Contents

What is a mental model?

Mental Model 1: Problem solving is search
Mental Model 3: Knowledge grows exponentially
Mental Model 4: Creativity is mostly copying
Mental Model 5: Skills are speci c
Mental Model 6: Mental bandwidth is extremely limited
Mental Model 7: Success is the best teacher
Mental Model 8: We reason through examples
Mental Model 9: Knowledge becomes invisible with experience
Mental Model 10: Relearning is relatively fast
What is a mental model?
A mental model is a general idea that can be used to explain many
di erent phenomena. Supply and demand in economics, natural
selection in biology, recursion in computer science, or proof by
induction in mathematics—these models are everywhere once you
know to look for them.
Just as understanding supply and demand helps you reason about
economics problems, understanding mental models of learning will
make it easier to think about learning problems.
Unfortunately, learning is rarely taught as a class on its own—meaning
most of these mental models are known only to specialists. In this
essay, I’d like to share the ten that have influenced me the most, along
with references to dig deeper in case you’d like to know more.
Mental Model 1: Problem solving is
search
Herbert Simon and Allen Newell launched the study of problem solving
with their landmark book, Human Problem Solving. In it, they argued
that people solve problems by searching through a problem space.
A problem space is like a maze: you know where you are now, you’d
know if you’ve reached the exit, but you don’t know how to get there.
Along the way, you’re constrained in your movements by the maze’s
walls.

Problem spaces can also be abstract. Solving a Rubik’s cube, for

instance, means moving through a large problem space of
con gurations—the scrambled cube is your start, the cube with each
color segregated to a single side is the exit, and the twists and turns
de ne the “walls” of the problem space.
Real-life problems are typically more expansive than mazes or Rubik’s
cubes—the start state, end state and exact moves are o en not clear-
cut. But searching through the space of possibilities is still a good
characterization of what people do when solving unfamiliar problems
—meaning when they don’t yet have a method or memory that guides
them directly to the answer.
One implication of this model is that, without prior knowledge, most
problems are really di cult to solve. A Rubik’s cube has over forty-
three quintillion con gurations—a big space to search in if you aren’t
clever about it. Learning is the process of acquiring patterns and
methods to cut down on brute-force searching.
Mental Model 2: Memory
strengthens by retrieval
Retrieving knowledge strengthens memory more than seeing
something for a second time does. Testing knowledge isn’t just a way
of measuring what you know—it actively improves your memory. In
fact, testing is one of the best study techniques researchers have
discovered.

Why is retrieval so helpful? One way to think of it is that the brain

economizes e ort by remembering only those things that are likely to
prove useful. If you always have an answer at hand, there’s no need to
encode it in memory. In contrast, the di culty associated with retrieval
is a strong signal that you need to remember.
Retrieval only works if there is something to retrieve. This is why we
need books, teachers and classes. When memory fails, we fall back on
problem-solving search which, depending on the size of the problem
space, may fail utterly to give us a correct answer. However, once we’ve
seen the answer, we’ll learn more by retrieving it than by repeatedly
viewing it.
Mental Model 3: Knowledge grows
exponentially
How much you’re able to learn depends on what you already know.
Research nds that the amount of knowledge retained from a text
depends on prior knowledge of the topic. This e ect can even
outweigh general intelligence in some situations.

As you learn new things, you integrate them into what you already
know. This integration provides more hooks for you to recall that
information later. However, when you know little about a topic, you
have fewer hooks to put new information on. This makes the
information easier to forget. Like a crystal growing from a seed, future
learning is much easier once a foundation is established.
This process has limits, of course, or knowledge would accelerate
inde nitely. Still, it’s good to keep in mind because the early phases of
learning are o en the hardest and can give a misleading impression of
future di culty within a eld.
Mental Model 4: Creativity is
mostly copying
Few subjects are so misunderstood as creativity. We tend to imbue
creative individuals with a near-magical aura, but creativity is much
more mundane in practice.

