Don't Tell Me What I Already Know

Priya asks about the consensus protocol, and the system knows something about her. She has worked on consensus systems for two years. She has given talks on the subject. So when it prepares her digest it leaves out the speaker’s definition of partial synchrony as something she plainly knows, and shows her the parts that look new. It is a considerate thing to do. It is also a mistake, because the speaker uses the term in a narrower sense than the one Priya works with. Her usual definition allows delays whose bound is unknown. In this talk the bound is known to the protocol, and everything the speaker later says about safety rests on that. Priya reads the digest, meets the claim about bounded delays, and concludes the talk contradicts what she knows. The system had removed the one piece of information she needed in order to read the rest.

The word was familiar. The meaning was not what she knew. That gap is the subject of this chapter. The last one fixed the source and the task and found that two people still need different things, because what is new to one is not new to the other. The question now is:

Given the same source and the same task, what would actually add something for this particular person?

It is a narrow question, and the narrowness is deliberate.

A model of what you know, and nothing else

Answering it needs a model of the reader. That is a large step, and it is worth saying at once how small this model should be. Three different things might be meant by “a model of the person”:

MODEL OF WHAT YOU KNOW
   ≠
MODEL OF WHAT YOU LIKE
   ≠
MODEL OF WHO YOU ARE

This chapter needs only the first. The system does not need to infer a personality, a taste, an identity, a demographic trait, a political view, a learning style, or the kind of person someone is. It needs something much smaller: what information does the system have reason to believe this person already knows, partly knows, or does not yet know? A person who knows a great deal about consensus protocols and dislikes diagrams is two facts, and only the first bears on whether to show the definition of partial synchrony. The second belongs to different questions, which come later and carry risks of their own.

The transformation here is:

same source
same task
different prior knowledge
        ↓
different useful representation

The phrase “adaptation to prior knowledge” will be used for it in preference to “personalisation”. Personalisation carries baggage: preference optimisation, profile building, recommendation, persuasion. None of that is needed to decide whether a definition should be shown.

Novelty is relational too

The last chapter said relevance is a relation among a unit, a purpose and a context. Novelty is a relation too:

NOVELTY(unit, person, context)

Take the definition of a fault from the talk. For Priya, who has worked on this for years, it is redundant. For Tom, who is writing a history and has read about the protocols in passing, it is useful reinforcement. For Mina, who is meeting the area for the first time, it is entirely new. The unit and the task are the same. The value differs because the people differ. That gives a third distinction alongside the ones already drawn:

new in the source  ≠  needed for the task  ≠  new to the person

A unit can be something the source introduces, central to what the reader is trying to do, and already part of what she knows. Each of those is a separate fact.

What the system may say it knows

Once a model of knowledge is in play, a few rules keep it honest, and they all come from one observation: the system’s picture of what someone knows is built from evidence, and evidence is not the same as the thing it is evidence of.

evidence about knowledge  ≠  knowledge itself

Because of this, a knowledge state should be graded and not binary. The simplest useful version has three values:

KNOWN      strong evidence the person already knows this
PARTIAL    evidence of some understanding, not enough to rely on
UNKNOWN    insufficient evidence either way

Evidence can also conflict, a recent mistake against an old demonstration, say. In that case the honest state is uncertainty, and for the purposes of deciding what to omit it can be treated like the third value.

Several smaller separations protect the model from pretending to more than it has:

not observed  ≠  not known
mentioned before  ≠  mastered
familiar term  ≠  known meaning

The first guards against assuming ignorance: Priya’s expertise left no trace in the logs the system can see. The second guards against assuming mastery: a topic that appeared in something she opened last year has not thereby been understood. The third is the failure that opened the chapter. A word appearing is not the same as the person holding the meaning the source gives it.

Where does the evidence come from? Roughly, in descending order of strength: the person said she knows it, she used it correctly in her own work, she answered a question about it, she asked for an explanation of it earlier, she corrected the system about it, and she has repeatedly worked with material that uses it. Each of these is worth something. None is certain. And they decay: something demonstrated two years ago may be rusty now. The weakest evidence of all is simply that the person was exposed to something. Having had it shown is not the same as having read it, having read it is not the same as understanding it, and understanding it is not the same as remembering it a year later.

A graded state recorded this way is very different from a dossier. It records only what is needed to improve the communication: this concept, this evidence, this date, this level of confidence. It is visible to the person and open to correction, in plain terms: the system believes you know embeddings and cosine similarity; it is unsure about product quantisation; correct this, forget this, show the evidence. Knowledge about someone’s knowledge that she cannot see or fix is a hidden decision-maker. Questions about how long such records may be kept, and who may see them, are taken up when the book reaches the person in full, and nothing here depends on settling them.

Known is not safe to omit

The failure that opened the chapter generalises into a rule. A system should not drop something merely because it looks familiar, if later material depends on it.

Consider three units in a chain, where each depends on the one before:

A  →  B  →  C

Suppose the system believes the reader knows A and B, but is unsure of B’s exact role in this source. If it skips B, then C may become incomprehensible. A reader who knew A and B in some sense is stranded, because the meaning that C needs is the one the source gives to B, not the one she carries.

So suppression should require a conjunction, never a single flag:

probably known  AND  not needed to interpret what is shown
        ↓
safe to omit

and the presentation should follow confidence and not just a yes or no. High-confidence knowledge of something nothing else depends on is a candidate for omission. Medium confidence suggests compressing it or offering it collapsed. Uncertainty means keeping it. And anything that later material depends on stays, or at least stays within one click. This is the preservation obligation from Part I turned on the reader: novelty filtering must preserve the prerequisites reconstruction requires.

known  ≠  safe to omit

Two errors are possible, and they are not equally costly. Showing something the reader knew wastes a little of her attention. Hiding something she needed removes information she cannot know is missing and may mislead her about everything after it. The second is worse, so when the evidence is thin, ties go to showing.

What this is not

A few boundaries keep the narrow model narrow.

I know this  ≠  I like this
I have seen this  ≠  I know this
I often open this  ≠  I need more of this
I rarely open this  ≠  I already understand it

Interest and knowledge are different variables. A system that tracks what someone clicks is learning about her interests, and using it to decide what she knows will go wrong in both directions: someone who reads five articles on a topic out of curiosity may understand none of it, and an expert who never reads about her own field will look like a novice. The logic of recommendation, which gives people more of what they engage with, must not take over a layer whose job is to find what adds something. Interest modelling is a different job with different variables.

Where this leaves us

Personal novelty is the relation between a unit, a person and a context, and it needs only a narrow model of what the person knows: graded, evidence-based, scoped to the concept, and visible. The evidence is not the knowledge, a familiar word is not a known meaning, and known is not safe to omit, since later material may depend on the meaning the source gives. A system that treats novelty as a filter will sometimes bore the reader, which is cheap. One that treats it as permission to delete will sometimes strand her, which is not.

Even this is not enough. Two people can know exactly the same things and still need different representations, because one is scanning on a phone between meetings and the other is reading carefully at a desk, because one is trying to remember and the other to act, because one is interrupted and the other is not. The task and the prior knowledge describe what the person is after and what she has. They do not describe the situation she is in right now. The book returns to that, because it is where adaptation starts to need real care about what a system may infer. Before it does, something simpler is missing from this part’s account. Everything so far has happened because someone asked. But a person’s interests persist when she is not asking, and the most valuable thing the system could tell her may be something she has not yet thought to look for.