What survives the squeeze.
An AI is built from lots of stored numbers. Drag the dial to give each number fewer possible settings — that's the squeeze that shrinks it to fit a phone. The strip below stands in for those numbers. Watch what the squeeze does to them — and, further down, to the answers.
Narrated by a synthetic voice — not a recording · ~0.5 MB
a toy stand-in — built to show the shape, not a real AI
65,536 possible settings each number can be set to
Try it — sweep your cursor across the page, or drag the dial; distinct colours collapse into shared bandsTry it — drag the dot below; distinct colours collapse into shared bands
Every square is still here. All that changed is how many different shades they're allowed to be — the squeeze coarsens, it never deletes.
The catch: a smaller AI isn't simply a worse AI. Almost everything it does can look fine — until it hits the one kind of answer that was already a close call. This page lets you find that edge yourself.
Scroll — so what are these numbers? ↓What does shrinking an AI actually change?
An AI is built out of billions of stored numbers — roughly one for every tiny connection inside it. These numbers are what it has learned; together they are what the AI "knows."
When the AI is first built, each number is written out in fine detail — about sixteen bits of room apiece, far more than it needs just to answer questions. So to make the AI smaller and cheaper to run, you write every number in less space: round each one to a shorter list of allowed values. The count of numbers never changes — only how finely each one is written down. What that costs the AI's answers is exactly what this page lets you test.
What is illustrative: the coloured strip, the six test questions and the point where one flips are hand-picked teaching values. The rounding itself, and the storage arithmetic, are real. This shows the shape of the trade, not the behaviour of any one AI.
The strip here uses the simplest possible squeeze — round every number to the nearest evenly spaced mark, nothing cleverer. Real shrinking is more careful: different parts of the AI are rounded on different scales, the rounding is tuned against real data, the most sensitive numbers are kept wide, and the AI is given a chance to adjust before it ships. All of that is left out here on purpose, so you can see the bare mechanism.
You're reading the pocket version — the strip and the dial both work under a thumb. On a desk, your cursor alone drives the hero above.
First, look at the AI before any squeeze. ↓How much room does each number need?
Look closely at the strip on the right. At the widest setting — sixteen bits of room — almost no two squares are exactly the same shade. Thousands of finely separated values, each number free to be whatever it learned to be.
An AI is usually built, and often run, at this width, because working the numbers out in the first place needs that much room. Answering questions afterwards doesn't always.
At full width, the rounding is so fine you can barely see it.
Next, do the squeezing yourself. ↓What does the strip lose as you squeeze?
Slide the dial from sixteen down toward two and watch the grain drain out of the strip. Squares that used to hold their own distinct value start landing on the exact same stored number — that's a band. Nothing is deleted: the same squares are all still there, each one just allowed fewer possible values.
The thing most people get wrong: shrinking an AI doesn't tear pieces out of it. Every number stays. It only gets a blunter set of values to round to — and, as you'll see next, that bluntness doesn't hurt every answer equally.
Try it — drag the dial left (or sweep your cursor) to lower the setting and watch the bands formTry it — drag the dial left to lower the setting and watch the bands form Move over any square to see the value it actually heldTap any square to see the value it actually heldEvery square remains — but at the low settings there are only a handful of stored values left for them all to share.
You've seen the whole squeeze. Next: the simple rounding rule behind it ↓Why does squeezing blur the numbers together?
Here's the whole rule. Line up the smallest and largest value a number might need, then lay a ruler across that gap with a fixed number of evenly spaced marks — and round every number to its nearest mark. Fewer bits of room means fewer marks on the ruler.
At sixteen bits the ruler has 65,536 marks; drop to four and it has just sixteen. The number's importance hasn't changed — only how precisely it can be written down. When two different numbers fall between the same pair of marks, they round to the same value. That's a band.
Fewer marks, coarser ruler, more numbers colliding on the same value.
Try it — drag again and watch the ruler below the dial lose its marksTry it — drag again and watch the ruler below the dial lose its marks Now test what that rounding does to real answers. ↓What breaks first — and why that one?
Six little test questions, re-answered at whatever setting you land on. Each has a margin — how far ahead its best answer was over the runner-up. Squeezing adds a little rounding error to every number. An answer only changes when that error grows bigger than its own margin — so the close calls go first.
What is illustrative: any flip you see is worked out live, not scripted — it comes from comparing the rounding error at your setting against each question's margin. It's a teaching stand-in for the idea, not a peek inside a real AI.
Try it — drag from four bits down to two and watch the listTry it — drag from four bits down to two and watch the listOne flipped — the question whose top two answers were already neck-and-neck. The other five didn't get lucky; their margins were simply wider.
A narrow margin broke first. Now see what all that squeezing actually buys. ↓What does the squeeze buy you?
The storage shrinks almost exactly in step with the setting. An AI with seven billion of these numbers packs to about 14 GB at sixteen bits — already too much for most laptops to hold comfortably, let alone a phone. Squeeze to four bits and the same numbers pack to about 3.5 GB. Now it fits.
All seven billion numbers are still there — each one just takes fewer bits to store.
This meter counts only the numbers themselves, not the extra working memory the AI needs while it's actually running.
The saving is in packed storage; it is not the AI's total memory while it runs.
Carry the trade, not the toy numbers. ↓every number rounds to the nearest of these 65,536 marks
the stored numbers only, not the extra room the AI needs while running · marker at 4 GB — roughly where an AI starts feeling phone-sized