AI Models Ace Complex SRE Tasks But Fail At Basic CSS In 2026

also wild, i needed something that grafana didn't do and i genuinely thought, with no sense of irony "i should probably just build my own, it will look and work better anyway' . software in 2026 is wild man

@_xjdr reflect on the quality of the engineers writing the code in the training set and be enlightened, imo

@_xjdr there are many ways to center a div but only one way to properly log in distributed systems

@meowbooksj for sure. also probably exactly 0 high quality training examples of 'center this div optimally' vs probably a mountain of reasonably high quality telemetry examples

@_xjdr genuinely most data pipelines get fixed with a kafka stream loading to a bucket with icerberg format I am not kidding, the issue is that it's an absolute pain to manage kafka

@_xjdr the frontier is jagged

@_xjdr if you use AWS MKS is decent enough but it still sucks balls

@_xjdr and then plug streamlit on top and you're good to go!

@_xjdr somewhere along the way the joke ended and this became very real

@_xjdr don't know about center a div but agents are bad at writing good frontend code especially react

@_xjdr i'm having to do some ad hoc gpu cluster administration right now and codex is like having a few very extremely literal SREs (but I repeat myself) working for me, it's great.

@xlr8harder @_xjdr everything is just a skill issue

@_xjdr tbf how often do you realistically want to center a div vs deploying some custom otel stack on k8s?

@_xjdr 现在模型很像会做难题但偶尔卡在常识题上。越是结构化、约束清楚的任务越强，越是前端细节和人类审美，越容易离谱。

@_xjdr (with spark under the hood)

@_xjdr Btw Claude can center the div properly. This is purely a codex issue

But Codex is my SRE. Every time I get an alarm codex diagnoses the issue faster than I could

@_xjdr software has become quite malleable to our needs imo