Draft: Intelligence Per Token
blog-intelligence-per-token.md
For the full content of this gist, refer to https://gist.github.com/jokull/2fef16038f1b25c2df90a6478aac07d4
Jökull Sólberg Auðunsson@jokull
blog-intelligence-per-token.md
For the full content of this gist, refer to https://gist.github.com/jokull/2fef16038f1b25c2df90a6478aac07d4
Hot take: TUIs are a phase, and GUIs that basically emulate TUIs are a phase too. We just don't know how the IDE of the future will look like.
He's talking about the arc of dev tools. Terminal, then Notepad++, then richer IDEs, then VSCode. Each phase building on the last, borrowing metaphors, adding chrome. What comes next?
But something about the question stuck with me.
Unix evolved under constraints. Teletypes. Tiny memory. Slow connections. Those limits forced everything through text — and text turned out to be strangely universal. Decades later, everything still speaks it.
Over time, Unix grew its own dialect. Not quite a programming language. More like a pidgin for dynamic tasks. Ad-hoc pipelines. One-liners you'll never run twice. A language for the unrepeatable.
Most syntax wants to be repeated. Code you'll run a thousand times. Workflows you'll click through daily. The shell is different. It's for the one-off. The exploratory. "Let me just see what's in here."
And that made it weirdly ergonomic. Terse commands that feel cryptic at first, but compress intent in a way that becomes natural. ls -la | grep foo | wc -l isn't elegant, but it says exactly what you mean with minimal ceremony.
Now stretch this a bit.
Context windows are the new RAM. Every token matters. An LLM parsing your intent has limited attention — and anything spent on UI description is attention not spent on the problem.
Pipes and flags pack more meaning per byte than any dialog box. An AI can chain jq | grep | awk more reliably than clicking through nested menus. It reads your intent, not your mouse movements.
The shell was always a kind of prompt. A way of expressing intent to a system that could execute it. We just needed something on the other end that could reason about it — not just parse it.
So what's the IDE of the future? Not a new interface. The coding agent prompt.
Same constraints that shaped Unix. Same terseness. Same pidgin for the unrepeatable. Just a different listener — one that doesn't just execute, but thinks.
All about intelligence per token.