It contains a helpful insight that there are multiple modes in which to approach LLMs, and that helps explain the massive disparity of outcomes using them.
Off topic: This article is dated "Feb 2nd" but the footer says "2025". I assume that's a legacy generated footer and it's meant to be 2026?
For example, there was a case of how Claude Code uses React to figure out what to render in the terminal and that in itself causes latency and its devs lament how they have "only" 16.7 ms to achieve 60 FPS. On a terminal. That can do way more than that since its inception. Primeagen shows an example [0] of how even the most terminal change filled applications run much faster such that there is no need to diff anything, just display the new change!
The people that could make terminal stuff super fast at low level are retired on an island, dead, or don't have the other specialties required by companies like this, and users don't care as much about 16.7ms on a terminal when the thing is building their app 10x faster so the trade off is obvious.
Not that everything we want an agent to do is easy to express as a program, but we do know what computers are classically good at. If you had to bet on a correct outcome, would you rather an AI model sort 5000 numbers "in its head" or write a program to do the sort and execute that program?
I'd think this is obvious, but I see people professionally inserting AI models in very weird places these days, just to say they are a GenAI adopter.
Maybe one day that will change