something more interesting: AI can help produce software that ships, runs, and keeps getting better, as long as the feedback loop is strong enough. That is a much less romantic story than magic. It is also the one that matters.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI coding that gets ignored because it’s less glamorous than instant genius: verification beats vibes. The industry keeps pretending the model is the product, when the boring checkers and constraints are what make the whole thing usable. Open or closed, the winning setup is the one that can prove itself without acting like a caffeinated intern.
Read more about this at: Simon Willison’s Weblog
