Post

Running a post-hoc multi-hour design session for my team today, in response to my 3-hour review of an MR (concluding it is obscenely heavy for what it does), opened by someone who hasn't had the opportunity to learn the language or codebase, because we rely on LLMs to ship features faster... hmmm

“Most of us have taken prod down at one point or another.” hey now, there’s no reason to get personal like this

simonwillison.net/2026/Jun/11/... This is a great example of what I'm talking about. AI enthusiasts would likely get lost in everything the agent is doing; "look how productive it is! It's inventing whole debugging flows!" But step back, and you see it burned $12 in tokens on a 2-line CSS fix.

Great piece something else to consider is the distinction between declarative & procedural knowledge & how it is the degradation of procedural knowledge that dampens a junior’s & senior’s coding skillset. Expropriation risks from cognitive offloading is another significant issue.

Precisely why I didn't listen to boosters who from last Dec claimed "this is a game-changer" and yet aren't held responsible on a daily basis for their code by other engs. Your hobby startup or optimization of the smallest repo at your employer doesn't count. Show me "sprint" 🤢 work that shipped.

Aw thank you so much, Jim! Likewise, I can't express how much that means to me coming from you.