Home / What a language model is, without the jargon

The myth of the magic setting

Two different answers to the same question do not signal a fault: they come from an internal setting, temperature, which adjusts the risk taken in choosing the next word, and this setting does not fix a poorly phrased request, the request itself can.

Ask Claude the same question twice, in two separate conversations, and you will often get two different wordings for similar content. This is not a flaw or pure chance: it is the effect of an internal setting, called temperature, which adjusts how much risk the model takes when several fragments of text have a similar probability to one another. Understanding this mechanism stops you looking, in some checkbox, for the cause of a disappointing answer.

What temperature actually adjusts

When several next words are roughly as probable as one another, a low setting pushes the model towards the most probable choice, making it more stable and more predictable from one run to another. A high setting lets it sometimes keep a less probable choice, which makes the answer more varied and more exploratory, at the cost of lower consistency. This setting is finely controlled by people who use Claude's programming interface, to build a product or a service. In the public interface at claude.ai, what you choose directly is the model and a level of effort, from the fastest to the most thorough: a slider that bears on the depth of reasoning engaged to answer, not on this risk taking in the choice of words.

What a setting does not fix

The costliest confusion is looking, in some setting, for the solution to a disappointing answer. A vague request remains a vague request after several surface reformulations: the words change from one answer to the next, the weakness stays the same, because the information missing from the request is no more present in the setting that produced the answer. Two answers obtained from the same vague prompt can thus differ in their wording while sharing exactly the same gap in substance, a situation this lesson's figure illustrates with a concrete example drawn from a request for an event tagline.

The reflex to keep

Faced with an answer that does not fit, the useful question is not which setting to change, it is spotting what the request itself left unspecified, the intended audience, the expected tone, the desired length, or the final use of the text produced. Rephrasing the request to add the missing element changes the result more than sending the same request again in the hope of a more favourable draw. The lesson on clarity and the why takes up this reflex and gives the method for completing a request that misses its target.

Request: write a tagline for our event.
Answer: A friendly evening awaits you on 10 October.

Request: write a tagline for our event.
Answer: An event not to be missed, on 10 October.