Technology

I was just planning to do some sort of write up on this topic, although it will be internal only.

Of the three projects I’m currently on

• existing code base where AI sometimes has good ideas but almost never able to implement them successfully. This is legacy code, all human generated, and is probably too tightly coupled. Test framework is tightly coupled to the environment so ai cannot run it
• new tool implementation to give cheaper and faster context across all repos (Spotify Backstage)
• new code base almost entirely ai generated. Much more loosely coupled. There is no test /mock framework available, so it’s all scripts, which the ai is able to run at will to refine its guesses

There’s definitely distinct conditions where ai can be the right tool and can succeed vs when it can’t. In managements blind rush to vibe code everything, they need to better understand where it works and where it doesn’t

In particular, functionality I’m working on this week

• existing code base ”modify function x to cover scenario y” at best gives a useful strategy
• new code base “implement function x similar to existing code base, but that also covers scenario y” seems to work

It can be like that, and other times it can spit out perfectly good work that similar jobs have taken me a week in the past in under an hour. Depends on the task, depends a bit on how you ask, and depends on the model. Claude Sonnet/Opus 4.5 and higher have gotten pretty good about actually saving time and producing useful results.

Except when trying to draw 3D cats in Blender, I can say with authority now: Opus 4.8 on High/Max thinking still can't pin the tail on a cat correctly without a half dozen very specific annotated views showing it where it goes. It claims it's having trouble because the cat is lying down... our cat lays down 22 hours a day, suck it up and learn what they look like already.