Attention in a language model.

Step 1 of 6

Tokens

A large language model, like those behind chatbots, is a huge neural network with one task: take text in and predict the next token. It does that again and again, token after token, and so it writes. The whole model is neurons and weights, as in the last Lessons, only billions of them.

On the rack is an English sentence, a tile per token, with its number in the vocabulary. Behind each tile is a column of lit cells, the token’s vector. The numbers here are made up for illustration, but they work as in a real model.

Step 2 of 6

Attention

Before it predicts, each token gathers information from the tokens before it, through an attention head: it gives each earlier token a weight, the shares add up to 100%, and its vector is updated by their weighted blend. It cannot look at the tokens after it: they have not been written yet. The Attending token is raised, and smoked glass covers what comes after it. Predict, and then Look inside will draw arcs, as thick as the attention.

Step 3 of 6

Sharp attention

In most places attention spreads over several tokens. Change the Sentence and the Attending token, and look for a place where one token takes nearly all the attention.

Step 4 of 6

Copying

In the zorblat sentence the last token is zor, for the second time. Look where it is looking, and at what came after the first zor.

Step 5 of 6

The next token

After attention, the model gives each possible token a probability of coming next. The wheel at the end of the rack is divided by the probabilities, with one more hatched slice for the long tail of everything else. Then it samples: spins the wheel and picks. The Temperature decides how: it raises each probability to the power $\dfrac{1}{T}$⁠ and rescales them. Predict what a low temperature will do.

Step 6 of 6

Temperature

At temperature $1$⁠ the wheel is the model’s distribution as it is. Tune it until the leading candidate gets the chance in the question.