Chaos and Order

💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

Introduction

The most common way of explaining an expert is by quantity. They know more, they memorised more, they have done more.

That explanation is only half right. And the wrong half is the more useful one in practice. The direction the material accumulated since the 1970s chess experiments points in is this: the decisive difference between expert and beginner is not the amount of stored information but what is seen when both look at the same scene.

This post follows that evidence. It starts with chess, moves to sport and language, and at the end sets out what the finding demands of how we practise. Somewhere in the middle there is a point where I stop because I could not find a source to cite. I have left that place as it is.

1. The five-second experiment

Show a chessboard for five seconds and take it away. Then ask the person to reproduce the arrangement they just saw on an empty board.

On this task chess masters overwhelmingly outperform beginners. That result on its own is not surprising. People who are good at something are good at it. What is interesting is the interpretation attached to the result. For a long time people read it as a difference in memory. Masters have better heads, so they hold more.

The paper Chase and Simon published in Cognitive Psychology in 1973 broke that interpretation (Chase & Simon, 1973 bibliographic record, checked 2026-08-16). I should disclose one thing here. I did not obtain the full text of this original paper this time; I confirmed only the bibliographic information. I read the content through later studies that cite this paper.

The piece Bartlett and colleagues wrote in Frontiers in Human Neuroscience in 2013 summarises the classic like this. Chess experts remember normal arrangements far better than random ones, and on normal arrangements their recall is far greater than that of beginners or lower-rated players (Bartlett et al., 2013, read 2026-08-16).

2. The random arrangement as the decisive condition

The heart of the experiment is the second condition. The same number of pieces is shown, scattered at random into positions that could not arise in an actual game.

Here most of the master's advantage disappears. If they had remembered well because they had good memories, they should remember the random board well too, and they do not. Only one explanation is left. What the master remembered was not the individual pieces but the shapes they already knew.

Bartlett and colleagues explain this advantage as a result of chunking ability rather than memory capacity. Thanks to knowledge structures stored in long-term memory, the expert can encode the whole board as a relatively small number of patterns, while the beginner, having no such structures, cannot do that bundling.

An important qualification attaches here. Popularly, this experiment tends to be summarised as "on a random board, master and beginner are the same," but later research is more precise than that. According to the Gobet and Simon data of 1996 that Bartlett and colleagues cite, recall on random boards also correlates positively with playing strength. It is just far weaker than the correlation on normal arrangements.

So the accurate sentence is this. On random arrangements the expert advantage does not disappear; it shrinks greatly. Saying it vanishes entirely is an exaggeration; saying it is preserved intact misses the original finding.

The design of the experiment is itself a tool. When someone claims to possess a particular ability, you can find out whether that ability is really general or attached to a particular pattern by constructing a condition with the pattern removed. The random board is that removal condition. And the fact that under such removal conditions the expert advantage shrinks greatly is exactly the same story as the absence of far transfer we saw in the previous post, wearing a different face. Expertise is attached to a domain.

Seen this way, you also notice how much everyday self-assessment happens with no removal condition at all. The tools you always use, the kinds of problem you always handle, the same opponents. The sense of being good inside that is accurate information, but you have to read narrowly what it is information about.

3. Chunks, and the corrections that followed

The study Gobet and Simon published in Memory & Cognition in 1996 made this picture precise (Gobet & Simon, 1996, read 2026-08-16).

Three things stand out.

First, the comparison between computer simulation and human experiment matched the existing estimate that a chess master stores about 50,000 chunks in memory. More important than the figure 50,000 is the unit. The unit of storage is not a piece but a pattern.

Second, when arrangements were mirror-reflected, recall got worse. The authors wrote that this implies each chunk represents a specific pattern in a specific location. In other words, expert knowledge is not abstract principle but, to a considerable degree, concrete and location-bound form.

Third, the authors judged the original chunking theory insufficient to explain the observed recall and proposed a larger structure, the template. The very fact that the theory was revised once matters. Chunking is not a finished answer but a framework still being worked on.

The capacity side has to be looked at alongside it. The review Gilchrist wrote in Frontiers in Psychology in 2015 sets out that Miller's classic limit of five to nine was later revised to four or fewer, independent of chunk size (Gilchrist, 2015, read 2026-08-16).

Overlay the two and the picture sharpens. The number of chunks a person can handle at once does not increase. What increases is the size of the content that goes into one chunk. Expertise is a matter of enlarging the vessel, not of adding vessels.

4. Sport: the gaze that arrives before the ball

The same structure is observed in athletics. A large part of what we usually call reflexes is in fact prediction.

The meta-analysis Mann and colleagues published in the Journal of Sport and Exercise Psychology in 2007 quantified the difference between experts and non-experts across 42 studies and 388 effect sizes (Mann et al., 2007, read 2026-08-16).

The results ran along two lines. Experts led on response accuracy and response time. And systematic differences appeared in visual search behaviour. Experts had fewer fixations, each fixation lasted longer, and the interval called the quiet eye was longer.

A qualification attaches here too. The authors reported that sport type, the research paradigm used, and the mode of stimulus presentation significantly moderated these relationships. That is, there is no single fixed expert gaze pattern but a tendency whose size varies with conditions.

Still, the direction is the same as in chess. The expert is not moving faster; the expert is looking earlier. Not shifting the gaze around means the judgement about where to look has already been settled.

5. Language: chunks, not words

The same structure shows up in language. When you learn a foreign language, the stage of assembling words one at a time and the stage where whole pieces come out are qualitatively different.

