What a Course Cannot Buy

A gap announces itself at an inconvenient moment. Somebody puts a design in front of you that turns on a piece of technology you have read about and never used, and you have to say something useful about it by Thursday. You know the vocabulary. You have read the launch post and two summaries and could hold up your end at a conference. What you cannot do is say which of the three options will hurt in eighteen months, which is the only question anybody asked.

The reflex is to go and learn it, and the reflex after that is to reach for the thing that looks like learning: a course, a book, a weekend of videos at one and a half speed. I have done this many times and paid for others to do it, and the result is consistent in a way I found uncomfortable once I noticed it. The material makes sense while it is in front of you. Two weeks later you recognize every idea and can produce almost none.

That distance between recognizing and producing is not a matter of effort or attention. It is the whole subject. Recognition is what exposure gives you, it arrives quickly, it feels exactly like understanding from the inside, and it is worth very little in front of a blank editor and a decision. Recall under something like real conditions is what you needed, and nothing produces it except being made to do it.

So the sorting question about any method, any course, any team learning plan is narrower than it looks: at what point does this make me produce something with the material out of sight. If the answer is never, what was bought is a map. Maps are useful. Nobody ever crossed a river on one.

Being tested beats studying, in the setting where it was tested. Roediger and Karpicke's 2006 experiments had students read short prose passages and then either restudy them or take a recall test, with retention measured minutes, days and a week later. Restudying won immediately and lost badly at a week. That is a robust result about undergraduates and prose passages in a lab, and the leap from there to learning a storage engine is mine, made on judgment rather than evidence, because nobody has run that study.

The signal you are using to judge your own learning is the wrong one. What people report when they rate a course is fluency: it went down easily, the examples made sense, nothing snagged. Ease of processing feels like knowing and is close to unrelated to it, which is why rereading is the most popular study method and one of the weakest, why the courses rated highest are often the ones that demanded least, and why a training budget steered by satisfaction scores buys comfort.

Spacing is the one intervention that is nearly free. The 2006 meta-analysis by Cepeda and colleagues found that spreading the same total study time across separated sessions beat massing it, across a large number of verbal recall experiments. Same hours, better retention, purely from the arrangement. Verbal recall in a laboratory again, but the practical version costs nothing: four short sessions across a month rather than one Saturday, with a real attempt between each.

The practice literature is thinner than the slogan built on it. The ten thousand hours idea has a 2014 meta-analysis under it that is rarely quoted alongside: deliberate practice accounted for about 26 percent of the variance in performance in games, 21 in music, 18 in sports, 4 in education and under 1 percent in professions. That last figure is the one that concerns anybody reading this. It is correlational, built from self-reported practice logs and definitions that vary between studies, and Ericsson disputed it in print. I would not read it as practice not mattering. I would read it as the hours not being the interesting variable.

A good course buys vocabulary and a map. You learn what the parts are called, roughly how they fit and which questions are the standard ones, which is genuinely valuable because it is the layer that makes documentation readable. Reading a specification with no map is slow and demoralizing. The mistake is not buying the map. It is stopping there, which is easy, because the map is the part with a certificate at the end.

A conference buys a sense of what other people are worried about, which documentation cannot give you at any price, and it decays inside a month unless somebody comes back and acts on it. A certification buys a hiring filter, a shared floor of terminology and, honestly, evidence to a customer or an auditor that the floor exists. That is a real thing to want, it is not the same as being able to do the work, and everybody in the transaction knows it.

Organizations buy the map and not the rest because a course has a price, a vendor, an invoice and a completion record, and the alternative has none of those. Nobody has been questioned for approving training. What works is a few engineers spending real weeks building something awkward that a real user will complain about, which on a spreadsheet is indistinguishable from a few engineers being slow.

Build the smallest real thing, where real means somebody complains. Not the tutorial project, which has had every hard question answered in advance by whoever wrote it. Something with a real constraint and a consumer who will notice when it is wrong. The constraint is the whole mechanism: it generates the questions a tutorial answers before you can ask them, and it makes you produce with nothing in front of you, which is what everything on this list has in common.

Read the primary source, which is boring for twenty minutes and then permanently faster. The specification, the RFC, the release notes, the actual documentation, rather than the article about them. Nobody does it because the first twenty minutes are unrewarding next to a blog post, and the reason to do it anyway is that a blog post is one person's compression of the source, with their emphasis and their mistakes, so after two of them you are learning a summary of a summary. It is the highest-leverage habit here and the one I most often fail to hold.

