Why Trying to Say the Smartest Thing Made the Machine Say Nothing at All

Give a text-generating model the opening of a story and ask it to keep going. Sometimes the result is lovely. Sometimes, thirty words in, it locks: the same sentence, then the same sentence again, then again, until you stop it. A scratched record.

The instinct is that the machine is being careless and needs to be more careful. That instinct was built into the software. At every position, the model produces odds across every word it knows: some chance of "the," a smaller chance of "a," on down a list tens of thousands long. The standard rule was to take the top one, every time. A fancier version, beam search, looked ahead and kept whichever whole phrasing scored highest. Both amount to a single instruction: at every step, be as sure as you can.

That instruction is the disease. Text made by always taking the top word repeats itself 73.66% of the time. Nearly three-quarters of it is the model saying what it has already said. Beam search, weighing whole phrasings, gets that down to 28.94%, which is still ruinous. Human writing, measured the same way, repeats 0.28% of the time. The looping isn't a flaw sitting on top of the strategy. It is the strategy.

The obvious opposite is to stop ranking and roll dice, weighted by the odds. Repetition collapses to 0.22%. The prose also wanders into nonsense, and the reason is a long tail: the model spreads a thin film of probability over tens of thousands of words it has no real opinion about. Each one is nearly impossible. Added together they are likely enough that, roll after roll, one eventually comes up. A single absurd word poisons everything after it, because the model keeps writing from what it just wrote.

Nucleus sampling threads between the two. At each position, take the likeliest options from the top until they add up to 95% of the probability. Throw the rest of the vocabulary away, then pick at random from what survives. The pool is not a fixed size. Where the model is confident, it holds one or two words and there is nothing much to gamble on. Where the sentence could go anywhere, hundreds qualify. The model's own uncertainty decides how much freedom it gets.

Judged by people on quality and variety together, it scored 0.97, against 0.67 for pure dice-rolling and 0.08 for a popular method that shortlists a fixed number of words no matter what. Its statistical fingerprint — how surprised the model is by its own output, how much it repeats, how wide its vocabulary ranges — landed nearer to human writing than anything else tested.

None of this makes the machine truthful. It was tested on a single model continuing open-ended passages, and it still says wrong things with perfect composure, confusing a whale for a bird. The 95% cutoff is a dial someone picked; it holds up across a range of settings, but no one has shown a best one. So the lesson is narrower and stranger than "randomness helps." Certainty is a reasonable thing to want at any one step and a destructive thing to demand at every step. What rescues the sentence is knowing where to loosen your grip, not how much.