People hear "dice roll" and picture six faces, equal odds, pure chance. That metaphor is why the magic of language models survives, because a fair die would only ever produce gibberish. To see the machinery as it actually is, you have to upgrade the die.
A six-sided die is uniform: each face has the same 16.6% chance of landing up. A 100,000-sided fair die, one face per word in the vocabulary, would simply emit a random word from the dictionary every time. The result is pure noise. Language requires the weighting.
Imagine a die with 100,000 sides — one for every token the
model can emit. The die is magically weighted, and the
weighting changes in real time based on the words that came
before. Given “The capital of France is”, the
face for Paris carries almost all the mass. The face
for banana still exists; it just weighs nearly
nothing.
What you are watching is the die obeying gravity.
"Temperature" is a single number that re-shapes the distribution before the roll. Turn it down and the model always picks the heaviest weight: safe, repeatable, a touch robotic. Turn it up and the lighter weights get a chance. Turn it up far enough and the long tail, the off-script tokens, starts to land.
If the loaded die lands on Paris 97 times out of 100, we call the system a genius. But on the 98th roll, the 1.5% face lands up and the output is London. To you, that's a lie. To the machine, the math worked exactly as designed.
We just said: some rolls land light, and those are hallucinations. That's the comfortable version of the story. The hard version goes further, and once you see it, you cannot unsee it.
The machine has no causal model of the world. No filing cabinet of facts. No internal stage where it consults reality before speaking. It only has weights and probabilities. When it says Paris, it is outputting a high-probability token sequence. No fact is being accessed. When it says banana, it is doing the exact same thing. We call one of these accuracy and the other a hallucination. The mechanical process is identical. It is always hallucinating. Sometimes the hallucination happens to align with consensus reality, and we reward the machine for it.
The AI never sees reality. It only calculates the proximity of tokens.
{/* The introspection trap — a live demo. */}When you ask the model “How do you work?” it does not run a diagnostic on its own math. It cannot — there is no internal mirror to consult. It does what it always does: predicts the most plausible next tokens. Because its training data is saturated with articles, blog posts, and transcripts about how AIs work, the output sounds exactly like an engineer explaining an AI. It is a simulation of an explanation. Try it:
This is why it doesn't make sense to ask an LLM to explain itself.
It cannot introspect. It can only predict what introspection sounds like.
Any explanation the model gives for its own behavior is a post-hoc rationalization. The bureaucrat persona steps in after the fact to write a polite, plausible memo about a process it has no access to. Treat it as confabulation.
The Empty Glove can sound like a verdict: therefore, never use the tool. That's the wrong reading. The right reading is more useful and more honest.
Hallucination is the engine.
And many tasks want a probability engine. They want plausible structure, fluent shape, a draft to react to. The trick is matching the task to the tool. Three buckets:
As a brainstorming partner, the model's tendency to invent plausible structures is its greatest feature. Ask it for ten possible angles on a problem. Three are bad, two are obvious, five are useful. You cut what you don't want and keep what you do.
You'd rather start with seven good ideas you didn't have to generate than spend twenty minutes staring at a blank page. The hallucination is scaffolding you didn't have to build.
Verification as the default.
The machine is your drafter. Every output is a draft.
Every draft must be checked against ground truth before you stake anything on it. The model produces a shape. You bring it into contact with the world. That seam, between what the die generated and what is actually so, is where the human has to stand.
The loaded die is a complete mechanical vocabulary. It lets you describe what the machine is doing, why it usually works, and why it sometimes fails, with no reference to understanding, intention, or mind. Three questions, three cleaner answers:
Where the weights come from, what they are, and how they get cast. There is no magic and no mind here, only a casino that has been rigged for so long it has started to look like an oracle. Pair this exhibit with the other Glass Box presentations on AI literacy.