The eye-movement study Carrol and Conklin published in Language and Speech in 2020 compared three kinds of formulaic expression: idioms, binomials, and collocations (Carrol & Conklin, 2020, read 2026-08-16).

All three had shorter reading times than control phrases. And while a substantial part of this processing advantage was explained by whole-string frequency, the additional factors at work differed by type. For idioms, frequency, familiarity, and decomposability; for binomials, predictability and semantic association; for collocations, mutual information.

What this result says is simple. What has frequently appeared together is processed as one. Part of being good at a language is not having a large vocabulary but having a larger chunk unit.

6. Code: here the citations stop

There is a widespread story that programmers see code the same way. Studies in the form of a direct transplant of the chess experiment: show normal code and scrambled code, and only the skilled show an advantage on the normal code.

I could not reach those primary sources this time. Secondary summaries were available in several places, but I will not present as evidence content I have not read in a verifiable form. So I am leaving this item empty, without citation.

Instead I will note only this. That the skilled read code in units of structure rather than lines is an empirically plausible analogy, not a claim I can supply evidence for in this post. It is better to keep analogy and evidence apart.

7. So how should practice change?

If perception is the centre of expertise, the goal of practice becomes making patterns rather than putting in more information. Three implications follow. These implications are proposals that follow logically from the research above; they are not each separately verified prescriptions.

First, you have to repeat situations, not single items. The sentence in which a word actually appears rather than a word list; the phase in which a technique becomes necessary rather than the technique alone. Patterns are made only inside context. The reason the master's advantage shrank on the random board is precisely that context was removed.

Second, you have to distinguish normal conditions from distorted conditions. If you want to check your own level, you have to try it in situations that are not the familiar shape. Being good only on normal arrangements means you have learned the patterns, and that is a good thing. You just must not mistake it for general ability.

Third, you have to enlarge the chunk deliberately. Practising again what is already automatic is waste, but practising the bundling of what you are still processing piece by piece has a direct effect. In a foreign language, whole expressions that frequently travel together; in code, naming recurring structures.

Fourth, you can diagnose your own stage by chunk size. Let me explain with a constructed example. If, listening to English in a meeting, each word arrives separately and you miss the next sentence while assembling the meaning, you are still at the single-item stage. If a whole sentence comes in as one chunk and you can instead spend attention on the speaker's stance or the logical structure, your chunk unit has grown. You can make the same diagnosis in a code review. Are you reading line by line, or does the kind of structure this part is come into view first? This is not an actual measurement instrument but a criterion constructed to put your own state into words.

The third and fourth items share a trap. When the chunk grows, performance in that domain becomes noticeably easier, and that ease feels like a general rise in ability. What has actually happened is the automation of a specific pattern, and step outside the pattern and you return to something close to a beginner's processing speed. Ease is a signal of progress, but not a signal of range.

8. Evidence grade by claim

Claim	Evidence grade	Source
Chess experts recall normal arrangements far better than beginners	Replicated	Chase & Simon 1973, Bartlett et al. 2013
The advantage shrinks greatly on random arrangements	Replicated	Bartlett et al. 2013
A weak positive correlation remains even on random arrangements	Replicated	Gobet & Simon 1996, cited in Bartlett et al. 2013
A master's store is estimated at about 50,000 chunks	Estimate, model-dependent	Gobet & Simon 1996
Short-term memory capacity is four or fewer, independent of chunk size	Replicated	Gilchrist 2015 review
Sport experts have fewer and longer fixations	Replicated, with moderators	Mann et al. 2007
Formulaic expressions carry a processing advantage	Replicated	Carrol & Conklin 2020
Programmers chunk code the same way	Source not obtained	No citation

What has not been verified

The largest gap is section 6. On chunking in code reading I could not supply evidence.

Second, section 7 of this post consists of proposals derived logically from the research. I did not directly cite a comparative experiment showing that pattern-centred practice beats item-centred practice. Plausible and verified are different things.

Third, chunking theory itself is not a final version. That Gobet and Simon proposed the template came out of the inadequacy of the original theory in the first place, and Gilchrist's review points out that the method of measuring chunks is itself still contested. The existing methods of inferring chunks from recall order see only the outcome and not the process of formation. In other words, the core concept of this field still carries a measurement problem.

Finally, perception is not all of expertise. The chess experiment is a recall task, and being good at recall and playing a good move are different matters. That perception is an important axis and that perception is the only axis have to be kept apart.

Closing

An expert is not someone who remembers better what the beginner sees, but someone who sees something different in the first place. Where the beginner counts pieces, the master counts shapes.

The reason this matters practically is that it changes the direction of effort. Trying to memorise more is trying to add vessels, and that does not work well. Coming to recognise the shapes that appear often, by contrast, is enlarging the vessel, and that does work.

Fourteen Things Table Tennis Taught Me uses the word feel. The body knowing where the ball will go before you do. Translated into research language, that is exactly the subject of this post. A word that came out of experience and a word that came out of the laboratory are pointing at the same thing.

Previous post: What Studying Well and Playing Sport Well Share Is Narrower Than It Looks — what transfer research knocks down.
Next post: The Structure of Practice — spacing, retrieval, mixing. And why practice that goes well and practice that lasts differ.
Memory Lab — a tool for practising memorisation in bundled chunks.

References

Below are the sources checked directly on 2026-08-16 while writing this post. The links are the addresses actually read, and some are public API responses from Europe PMC and Semantic Scholar.