The Loaded Die.
Why a probabilistic language model can be both mathematically random and shockingly accurate, and what happens on the one roll in a hundred that lands light.
① Standard die · uniform
Every face has equal weight. The output is grammatically empty.
② Loaded die · weighted by training
100,000 faces. The weights shift with every preceding word.
The die for “The capital of France is ___”
The die is just obeying gravity. A weight of 94% on one face means roughly 94 of every 100 rolls land there, indistinguishable, on average, from "knowing the answer."
The weights come from trillions of examples.
The factory that built this die read nearly every sentence humans have written down. Its imbalance is the entire product.
The temperature knob.
A single number that reshapes the distribution before the roll. Down: the heaviest weight always wins. Up: lighter weights, and eventually the long tail, get a turn.
The dice roll is removed. The system simply picks the heaviest weight every time. Safe, repeatable, a touch robotic.
The die rolls as cast by training. Heavy weights usually win, but the 1.5%, the 0.4%, the 0.0001% all eventually land.
Lighter weights climb. The tail begins to surface. Output becomes "creative," and fabrications become inevitable.
A hallucination is a die roll that landed light.
Roll the die 100 times. About 94 dots land on Paris. A few land on London, on beautiful, on something from the long tail. To a human reader those look like errors, even though every dot is the system working perfectly.
But push harder. The reframe gets worse.
If it's always hallucinating, why use it at all?
Hallucination is the engine. Many tasks want a probability engine: plausible structure, fluent shape, a draft to react to. The trick is matching the task to the tool.
Tasks that need plausible structure. The hallucination engine is exactly the right tool for the job.
- Translate this email
- Refactor this code
- Brainstorm ten titles
- Reformat as a CSV
- Soften this paragraph
Shape will be convincing. Every factual claim inside must be checked against ground truth.
- Summarize a paper
- Draft with citations
- Explain a historical event
- Technical documentation
- Synthesize sources
Here, the job calls for an answer rather than a sentence. Use a database, search engine, calculator.
- Current weather
- Numerical calculations
- Phone numbers / look-ups
- Stock prices
- Verify a citation exists
Never ask an AI to explain itself.
It cannot introspect. It can only predict what introspection sounds like. Treat any self-explanation as confabulation.
Verification as the default.
The machine is your drafter. Every output is a draft that must be checked against ground truth before you stake anything on it.
A highly effective shape machine. Still, in the end, just a die.