Part 2 · Intuition

Bidirectional sight

A subtler gift: when the typewriter predicts a token, it can only look backward — the future doesn’t exist yet. A dLLM filling in a blank sees committed tokens on both sides. Predicting the blank in “the ▢ sat on the mat” is much easier when “sat on the mat” is already visible. Architecturally this is a small change: drop the causal attention mask so every position attends to every other — the transformer itself is otherwise the same machine.