Full-Bandwidth Transformer

Every decoding step computes a full hidden state, then squeezes it down to a single token id before the next step can use it. Training fills the token×layer grid one row at a time; inference fills it one column at a time. Hover any cell to see what it can and can't read — and how much of that 17-bit ceiling is going unused.

A teacher-forced pass can’t see its own latents, so training runs the pass again — handing the top-layer latent from the end of one pass back to the start of the next. Watch the exact prefix grow one token per pass.