The longer you speak, the heavier your words become.
The attention tax
One conversation. Two costs to follow.
"Attention" just means the AI looking back over every earlier word each time it adds a new one — and that looking-back is the tax.
Watch one AI chat get slower and more costly with every word it adds — and see, step by step, exactly why.
A guided walk through the notes and the cost · synthetic voice, not a recording · ~0.5 MB
You have probably felt it. A long AI chat starts quick, then drags — each reply slower than the last, and if you pay by length, pricier too. The odd part: the AI is not thinking any harder.
The catch: with every new word, the AI re-reads a growing pile of notes — one note for every word already said. The longer the chat, the taller the pile it has to re-check each time. This page shows you that pile growing, and what it costs.
Follow two costs as they climb: the space the AI's notes take up, and the effort it spends choosing each new word. You'll stretch one conversation and watch both respond. Scroll the page, and drag the sliders with your cursor.Scroll the page, and drag the sliders with your thumb.
You're reading the pocket version — everything works under a thumb, but on a desk the whole page reacts to your cursor. Worth a second visit.
What does a growing chat actually cost?
Every time the AI replies, it is handed the whole conversation so far — from the very top — as one long block of text (everything it can see right now is its context window). Before it can add a single word, it has to work through all of it.
What is the AI actually keeping?
Reading that whole conversation from scratch each time would be slow. So the AI does the reading once, and keeps a short note for every word — a small bundle of numbers. That growing pile of notes is its scratchpad for this one chat (engineers call it the KV cache).
It isn't remembering or understanding. Each note is just numbers, jotted down once when its word arrived and never touched again — a pile that simply takes up space. On the right, every word kept adds one line to the stack, and the meter labelled "Space the notes take up" rises exactly with it.
The more of the conversation it keeps, the taller the note stack — and the more space it takes up. Twice the words, roughly twice the storage space.
Now: what it costs to consult that receipt ↓What does choosing one word actually take?
Here's where the cost hides. Hit the button, and watch the glowing lines: that sweep is the AI choosing a single new word.
To pick it, the AI glances back at every note in the stack — weighing what it is about to say against each earlier word in turn (this glancing-back is what engineers call attention). It's easy to picture the AI reading the chat once and then coasting. It doesn't: it runs this whole check again from scratch for every new word.
Can throwing away old notes fix it?
Here's the repair everyone reaches for: drop the oldest words to keep the pile of notes small. Drag the notes away and watch what it costs you.
So here's the habit worth keeping. When an AI chat feels slow, or you're paying by length, picture the note stack: everything still in the conversation gets re-checked for every new word. Keep what matters, trim what doesn't — and when the pile of notes outweighs the point of the chat, start a fresh one. Just know that whatever you cut, the AI genuinely can no longer see it.
Narration transcript (synthetic voice)
0 · Scene setting (plays when you turn sound on)
You're in the lab. One AI chat, and two costs that climb with every word it adds. The controls are yours to drag.
1 · The hook
A long chat doesn't drag because the AI is thinking harder. It drags because a pile of notes keeps growing — and it re-checks all of them for every new word.
2 · Grow the chat
Drag the first slider to the right. Stretch the conversation out. Keep your eyes on the two meters.
3 · Grow payoff
Both meters climbed together. Storing the conversation is one cost. Choosing each new word is the other.
4 · The note stack
That stack on the right is its notes — one line for every word so far. Just numbers, written once. The taller the stack, the more space it takes.
5 · Watch one word
Hit the button and watch. Those lines are the AI choosing a single new word — glancing back at every note in the stack.
6 · The sweep payoff
And it does that whole sweep again for every word it writes. A longer chat means more notes to check.
7 · The obvious fix
Drag this slider left to throw away the oldest notes. Watch both meters — and watch the start of the conversation.
8 · The cut payoff
Both costs dropped. But the opening vanished with them. The AI can't use what it no longer holds.
9 · Coda
Everything still in a chat gets re-checked for every new word. Trim what you don't need, or start fresh — just know that whatever you cut, the AI truly can't see anymore.