The Bill That Grows and the Bill That Does Not
Last timeDeciding What to Keep
A look-back model stores something for every token it has seen, so serving cost climbs with the conversation. A carried state is one fixed block, and that difference decides what can be deployed.
Everything up to here has been about whether a fixed state can be made good
enough. This lesson is about why anyone wanted it to be, and the answer is
visible on an invoice.
What a conversation costs to keep
A look-back model does not recompute its history. It stores two vectors for
every token it has processed, in every layer, so that each new token can consult
them. That store is the reason generation is fast, and it is also a quantity
that grows without limit as a conversation continues.
| 1 thousand tokens | 8 thousand | 64 thousand | 512 thousand | |
|---|---|---|---|---|
| a growing store, one ent | 0.5 | 4.0 | 32.0 | 256.0 |
| a carried state of fixed | 8.0 | 8.0 | 8.0 | 8.0 |
There is a second growth hiding behind the first. Each new token has to consult
every stored entry, so the arithmetic for one token also grows with the length
of the conversation. In practice reading the store out of memory, rather than
the arithmetic itself, is the slow part, so the per-token time climbs roughly in
step with how much has been said. A reply at the end of a long session is
genuinely slower to produce than the same reply at the start, and users notice
it before operators do.
The lesson stops here
6 more paragraphs to go
You have read the opening. The rest of the argument, the problems that check whether it landed, and the lines worth keeping at the end all come with a plan.
The first lesson of every course in the library reads the whole way through, free, so you can see exactly what the rest of them are.
See the planThe contentsThis is the reading half
Starting the course gives you your own copy of it. Every idea on every page has problems standing under it, marked with a reason rather than a tick, and any sentence you do not believe can be opened and argued with. None of that can happen on a page nobody owns.
The contents