Watch an AI answer a simple riddle. First it blurts out a fast, confident, wrong number. Then — as you give it room to work the steps out — the answer changes to the right one.
Sarah is twice as old as her brother.
In 4 years she will be 20.
How old is her brother now?
A teaching demo — the steps, answers and numbers are hand-written to show the idea, not measured from a real AI.
Slide the dot to the right to give the AI room to think — or use the arrow keysDrag the dot to the right to give the AI room to think — or use the arrow keys
The chain that thinks out loud
Watch a confident wrong answer become a worked-out right one — as you give it room.
Narrated by a synthetic voice — not a recording · ~0.4 MB
The first time I watched an AI “think out loud,” I assumed I was reading its mind. I wasn’t. I was watching it do something much more ordinary: it was working the problem out in rough, one line at a time, before it settled on an answer.
Ask an AI a question that takes two steps and give it no room to work, and it does what a rushed person does — it grabs the loudest-looking number and guesses. Give it room to jot the steps down, each one leaning on the last, and it can reach an answer it could never have blurted out in one go.
So here is one small riddle, held still. You decide how much room the AI gets. Drag it up from nothing and watch the working-out fill in — and watch the answer change.
What is made up — Every number on this page is a teaching prop: the amount of room, the lines of working-out, the “how often it’s right” curve and its percentages are all hand-written to show the shape, not measured from a real AI. What the demo honestly shows: questions that take a few steps can fail when an AI gets no room to work them out, and giving it that room can flip a wrong answer to a right one. And the working-out it shows you is room being spent — not a guaranteed record of what really went on inside. We’ll come back to that last part.
You’re on the pocket version — the slider works under your thumb, but on a desktop the whole page answers your cursor. Worth a second visit.
What can it do with no room to work?
Room set to nothing. The AI answers straight away — ten — and it sounds sure. It grabbed the biggest number in the riddle, twenty, and halved it. Twice as old, half the age. Neat. Wrong.
The right answer needs two steps this snap guess never took.
What do the first lines of working-out buy?
Drag the room up. Lines of working-out appear before the answer: first it works out Sarah’s age now, 20 − 4 = 16. Stop it here and it answers 16 — still wrong, but wrong in a more honest way. It simply ran out of room mid-thought.
Each line the AI writes is fed straight back to it as it writes the next one — the working-out is scratch paper that lets it carry a result forward instead of holding everything in its head at once. (Each little piece it writes is what engineers call a token; the whole run of working-out, a chain of thought.)
Try it — drag the slider and watch the working-out appear before the answerFirst step done · More room showed useful working-out, but a half-finished attempt still gave a wrong answer. Now test the second step.
Can one more step change the answer?
The working-out has stalled one step short of done. Everything it needs is already on the page: half of sixteen. The room is yours to give now.
Try it — give it the last step and watch the answerIt flips: eight. Same riddle, same AI — the only thing that changed is how much room it had to work. The thinking here isn’t a flash of insight. It’s a quantity you can hand over.
That’s the flip — now zoom out ↓Does this hold beyond one riddle?
One flipped riddle proves nothing on its own. So try a whole pile of them.
Run a hundred toy riddles like it and a shape appears. With almost no room, it gets few right — barely better than guessing. Give it more room, and more of the answers come out right — the line climbs.
Try it — drag the slider and watch the dot ride the lineDoes more thinking always pay off?
Push the room higher and, in this toy pile, the climb flattens out. Once a riddle’s two steps are covered, extra room here mostly buys repetition, not new right answers. This is a two-step riddle, so two steps is most of what it needs.
Notice — past the second step, the dot stops climbing.Reality check — Real results aren’t this tidy: depending on the task and the AI, extra thinking can keep helping, level off, or even make things worse. This line is just one made-up shape, chosen to show the flattening.
Is the working-out what really happened?
First — how could that even happen? The AI writes those steps the same way it writes anything else: one word at a time, each one picked to sound like believable working-out. It isn’t reading them off a recording of what it did inside; it is composing them. So the same knack that lets it write a smooth sentence lets it write tidy steps that look like the route to the answer without being it — the way a student who already senses the answer can write neat lines afterwards to justify it.
So here is the honest part. The tidy steps you just watched are working-out the AI spent room on — they are not guaranteed to be how it actually reached the answer. An AI can land the right answer by a route that looks wrong, or write neat working-out that isn’t the real path it took.
Read the working-out as work you paid for. Useful, often honest — but never proof of what happened inside.
Try it — press the button, then compare the two ways of working it outThe answer stayed right while the route went wrong. Treat the working-out as something you can read, not proof of the hidden path that produced it.
Sarah is twice as old as her brother. In 4 years she will be 20. How old is her brother now?
Same “8” — but suppose the AI got there like this:
and 10 − 2 = 8. → answer: 8 ✓
Right answer, two wrong steps — a fluke that happens to land. The working-out looked like real thinking. It wasn’t.
Sarah is twice as old as her brother. In 4 years she will be 20. How old is her brother now?
Same “8” — but suppose the AI got there like this:
Sarah is 20, so 20 ÷ 2 = 10, and 10 − 2 = 8. → answer: 8 ✓.
Right answer, two wrong steps — a fluke that happens to land. The working-out
looked like real thinking. It wasn’t.