In an impressive review of signi cant inventions, Matt Ridley argues

that innovation results from an evolutionary process. Rather than
springing into the world fully-formed, new invention is essentially the
random mutation of old ideas. When those ideas prove useful, they
expand to ll a new niche.
Evidence for this view comes from the phenomenon of near-
simultaneous innovations. Numerous times in history, multiple,
unconnected people have developed the same innovation, which
suggests that these inventions were somehow “nearby” in the space of
possibilities right before their discovery.
Even in ne art, the importance of copying has been neglected. Yes,
many revolutions in art were explicit rejections of past trends. But the
revolutionaries themselves were, almost without exception, steeped in
the tradition they rebelled against. Rebelling against any convention
requires awareness of that convention.
Mental Model 5: Skills are specific
Transfer refers to enhanced abilities in one task a er practice or
training in a di erent task. In research on transfer, a typical pattern
shows up:
Practice at a task makes you better at it.
Practice at a task helps with similar tasks (usually ones that
overlap in procedures or knowledge).
Practice at one task helps little with unrelated tasks, even if
they seem to require the same broad abilities like “memory,”
“critical thinking” or “intelligence.”

It’s hard to make exact predictions about transfer because they

depend on knowing both exactly how the human mind works and the
structure of all knowledge. However, in more restricted domains, John
Anderson has found that productions—IF-THEN rules that operate on
knowledge—form a fairly good match for the amount of transfer
observed in intellectual skills.
While skills may be speci c, breadth creates generality. For instance,
learning a word in a foreign language is only helpful when using or
hearing that word. But if you know many words, you can say a lot of
di erent things.
Similarly, knowing one idea may matter little, but mastering many can
give enormous power. Every extra year of education improves IQ by 1-5
points, in part because the breadth of knowledge taught in school
overlaps with that needed in real life (and on intelligence tests).
If you want to be smarter, there are no shortcuts—you’ll have to learn a
lot. But the converse is also true. Learning a lot makes you more
intelligent than you might predict.
Mental Model 6: Mental bandwidth
is extremely limited
We can only keep a few things in mind at any one time. George Miller
initially pegged the number at seven, plus or minus two items. But
more recent work has suggested the number is closer to four things.

This incredibly narrow space is the bottleneck through which all

learning, every idea, memory and experience must flow if it is going to
become a part of our long-term experience. Subliminal learning
doesn’t work. If you aren’t paying attention, you’re not learning.
The primary way we can be more e cient with learning is to ensure the
things that flow through the bottleneck are useful. Devoting bandwidth
to irrelevant elements may slow us down.
Since the 1980s, cognitive load theory has been used to explain how
interventions optimize (or limit) learning based on our limited mental
bandwidth. This research nds:
Problem solving may be counterproductive for beginners.
Novices do better when shown worked examples (solutions)
instead.
Materials should be designed to avoid needing to flip between
pages or parts of a diagram to understand the material.
Redundant information impedes learning.
Complex ideas can be learned more easily when presented rst
in parts.
Mental Model 7: Success is the best
teacher
We learn more from success than failure. The reason is that problem
spaces are typically large, and most solutions are wrong. Knowing
what works cuts down the possibilities dramatically, whereas
experiencing failure only tells you one speci c strategy doesn’t work.

A good rule is to aim for a roughly 85% success rate when learning. You
can do this by calibrating the di culty of your practice (open vs.
closed book, with vs. without a tutor, simple vs. complex problems) or
by seeking extra training and assistance when falling below this
threshold. If you succeed above this threshold, you’re probably not
seeking hard enough problems—and are practicing routines instead of
learning new skills.
Mental Model 8: We reason
through examples
How people can think logically is an age-old puzzle. Since Kant, we’ve
known that logic can’t be acquired from experience. Somehow, we
must already know the rules of logic, or an illogical mind could never
have invented them. But if that is so, why do we so o en fail at the
kinds of problems logicians invent?
In 1983, Philip Johnson-Laird proposed a solution: we reason by
constructing a mental model of the situation.
To test a syllogism like “All men are mortal. Socrates is a man.
Therefore, Socrates is mortal,” we imagine a collection of men, all of
whom are mortal, and imagine that Socrates is one of them. We
deduce the syllogism is true through this examination.

