The Favourite That Makes Itself
Last timeWhat the Experts Actually Learn
A copy that gets slightly more traffic early gets better, which earns it more traffic, which makes it better still. Left alone the arrangement collapses onto a handful of copies and wastes the rest.
A routed layer has a failure mode that is not a bug in anyone's code. It follows
from the structure, it happens reliably, and every working system contains
something specifically built to stop it.
The loop
Suppose that at initialisation one copy is very slightly favoured by accident.
Follow what happens.
What the collapse costs
Two separate bills, and they are worth separating because they fall on different
people.
The first is capacity. The whole proposition was that many copies hold more
knowledge than one. A copy that never receives tokens never learns anything, so
the parameters bought to hold it are inert. They still occupy memory, on every
machine, for the whole life of the model. The model has the deployment
difficulty of something very large and the useful capacity of something small.
The second is time. A batch is not finished until every part of it is finished,
so if one copy has been handed most of the batch while seven sit nearly empty,
the step takes as long as the overloaded copy needs. Seven eighths of the
hardware waits.
The lesson stops here
5 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