The few that wake
A giant AI — most of it asleep — and the handful that light up for a single word.
A modern AI can be enormous. Yet to answer one word, it wakes only a tiny corner of itself and leaves the rest dark. This page slows that down so you can watch it happen.
The problem it solves: making an AI “smarter” usually means making it bigger — and running a bigger AI for every word costs more time and more electricity. So the largest ones cheat, cleverly: for each word, nearly the whole system stays switched off.
Narrated by a synthetic voice — not a recording · ~0.7 MB
The word is “pitch” — a musical pitch, a sports pitch, a startup pitch. Drag it from music to startup.Drag it from music to startup with your finger. Same word, three meanings — two lights follow it, the rest stay dark.
The word stayed the same. What it meant changed the pair that lit.
The fix is almost a sleight of hand. Instead of one enormous brain that reads all of itself for every word, the AI is built as a big team of small specialists. For any given word, nearly all of them stay asleep — and a sleeping specialist costs nothing to run.
Only a few wake up, do their bit, and go quiet again. That is how an AI can be huge on paper yet stay quick and cheap, word by word.
Engineers call this design a “mixture of experts.” That name is the only piece of jargon you’ll need — everything else here, you can simply watch. The grid you’re about to meet answers to a word you can drag.
You’re reading the pocket version — everything works under a thumb. On a desktop the whole page also answers your pointer. Worth a second visit.
How much of this AI is actually awake?
Here are sixteen specialists — one AI, split into a team. It’s tempting to picture the whole thing lighting up and thinking hard about every word. It doesn’t.
A sleeping specialist costs nothing to run. It still takes up room — the AI is large even at rest — but it does no work until it’s woken. Huge when still, cheap in motion.
Notice how most tiles stay charcoal-quiet.
Next: who decides which ones wake ↓
Who decides which specialists wake?
Before a word reaches the team, a tiny sorter looks at it and scores how well each specialist fits — going by what the word seems to be about.
It doesn’t wake them all. It wakes just the best two. The stronger one (tangerine) leads; the runner-up (cyan) assists. Everyone else stays dark. (Here it’s two; some AIs wake one, some a few more.)
But what does the sorter actually look at? Before any of this, every word is turned into a short list of numbers — a kind of fingerprint of how that word tends to be used. Words used in similar ways get similar numbers. That list is what the sorter compares.
So the sorter isn’t really reading meaning. It’s doing fast arithmetic on those numbers and keeping the two highest scores — that’s the whole trick you’re watching.
Try it — drag the word below the grid from music to startup. Watch which two tiles light, and their scores.Just two specialists woke. The other fourteen stayed asleep — that’s the saving.
Next: what happens when everyone needs the same specialist ↓
What if everyone needs the same specialist?
Each specialist can only handle so many words at once. If a whole rush of text is about baseball, the sports specialist gets swamped.
When one fills up, the AI has a choice: drop the extra words, or send them to the next-best specialist instead. Watch the bar under the lit tile fill as a rush of same-topic words piles in.
Holding the word still stands in for that rush — hundreds of similar words arriving together, all wanting the same specialist.
Try it — hold the word still in one meaning to make the pretend rush force a spill.Do these specialists really exist?
We gave them tidy names — “The Motion Expert,” “The Sound Sleuth.” It’s the natural way to picture a team. But look inside a real AI and you won’t find neat job titles at all.
The real specialists are tangled together. They pick up strange, overlapping habits that match no human label. The sorting is pure arithmetic, not meaning — the names are ours, invented to make it make sense.
Watch the labels get struck through.
One line left ↓
The work bent to the next-best specialist — the word still got handled. Another AI might simply drop it instead.
This demo chooses to reroute. Real systems may reroute, drop, or share out the load differently.
Next: a confession about the names ↓