Read the code you already depend on. The library in your own dependency tree is the ideal teacher and it is free. Real code, a real problem, already in scope, and a running system to check your understanding against. Pick the thing you have used for two years and never opened. An hour in there teaches more about how a class of problem is really solved than a week of material written to be taught.

Explain it to somebody who will interrupt. The gap between understanding something and being able to explain it is where the holes are, and they are hard to find alone, because your internal account skips exactly the parts you skipped. Someone else's question does not. Which is why explaining is worth more to the explainer than to the audience, an awkward thing to admit after asking a colleague for thirty minutes.

Work at the edge, which is unpleasant, which is why it gets renamed. Doing again the part you can already do is comfortable and produces nothing. The useful zone is where you are visibly clumsy and the feedback is quick, and it is unpleasant to sit in. Most of what people call learning is a way of not being there: watching an expert, reading around it, planning the approach. All three feel like work and none is the part that costs.

A model is extraordinary at the first hour. It names the thing you do not know the name of, orients you in an unfamiliar codebase, produces a worked example in the shape of your actual problem, and clears the stall that would have eaten an afternoon. That is not a small thing and I use it constantly. In the terms of this piece it has demolished the cost of the recognition half, which used to be paid in hours of reading.

The difficulty is that it substitutes for precisely the step that does the work. The moment you would have had to produce the answer cold, from a model of the system in your own head, is exactly the moment it is easiest to ask. Each instance is correct: the task gets done, the code is fine, nobody reviews it differently. The aggregate is that the first hour keeps happening and the tenth never does, and nothing announces it, because the output looks the same either way.

I do not have a tidy answer. What I do is narrower: get oriented, then close it before the part I am trying to learn, and sometimes ask it to test me rather than tell me, which is retrieval practice with a cheaper examiner. This is not a discipline problem and I would rather not frame it as one. The tool sits on the recognition side of the line by construction, so the retrieval has to be arranged. How that plays out over years I do not know.

It is scheduled, or it does not happen. Learning that lives in the space left over after delivery does not exist, because there is no such space in any quarter I have run. Time set aside that is first to go when things get busy was never a policy, it was a preference. The honest version is a smaller commitment that survives a bad month: half a day, on the plan, defended when somebody comes for it, which is the only hard part.

Price pairing and rotation honestly. Both work, both belong to the older argument about how work gets allocated, and neither is free. Pairing costs about the hours it appears to cost and buys transfer that does not survive in a document. Rotation costs delivery speed, and pays only if it lasts long enough to be uncomfortable, because a two-week visit teaches vocabulary and nothing under it.

Make the first clumsy attempt cheap. The fastest way to stop a team learning anything is to make being visibly bad at something expensive, and it takes no policy to do it. One review where somebody is corrected sharply in front of others is enough for everyone watching to conclude they should take on only work they can already do. Whatever a manager says about growth after that is theater.

Hiring the skill is fast and teaches the organization nothing. Bringing in somebody who already has it is the right answer more often than people who enjoy growing teams admit, particularly under a deadline. It is a purchase rather than an investment: the capability arrives whole, sits in one person and leaves when they do, unless something deliberate happens meanwhile. Growing it is slow, unreliable and compounds. The mistake is picking the fast one every time and being surprised that nothing accumulated.

Not every gap is worth closing, and most are not. Depth costs weeks that come out of something else, and a manager with a shallow layer of everything is worse at the job than one who knows exactly what they do not know and asks well. I have reached a middling understanding of things I then never touched, and the honest accounting is that the reading was recreation under a professional-sounding name.

The test I use has two parts and neither is about interest. Will I have to approve a decision that is expensive to reverse, where this is the thing I would be wrong about. And will I have to weigh somebody else's claim about it repeatedly, where being unable to tell a hard problem from an unfamiliar one leaves me deferring to whoever sounds most certain. If neither holds, the map is enough, and buying the map is correct rather than lazy.

Which means this piece applies to two or three subjects at a time, not to a reading list. I would rather close one gap properly a year and leave twenty open, knowing which twenty, than keep a low-grade study habit running across all of them and produce none of it under pressure.

I want to be careful about how much of this is settled. The mechanism I lean on is measured in laboratories on materials nothing like the work, the practice literature is contested by the people who built it, and my own evidence is a few gaps I closed and rather more I told myself I had. What I trust is the negative half: I have never once been surprised to find that watching something did not stick.

So the test I keep is a small physical one. Shut the tab, close the book, turn the notes face down, and try to produce the thing. The discomfort of the four seconds that follow is the whole difference between having read about it and knowing it, and nearly everything a person can buy is designed, quite reasonably, to spare them those four seconds.