Adjust raw model outputs (logits) and watch softmax convert them into a probability distribution in real time.
Softmax converts a list of raw, unbounded numbers (called logits, the raw output of a model's final layer) into a valid probability distribution, values that are all positive and sum to exactly 1. This is essential any time a model needs to choose among multiple discrete options, classifying an image into one of many categories, or predicting the next word from an entire vocabulary. The formula is straightforward: exponentiate each logit, then divide by the sum of all the exponentials, this exponentiation is what guarantees every output is positive, and the division by the sum is what guarantees they add up to 1.
You could imagine simply dividing each logit by the sum of all logits to get something that sums to 1, but this fails badly with negative logits, and it doesn't capture the intuitive property that softmax has: larger logits should dominate the probability distribution more than proportionally. Exponentiation amplifies differences, a logit that's only slightly larger than another becomes substantially more probable after exponentiation, this is precisely the behavior you want when the model needs to express genuine confidence in its best answer.
Real softmax implementations, including the one powering this visualizer, subtract the maximum logit from every value before exponentiating. This doesn't change the final result mathematically (the max cancels out in the division), but it prevents a genuine practical problem, exponentiating a large logit directly can produce a number too large for floating-point arithmetic to represent, causing overflow errors. This small trick is a standard, essential detail in every production machine learning framework's softmax implementation, not an optional optimization.
Adjust the temperature slider and watch the effect directly. A temperature below 1 sharpens the distribution, the highest logit's probability increases further, making the model's choice look more confident and decisive. A temperature above 1 flattens the distribution toward uniform, all classes become more equally likely regardless of their original logit differences. This is exactly the mechanism behind "temperature" settings in AI chat interfaces and text generation APIs, low temperature produces more predictable, deterministic-feeling output, high temperature produces more varied, sometimes surprising output, both are the same underlying model, just with softmax's output distribution reshaped before sampling.
Every classification model with more than two output classes ends with a softmax layer. Every language model uses softmax to convert its raw output scores for the entire vocabulary into a probability distribution over "what word comes next," this is literally the mechanism by which GPT and Claude decide what token to generate at each step. And the attention mechanism inside transformer architectures, the core innovation behind modern LLMs, uses softmax internally to determine how much "attention weight" each token should receive relative to others.