ContentsThe library

Models That Carry a Memory

A Rule That Reads What Arrived Before Deciding

Last timeA Linear Recurrence Done Properly

A fixed decay rate treats every token the same, so the state fills with filler. Letting the current input set the keep and write amounts turns the state into something selective.

Everything so far treats the input as something to be absorbed. The slot keeps

its fixed fraction and adds what arrived, at every position, with no regard for

what arrived. That is a strange way to run a limited memory, and it is the last

thing standing between a fixed state and a useful one.

The problem with a constant rate

Consider a model holding one fact through a passage. The fact arrives, the slot

takes it, and then four hundred words of unremarkable text go by. Each of those

words costs the slot the same fraction of its contents as the fact did, because

the rate does not know the difference. By the time the fact is needed it has

been diluted four hundred times by material that mattered less than it did.

FIG 1A fact arriving, then four hundred steps of filler
stepstephow much the step matteredkeep factor actually usedshare of the fact still heldwhat happened
1110.991The fact arrives and is written in.
2500.10.990.61Fifty steps of low-value text have each taken a hundredth, and nearly forty percent is gone.
32000.10.990.13The keep factor has not changed, because nothing about it can change.
44000.90.990.02The question finally arrives and the fact is effectively gone. Nothing was wrong with the rate; it was simply applied to the wrong things.
4 steps
The third column is constant and that is the entire problem. It was chosen for an average over the whole corpus, so it is wrong for every individual stretch of text, and no single value could have been right.

This is not a small effect that shows up only on contrived tasks. Natural text is

enormously uneven in how much each token is worth remembering: a name, a number

or a negation can change the meaning of everything that follows, while long

stretches of connective material can be dropped with no loss. A rule that

treats those the same is throwing away most of the capacity it has. The models

in the previous lesson were already good at holding things for a long time; what

they could not do was decide which things.

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 contents

This 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

The rest of this course

  1. 01Everything You Remember Has to Fit in the Same Box
  2. 02The Same Small Rule, Run a Thousand Timesopening only
  3. 03Almost Right, Four Hundred Times in a Rowopening only
  4. 04Add Instead of Multiply, and Decide How Muchopening only
  5. 05A Thousand Waits That Did Not Have to Happenopening only
  6. 06Slots That Forget at Rates You Choseopening only
  7. 07A Rule That Reads What Arrived Before Decidingyou are here
  8. 08The Bill That Grows and the Bill That Does Notopening only

Read alongside