ChatGPT, Gemini, Claude, DeepSeek and others artificial intelligences (AI) seem simple: we write, we give them our request or question and we receive a response. However, behind this simplicity of use there are two parameters that greatly determine the quality of the response that we get. And they often go unnoticed.
We cannot manipulate the temperature and response length directly from the conventional chat. In fact, they are only touched from the API or using advanced tools, but the good news is that we can simulate their effect by resorting to the prompt appropriate.
Temperature affects the creativity of the response. And the length to which the AI model extends. The important thing is that by combining them we can completely transform the result of the same question.
What these two parameters actually do
The language models we currently use do not “know” the answers to our questions. They actually calculate, roughly speaking, word by word which is the most likely next word. This is all. In this context, the temperature decides to what extent the model deviates from that most probable option.
A low temperature (and therefore close to 0) almost always chooses the predictable. A high one, however, allows for less probable options, with more variety, but also with more risk of detours. Be that as it may, there is no “correct” temperature: depends on the task. To summarize a contract, explain a concept or translate a text, we want it to be low. But to search for names, write an e-mail or propose ideas, we are interested in just the opposite.
Almost no interface in AI applications allows us to touch this parameter directly, but we can simulate its effect using explicit instructions. If we write in Claude or ChatGPT:
“Give me a very conservative and literal answer, without speculating or adding your own nuances”
In practice we are asking for a behavior equivalent to a low temperature. If instead we write:
“Be as original as you can. I prefer unconventional, even risky, ideas over the most obvious ones”
The model will behave much more like it would with the temperature at maximum.
Something similar happens with response length, although the mechanism is different. The model generates text until it decides to stop or until it reaches a limit of tokens fixed in advance, and a longer answer does not equal a better answer. It almost never is.
The solution is simple: just include an explicit extension instruction in the ‘prompt’
Without instructions, most models tend to extend with nuances that we do not always need. On the other hand, if we are looking for in-depth development, a short answer will remain half. The problem does not lie in the model: it is that we have not indicated how much space it should occupy.
The solution is simple: just include in the prompt an explicit extension instruction, from “answer me in a single sentence” to “develop it as deeply as possible.” And the two parameters are also combined: a low temperature and a short response give surgical responses, ideal for technical purposes. However, a high temperature and generous length enable an AI that expands and takes risks, perfect for generating new ideas.
Know both parameters, even if it is indirectly through the itinerary that we mark for the model in each promptallows us to ask exactly what we need instead of settling for what AI delivers by default.
Image | Generated by Xataka with a prompt created by Gemini and submitted to ChatGPT
In Xataka | ChatGPT blocking mode: what it is, what it is for, who can use it and how to activate it
In Xataka | AI is replacing one of the most hated jobs in the world: the tailcoat collector
GIPHY App Key not set. Please check settings