Post B655qNX6H1P8oL1vN2 by thomask@social.octet-stream.net
 (DIR) More posts by thomask@social.octet-stream.net
 (DIR) Post #B655qNX6H1P8oL1vN2 by thomask@social.octet-stream.net
       0 likes, 0 repeats
       
       Blog: "Yes, yes, I'm one of those AI users now" https://octet-stream.net/b/scb/2026-05-08-yes-yes-im-one-of-those-ai-users-now.html
       
 (DIR) Post #B655qNhNenczKDq8OW by minoru@functional.cafe
       0 likes, 0 repeats
       
       @thomask I'm also one of those who was super-sceptical but now toys with local LLMs for work and fun. You say you're three times faster but the quality is the same -- how did you measure this? I'm thinking of replicating this study https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ to measure the speedup, but I don't have a good way to measure quality aside from my gut feeling. That feeling appears to be gullible: in https://github.com/Minoru/minoru-fediverse-crawler/pull/412, I generated some code and believed it looked okay, but self-review one week later found plenty of holes.
       
 (DIR) Post #B658RI3dnCzyS8ym6i by thomask@social.octet-stream.net
       0 likes, 0 repeats
       
       @minoru Right, it’s hard to test yourself scientifically. It just seems continuous though. A couple of hours ago I had to analyse an iOS and an android app, neither of which I wrote, to understand how UI lifecycles relate to Bluetooth lifecycles, via complex reactive middleware. It took 15 mins prompting to get the answers I needed and it would have taken hours for me to understand otherwise. Coding-wise I get roughly 150-300 decent LOC in a half-day manually so I have a rough comparison.