Johnson-Laird suggested that this mental-model based reasoning also

explains our logical de cits. We struggle most with logical statements
that require us to examine multiple models. The more models that
need constructing and reviewing, the more likely we will make
mistakes.
Related research by Daniel Kahneman and Amos Tversky shows that
this example-based reasoning can lead us to mistake our fluency in
recalling examples for the actual probability of an event or pattern. For
instance, we might think more words t the pattern K _ _ _ than _ _ K _
because it is easier to think of examples in the rst category (e.g.,
KITE, KALE, KILL) than the second (e.g., TAKE, BIKE, NUKE).
Reasoning through examples has several implications:
1. Learning is o en faster through examples than abstract
descriptions.
2. To learn a general pattern, we need many examples.

3. We must watch out when making broad inferences based on a

few examples. (Are you sure you’ve considered all the possible
cases?)

Mental Model 9: Knowledge

becomes invisible with experience
Skills become increasingly automated through practice. This reduces
our conscious awareness of the skill, making it require less of our
precious working memory capacity to perform. Think of driving a car: at
rst, using the blinkers and the brakes was painfully deliberate. A er
years of driving, you barely think about it.

The increased automation of skills has drawbacks, however. One is

that it becomes much harder to teach a skill to someone else. When
knowledge becomes tacit, it becomes harder to make explicit how you
make a decision. Experts frequently underestimate the importance of
“basic” skills because, having long been automated, they don’t seem
to factor much into their daily decision-making.
Another drawback is that automated skills are less open to conscious
control. This can lead to plateaus in progress when you keep doing
something the way you’ve always done it, even when that is no longer
appropriate. Seeking more di cult challenges becomes vital because
these bump you out of automaticity and force you to try better
solutions.
Mental Model 10: Relearning is
relatively fast
A er years spent in school, how many of us could still pass the nal
exams we needed to graduate? Faced with classroom questions, many
adults sheepishly admit they recall little.

Forgetting is the unavoidable fate of any skill we don’t use regularly.

Hermann Ebbinghaus found that knowledge tapers o at an
exponential rate—most quickly at the beginning, slowing down as time
elapses.
Yet there is a silver lining. Relearning is usually much faster than initial
learning. Some of this can be understood as a threshold problem.
Imagine memory strength ranges between 0 and 100. Under some
threshold, say 35, a memory is inaccessible. Thus if a memory dropped
from 36 to 34 in strength, you would forget what you had known. But
even a little boost from relearning would repair the memory enough to
recall it. In contrast a new memory (starting at zero) would require
much more work.
Connectionist models, inspired by human neural networks, o er
another argument for the potency of relearning. In these models, a
computational neural network may take hundreds of iterations to reach
the optimal point. And if you “jiggle” the connections in this network,
it forgets the right answer and responds no better than if by chance.
However, as with the threshold explanation above, the network
relearns the optimal response much faster the second time.1
Relearning is a nuisance, especially since struggling with previously
easy problems can be discouraging. Yet it’s no reason not to learn
deeply and broadly—even forgotten knowledge can be revived much
faster than starting from scratch.
What are the learning challenges you’re facing? Can you apply one of
these mental models to see it in a new light? What would the
implications be for tackling a skill or subject you find difficult? Share
your thoughts in the comments!

1 ese networks are trained via gradient descent. Gradient descent works by essentially rolling
downhill. Correct knowledge is like the gently-sloping bottom of a steep canyon—the correct
direction is down the canyon, but the sides are quite high. Unlike a three-dimensional space, as would
describe a physical canyon, most networks are in an extremely high-dimensional space. at means
any imprecision in the direction results in running up the side of the canyon. e result is that
networks typically slosh around a lot before getting to the bottom of the long canyon. However, when
you add any noise to the system, the “downhill” direction usually goes straight back to the optimal
point.
 My book, ULTRALEARNING, is now
available.

Being able to master hard skills quickly is essential to your work and life. is book will show you
how.

Combining stories of dramatic ultralearning feats with the detailed science on how to learn
anything e fectively, ULTRALEARNING will transform how you work, live and learn.

Learn More:
https://www.scotthyoung.com/blog/ultralearning/