The Improviser · Lesson 12 — The Arrangement ← Course

The arrangement — a band that's never all playing.

You've heard this your whole life. A singer out front, and behind them two or three parts at a time — never the whole band at once. That is the everyday version of the last page.

Listen to any record with a band on it. The singer carries the tune from the first bar to the last. Behind the singer, the band is rarely all in: a piano and a bass under the verse, the horns arriving for the chorus, strings for one bridge and then gone. At any moment most of the chairs are quiet — and you never notice, because what you hear is one band with one sound. The arrangement is the plan for who plays when, and it is how a large band stays light on its feet.

Map that onto the model and the pieces land neatly. The singer is the part that is always on: the inner ear (Lesson 8), the listening-back that every note passes through. It is shared by the whole band and never split. The sections behind the singer are the experts — the block after the inner ear, split into groups, two of which play each note. The listener hears one voice; the bill is paid by a duo.

Now the two places where the model is not a band, and both matter. First, an arranger writes the handoffs in advance, in ink, on the chart — the written arrangement every player reads from — and for musical reasons: horns for the chorus because horns lift a chorus. The router writes nothing in advance. It decides live, note by note, from what it is hearing in context — and the cues it uses were learned in training, not chosen. Second, the arranger's sections have names and jobs. The router's don't. Look inside a trained model and the groups it wakes together seldom line up with any topic, style, or instrument a person would name. They are simply the groups that training found useful to keep apart.

So keep the band picture, because it is right about the thing that matters: a large ensemble can sound whole while only a few parts play, and that is the entire trick. Just don't let it tell you that the model knows which section is the saxophones. It doesn't need to.

A singer and four sections. Play the tune with the parts handed out by an arranger's chart — written in advance, by the phrase. Then hand them out by a router, live — decided note by note. Watch the singer stay lit throughout, count the chairs playing, and read who was chosen and why.

Start with the arranger's chart: the handoffs are written before a note is played.
the tune
the standwaiting
Singer — the inner earalways on
chairs playing this notesinger + 0 of 4

A chart is a lookup. A router is a decision.

Same band, same sections, same "two at a time". The only difference is where the choice comes from: a table written in advance, or a small learned function of the note in its context.

# the arranger: who plays is written down, by phrase, before the gig
chart = {"verse": ["piano", "bass"], "chorus": ["horns", "bass"], ...}
playing = chart[phrase_of(note)]              # a lookup — same answer every night

# the router: who plays is decided from the note itself, right now
scores  = router(x)                          # x: this note in its context
playing = top_k(scores, k=2)                  # a decision — can differ for the same note elsewhere

# in both: the singer (attention) is always on; only the sections are chosen
y = singer(x) + sum(w * sections[s](x) for s, w in playing)

One voice, a few chairs, and a decision made live.

A real band already works the way a mixture of experts does: a singer who is always on, and an arrangement that has only a few sections playing under them at any moment, so a big ensemble stays cheap to run and still sounds whole. In the model the always-on part is attention, shared by everyone; the sections are the experts behind it, two per note. The differences are where the picture stops: the router chooses live, from context, by cues it learned rather than was given — and its sections have no names, no instruments, and no jobs a person would recognise.

The band tells you why it's cheap. The router tells you why it isn't a band: nothing is written in advance, and nobody in it knows which chairs are the horns.

Next lesson: The Review →
Go deeper — what the singer stands for, and what the sections don't optional

Why the singer is attention

Each layer of a transformer has two halves: attention, which looks back across the context, and a feed-forward block, which transforms each position on its own. In a mixture-of-experts model only the second half is split into experts. Attention is computed once for every token in every layer, by everyone — it is the shared, always-on part. That is why the singer never leaves the stage while the sections come and go.

Arranger versus router, precisely

An arrangement is a fixed mapping from position in the tune to parts. A router is a learned function from the token's current hidden state to a score per expert; the top-k are taken and their weights renormalised. Because the input is the hidden state and not the position, the same word can go to different experts depending on what surrounds it — and identical contexts route identically. Nothing is stored per song; the decision is recomputed every time.

Why the sections have no names

During training there is no signal that says "make expert 3 the music expert". The only pressure is to predict the next token well while spreading load across experts. What emerges is whatever partition happens to help — often groupings by token shape, position, or punctuation, sometimes nothing describable. Researchers who probe experts find weak, leaky tendencies at best. Naming them "the saxes" is a convenience for the reader, not a fact about the model.