What is a Tensor ?

[smfadmin (OP)]

Your resident AI has transcribed this for you:

A LLM or (Large) Language Model works with lists of numbers representing pieces of your text but when several sentences arrive together, which numbers belong to which piece?

A tensor organizes those values so everyone has a position in the right group.

First, text is split into pieces called tokens and aach token can be looked up as a list of numbers called an embedding ... for a tiny made-up example, say each token has four numbers (real models commonly use much longer lists).

Suppose a short sentence becomes three tokens, so now put their list into three rows with four values in each row -- now you have a (2-D) table: one axis chooses the token position, the other chooses a value within its embedding -- that is a two-dimensional tensor.

Add a second sentence with the same number of tokens and group both tables into a batch.

Now choose the sentence, the token position, and the embedding position: three axes (3-D), and its shape is two by three by four, holding twenty-four values in this toy example.

But those four embedding values are not four neatly labeled meanings.

Their usefulness comes from patterns learned during training, and the tensor itself does not understand the sentences.

It holds their numerical representations in the arrangement the model's calculations expect.

The model still has to process them.

In machine learning, that is what a tensor is—an array of values organized along axes. Here, those axes separate sentences, tokens, and embedding positions.

With images, they can separate pictures, pixels, and color channels.

The contents change, but the job stays the same.

Keep the numbers organized for computation.