What’s your gold standard for deciding whether AI-written code is ready for production? In Episode 42 of Talking Postgres, Simon Willison—creator of Datasette, co-creator of Django, and prolific open-source developer—joins Claire to share how AI is changing the way he builds software today. We dig into why his test for shipping AI-generated code is “Could I explain this to somebody else?”, how engineering management skills can be surprisingly useful when managing AI agents, and how the bottleneck is no longer writing code but understanding it. Also: personal credibility, deep research, the Winchester Mystery House, slop proxies, and Simon’s observation that “Features are cheap. That doesn’t mean you should build them all.”
Previously on Talking Postgres:
Links mentioned in this episode:
In this episode, we covered:
00:00 Intro & music
04:00 A week of work before breakfast
09:12 Red-green TDD makes agents exercise every line
14:42 A million lines mean nothing if you don’t understand it
18:28 Finding low-hanging fruit among open PRs & issues
22:07 Getting work done while walking the dog
23:01 Gold standard: “Could I explain this to somebody else?”
28:27 Slop proxies add no value at all
30:49 Aggressive nitpicking reviews
35:01 Everything in software engineering is about trade-offs
41:16 Racoon heist game is only fun for 1 minute 15 seconds
43:46 New Year’s resolution to be more ambitious
48:11 You have to learn to throw things away
51:57 Features are cheap. That doesn’t mean you should build them all
1:00:45 Engineering management skills are so useful
1:04:44 QA specialists should be having a great time
1:07:53 Writing is thinking, don’t outsource it
1:08:23 Skill atrophy is a choice you make
1:10:07 I believe in people who are motivated
1:11:55 Systems are not set up to deal with this volume
1:18:48 Research agents stopped being absolute garbage
1:21:49 The whole point of the “human in the loop”
1:26:10 Dream: I want there to be more small businesses