The lab — instruments, not essays.
Some mechanisms are better operated than described. Each instrument here teaches one piece of how AI actually works — scroll it, hover it, drag it — and links back to the essays that go deeper. The numbers inside are toy values; the shapes are the point.
The journey of a prompt
Press Enter and something happens before the first word streams back — an invisible document assembles, your sentence becomes tokens, joins a queue, and two clocks start ticking. Walk the whole hidden passage, step by step.
The answer that calls for help
Ask an AI to add. It can answer from memory — sounding certain, getting it wrong — or hand the sum to a calculator and read back something real. Watch it decide which, and see which digits you can trust.
The paths a sentence takes
At every step a sentence could turn a dozen ways. Two small dials tilt which word the model reaches for — drag them and watch the branches open wide or narrow to a single line.
The attention tax
The longer the conversation runs, the heavier each new word becomes. Watch two costs climb together and feel why a model's memory is never free.
Certain, fluent, wrong
One answer that sounds certain, reads beautifully, and is false. Take it apart word by word and find the exact place where fluency parted ways with the truth.
The illusion of memory
A long conversation quietly drops an early fact — and then, ask as you might, the model itself has nothing left to recall. Watch a memory slip past the edge of what it can still see.
The geometry of meaning
A machine files every word as a place on a map, where 'close' just means 'used alike'. Turn the view and watch which words count as nearest quietly rearrange themselves.
Where the sentence stands
Dog bites man, man bites dog — same words, opposite worlds. A hidden tag tells the model which came first; peel it off and watch the sentence lose its footing.
What survives the squeeze
An AI is millions of stored numbers. Give each one fewer settings — the squeeze that fits it on a phone — and watch most answers hold steady while one quietly flips.
The few that wake
A vast model sits mostly asleep. For every single word only a handful of specialists stir — drag a word along and watch which few wake to handle it.
The word before the word
Slide a cover along a hidden sentence and watch an AI guess each word before it appears — its surprise spiking exactly where the sentence turns strange. This is how it learned to write.
The style switch
Ask the same plain question twice — once of a raw model fresh off the internet, once after it was coached to be helpful — and watch the same machine answer like two different minds.
The fast learner
Tuck two examples into your question and the model's next answer snaps into line — a new trick picked up on the spot, while nothing inside it actually changes.
The chain that thinks out loud
Give an AI a riddle and it blurts a fast, confident, wrong number. Give it room to work the steps aloud and the answer rights itself — until you watch extra thinking stop paying its way.
The cut that changes the answer
Drag a single cut through one safety sentence and watch a search tool pull back the wrong half — handing the AI the opposite of what the rule actually says.
The word it cannot spell
Type a word and watch an AI that writes flawless essays fumble a child's task — counting its letters — because it never sees the letters at all.
Instruments are built in the open and reviewed before they ship. If one taught you something — or misled you — tell me.