How does an AI pick up a new trick
without learning a thing?
The fast learner
Show an AI a couple of examples inside your question, and watch its next guess fall into line — without a single thing changing inside it.
Narrated by a synthetic voice — not a recording · ~0.7 MB
An AI like this is built once — from a huge pile of text — and then it's finished. Nothing inside it changes after that day. It can't pick up a new fact, a new rule, or a new habit by rewriting itself. Think of it as set in glass the moment it's done.
And yet — show it a couple of examples of what you want, and it catches on at once. That's the puzzle this page pulls apart: how can something set in glass still pick up a brand-new trick the instant you show it one? (Researchers have a name for this — learning from the examples inside your question — but the plain fact is what matters here.)
Scroll gently — the panel stays pinned beside you the whole way down.Scroll gently — the panel stays pinned above you, and you tap the examples into it as you go.
One honesty note. Every percentage on this page is a made-up teaching number, hand-picked to show the shape of what happens — not a real measurement (exactly how this works inside is still being researched). What's real, and what you can try with any AI: it's finished and unchangeable, it reads your whole question at once, and its next guess sharpens when your examples agree and splits when they disagree. The one thing to take away: examples in your question change what it says without changing the AI itself.
You're reading the pocket version — everything works under a thumb, but on a desk the background of dots answers your cursor too. Worth a second visit.
What does the AI do when your question gives it no clue?
Look at the panel. We've typed the start of a translation — the word cat, an arrow, and a blank — and asked the AI to fill in the blank. That's every scrap it's been given.
With nothing to go on, it can't tell what the arrow is even asking for. A sum? A word game? A riddle? So its guesses come out as a fog — a scatter of common words and noise, none of them sure. The bars show how strongly it leans toward each word (made-up teaching numbers).
Next: give it one example ↓Can a single example move a mind that can't change?
The AI reads your whole question at once — every example you've added, plus the blank — before it makes a single guess. So an example you drop in simply becomes more of what it's reading.
The first example is sea → mer — mer is the French word for sea. That single French clue is the only hint it now has about what you want.
Try it — drag 'sea → mer' into the question and watch the bars.Try it — tap 'sea → mer' to drop it in, and watch the bars. Next: a second example ↓What do two examples that agree do?
One example raised a suspicion. What does a second one, pointing the same way, do?
The second example is dog → chien — chien is also French. So now both clues point the same way: two French examples, agreeing. When your examples agree, they pull the blank hard in one direction.
Here's the part most people get wrong. You didn't teach it anything — nothing inside the AI changed. Your examples are just more words for it to read, and they steer it toward a pattern it already picked up long ago, from the huge pile of text it was built on.
Try it — add 'dog → chien' and watch the guess snap into line.Try it — tap 'dog → chien' to add it, and watch the guess snap into line. Next: break it ↓If it had learned the rule, could one odd example break it?
Nothing was written into the AI — so whatever just sharpened has nowhere to live. Watch what one example that disagrees does to it.
The third example is bird → pájaro — pájaro is Spanish, not French. Now the clues disagree: two French, one Spanish. A mismatched example doesn't settle the blank — it pulls it two ways at once, and the guess splits.
Try it — drag in 'bird → pájaro', then drag it back out.Try it — tap 'bird → pájaro' to add it, then tap it again to take it out. Last: what this means for the questions you type ↓The AI never learns your trick.
It just follows the examples in front of it.
So here's the habit worth keeping. When you paste examples into a chatbot to show it what you want, you're not teaching it anything — you're steering it. Choose those examples as carefully as you choose the question itself: keep them consistent and its answer sharpens; slip in one that disagrees and its answer splits and turns vague. When an answer comes back muddled, check your own examples first — a stray one may be pulling it two ways at once.
Narration transcript (synthetic voice)
0 · Scene setting (plays when you turn sound on)
You're in the lab. This AI was built once, from a mountain of text — and then it was finished. Nothing inside it can change now. You're about to hand it a brand-new trick anyway, and watch it catch on.
1 · The void
Look at the panel. Just the word cat, an arrow, and a blank — that's all it's been given. The guesses underneath are a fog of common words, none of them sure.
2 · One example
Drop in the first pair — sea, mer — and watch the bars. One example was enough to raise a hint of French, but it hasn't settled on anything yet.
3 · The snap
Add a second pair that agrees — dog, chien. There's the snap. And notice what did the work: the examples you added, nothing else. Nothing inside the AI changed.
4 · The contradiction
Drop in the Spanish pair and the focus splits apart. Take it back out, and the focus snaps straight back. Nothing was ever stored; it was only ever reading.
5 · What to keep
Examples in your question don't teach the AI — they steer it. Keep them consistent, and the answer sharpens; slip in one that disagrees, and it turns vague.
RK · a quiet place for long reading