You're viewing the front-end.social public feed.

Renke Meuwesemeuwese@mastodon.social
Nov 7, 2026, 10:18 PM
@bert_hubert @cwebber wow! The inclusion of the experiment on reviewing quality is very useful! I can't believe that research isn't more famous in general... but it explains something quite fundamental. I've always felt this vague feeling that code review isn't what we say it is. The experiment proves that, and makes clear that reviewing AI generated code is never going to give us what we want from it: a guarantee that we aren't introducing weird new bugs we don't understand.
💬 1🔄 1⭐ 0

Replies

Matt "msw" Wilsonmsw@mstdn.social
Nov 8, 2026, 1:11 AM
@meuwese @bert_hubert @cwebber there is quite a bit of literature on "modern code review", and if the practices are reliable for identifying defects. Generally, they are not particularly good for that.
But there are other benefits. For me, they help refresh my mental "theory of the program" and give me opportunities to exchange information and knowledge with others who are working on the same program (or similar programs).
💬 1🔄 0⭐ 0
Renke Meuwesemeuwese@mastodon.social
Nov 8, 2026, 1:24 AM
@msw @bert_hubert @cwebber for sure. But those benefits don't apply when you're reading LLM output. The machine doesn't understand the code any more than a toaster can taste the toast. If you develop a mental model of how LLM generated code is meant to work, you're just confusing yourself, because there is fundamentally no intention there.
💬 1🔄 0⭐ 0
Matt "msw" Wilsonmsw@mstdn.social
Nov 8, 2026, 2:02 AM
@meuwese @bert_hubert @cwebber there may be intent, or there may not be. It depends on the other humans involved, and what methods and strategies are used by them.
I liked this article from ACM Queue over the summer...
https://queue.acm.org/doi/10.1145/3807966
💬 0🔄 0⭐ 0