What Is Attention Actually Comparing?

Open with the mechanical feeling of attention: a token is read, and the model's 'focus' shifts. Pose the driving question visually with two arrows pointing at each other — what is on each side of the comparison?

• Attention is often described as 'looking back' at previous words
• But the actual operation is more specific than that
• Driving question: what two things are being compared when a head scores a match?

A single forced-choice question: when the model attends from the word 'sat' in 'The cat sat on the', what is it actually comparing? Commit to an intuition before the mechanism is shown.

• Choose the most concrete comparison you think is happening
• No penalty for being wrong — the point is to surface the assumption

A small, focused simulator: pick a token, and watch three vectors light up — the query (blue), one key from the bank (orange), and the value (green). The learner can swap keys and see the dot-product score change in real time.

• Every token produces three vectors: Q, K, and V
• The query is the seeker; the key is the address; the value is what gets returned
• Only the query is compared — against keys, not against words

Show a real-looking attention matrix as a heatmap: rows are queries, columns are keys. Highlight the row for 'sat' and let the brightest cell point to 'cat' — surprising, because 'cat' is a noun, not a verb. The evidence is that the match is not syntactic.

• Rows are queries, columns are keys, cells are dot products
• Bright cells reveal where the model 'attends'
• The match can be semantic, positional, or syntactic — not always what grammar would predict

Step through the pipeline: token → embedding → linear projection into Q, K, V. The comparison happens in the projected space, not on the string. Show that two very different words can have nearby vectors, and two identical words can have different vectors in different positions.

• Tokens are mapped into a high-dimensional space by learned matrices
• Q and K live in that space; similarity is geometric, not lexical
• Same word, different position → different Q, different K, different score

A manipulable scene: the learner can change the sentence and watch the attention row for the new verb shift. Predict where the brightest cell will land, then verify. The principle should hold across a new example.

• Apply the same Q-vs-K comparison to a fresh sentence
• Predict first, then observe
• The attention pattern is not about matching words — it is about matching learned vectors

A short, sharp boundary scene: attention does not compare words to words, does not store facts, does not 'understand' meaning. It is a differentiable, learned similarity lookup — and saying so precisely prevents overclaiming.

• Attention is a soft key-value retrieval, not a database
• Heads do not 'know' facts; they pattern-match in projection space
• The comparison is purely geometric: how aligned two vectors are

Resolve the driving question directly: attention compares a query vector to a bank of key vectors via dot product; the scores softmax into weights; the weights mix the value vectors. Close by naming the original intuition that was wrong.

• Compared: Q vs K (both are learned projections)
• Produced: a scalar similarity score per key
• Used: to weight the values that are summed into the output
• The first guess — comparing words to words — was the layer of illusion