I don't know much about Ira Glass and I'm not going to be a 5 minute wikipedia expert about it, so maybe I'm missing out on very relevant philosophy (I hope someone links the must read thing), but those would be very intentionally inverted meanings of the taste/skill dichotomy.

LLMs are good at things with a lot of quantity in the training set, you can signal boost stuff, but its not perfect (and its non-obvious that you want rare/special/advanced stuff to be the sweet spot as a vendor, that's a small part of your TAM by construction).

This has all kinds of interesting tells, for example Claude is better at Bazel than Gemini is, which is kind of extreme given Google has infinite perfect Bazel and Anthropic has open source (really bad) Bazel, so you know Gemini hasn't gotten the google4 pipeline decontamination thing dialed in.

All else equal you expect a homogenizing effect where over time everything is like NextJS, Golang, and Docker.

There are outlier events, like how Claude got trained on nixpkgs in a serious way recently, but idk, maybe they want to get into defense or something.

Skill is very rarely the problem for computers, if you're considering it as district from taste (sometimes you call them both together just skill).