What a good explanation is
Sit down with someone who says they understand what an API is. Ask them one plain question: why does a backend exist at all? Why not run everything in the browser?
They start strong. Then, somewhere in the second or third sentence, the explanation stops being theirs. It turns into a line they remember from a lecture. You can hear the exact moment it happens. The words keep coming, but the thinking has stopped.
I have watched that moment in seven cohorts and about 2,500 students, and I have caught myself in it more times than I would like. The difference between the two states, the sentence that is yours and the sentence you are repeating, is what David Deutsch’s The Beginning of Infinity is about. It is a long book with a lot in it. This is the part I teach.
Knowledge is a guess that survived
The story most of us were taught is that knowledge comes from looking. Observe carefully, collect enough examples, and the pattern reveals itself.
Deutsch, following Karl Popper, says that is not how it works, and never was. Observation on its own produces nothing. You have to guess first. Somebody proposes an explanation, and then everybody tries to break it. What survives is knowledge, for now. What does not is discarded. There is no other source. Not authority, not data on its own, not feeling certain.
I did not learn this from a philosophy book. I learned it from watching builders. Every product document my students write is a stack of guesses: who the user is, what their problem is, what done looks like. I make them mark every line as an assumption, because that is what it is until something outside the room confirms it. Then they go and test. A belief, exposed to reality, corrected, exposed again. That loop is the whole method, and it is also, as it happens, the loop a neural network runs when it adjusts its weights.
The builder who skips the loop does not look careless. He looks competent. He ships fast and polishes the screens and never once talks to the person he is building for. The trap is shaped exactly like competence. That is why it works.
A good explanation is hard to vary
So if everything is a guess, what makes one guess better than another?
Deutsch’s answer is the most useful sentence in the book. A good explanation is hard to vary while still explaining what it set out to explain.
His example is the seasons. The ancient Greeks explained winter as the grief of Demeter, whose daughter was taken to the underworld for part of each year. The explanation fits. It also fits anything else. If winter had come twice a year, the myth would have been adjusted. If the southern hemisphere had been discovered to have summer during Greek winter, the goddess would have been given a second grievance. Nothing in the story does any work, so the story can be bent to any shape, so it explains nothing.
The axial tilt of the Earth is different. Every part of it is load-bearing. It says the hemispheres must have opposite seasons at the same time, and that regions near the equator should barely have seasons at all. Change one detail and the whole account breaks. That is what makes it an explanation rather than a story.
You already know the bad kind. A horoscope. “You sometimes hold back from putting your work out there.” True of every human who has ever lived, and it names nothing. Most explanations in business are horoscopes: the market was not ready, the team lacked focus, the culture was wrong. They can be fitted to any outcome after the fact, which is exactly why they teach nothing. When my first company failed I had explanations like that for a long time. They felt like understanding. They were a way of not looking.
The test I use now, on every claim including mine: what does this forbid? If an explanation does not rule anything out, it is not an explanation yet. A claim specific enough to be wrong is the only kind that can be right.
Names are not explanations
Richard Feynman’s father taught him that you could know the name of a bird in every language on Earth and know nothing about the bird. You would only know what humans call it.
This is the same rule from a different side. A name is the most easily varied thing there is. I can say “retrieval-augmented generation” or “event-driven architecture” or “product-market fit”, and nothing I have said can be wrong, because nothing I have said does any work. The explanation is the thing underneath: why the model hallucinates, why the pieces must be separated, what the customer actually did last week.
So in my classes nobody is allowed to say “database” until they have felt the problem a database solves. We start with a program that forgets everything when the tab closes, and we work out what would have to exist for it to remember. By the time the word arrives, they have already invented most of the thing. A framework you derived is a framework you can trust under pressure. A framework you memorised is a slide you once saw.
Which is also why “I understand it” is the least reliable sentence a person can say. Recognition feels like understanding from the inside. The only test is to put the name away and rebuild the thing from the problem. Where you stumble is where the explanation stopped being yours.
Good explanations reach
Here is something I noticed only after years of teaching it. A good explanation always ends up explaining more than it was made for.
I use one lens for the whole technology stack: every hard problem is an interface, a contract between two systems that do not share a language. A user and a machine. A browser and a server. A program and a database. A probabilistic model and a deterministic tool. I did not design the lens to cover all of those. It was made for one problem and it kept turning out to cover the next one, uninvited.
Deutsch calls this reach, and he argues it is the signature of a real explanation. Rules of thumb do not reach. “Do what worked last time” is true exactly as far as last time and no further. A recipe tells you what to do; it does not tell you why it works or when it will fail, and so it cannot be improved by criticism, only by more trial and error. An explanation can be argued with. That is its advantage.
If you want to know whether you understand something or merely have a procedure for it, ask whether it has reached anywhere yet.
Why this matters more now
For most of history a fluent explanation was expensive to produce, so fluency was a decent proxy for understanding. That proxy is gone.
Ask a model why your product failed and it will give you a structured, confident, well-organised answer in four seconds. Ask it which track to choose in my course and it will produce a full decision framework with assessment questions and a checklist, recommending tracks that do not exist. The right template and the wrong data. That is what a hallucination is, and it is also what a bad explanation has always been: the shape of understanding with nothing inside that can be checked.
The model has made recognition free. Any term, defined instantly, forever. Whatever was worth anything about knowing the word is now worth nothing. What remains scarce is the thing that was always the real work: knowing what would prove the answer wrong, and being able to rebuild it from the problem when nobody hands you the words.
You are going to spend the rest of your working life judging explanations you did not produce. You can only judge what you could have derived. I have argued elsewhere that this is why the struggle to understand is not optional, and the argument is the same one Deutsch makes: there is no shortcut from a guess to knowledge that goes around criticism.
Problems are soluble
The part of Deutsch that people find hardest to accept is the optimism. He holds that problems are inevitable, and that problems are soluble. Not some of them. Any problem that is not forbidden by the laws of physics can be solved, given the right knowledge, and knowledge can be created.
This is not cheerfulness. It follows from everything above. If knowledge comes from guessing and correcting rather than from some fixed stock of truth handed down, then there is no limit built into the process. A problem you cannot solve today is a missing explanation, not a wall. The failure of the first company was not a verdict on me. It was a set of explanations I had not found yet.
Most of the people I teach arrive believing the opposite about themselves. “I am not technical.” “I need more clarity before I can start.” Those are explanations too, and notice how easy they are to vary. They fit every outcome and forbid nothing. The replacement is not confidence. It is a smaller, harder question: what is the smallest test I can run, and what would it prove wrong?
The whole thing in three questions
When someone hands you an explanation, or when you catch yourself about to offer one:
- What does it forbid? If the answer is nothing, it is a story, not an explanation.
- Could I rebuild it from the problem, without the name? If not, you are repeating, not understanding.
- Has it reached anywhere? If it explains only the thing it was made for, it is a recipe. Useful, but not yet knowledge.
Deutsch’s book runs to nearly five hundred pages and goes places I have not followed him, into the multiverse and the objectivity of beauty. You do not need any of that to use the core. Guess, expose the guess, keep what survives. Prefer the account you cannot bend. Trust the explanation that reaches.
Awareness is free now. Explanation is the work.