The Name Is the Analogy

Two ways to know a thing

Aristotle called it the archê: the beginning, the first point from which a thing is, or comes to be, or is known. A first principle is the floor you hit when you keep asking why and there is nothing underneath. Everything above that floor is built. Most of what any of us knows, we did not build. We borrowed it.

Borrowing has a name. Musk gave the clean version in a 2013 TED interview: you can boil a problem down to its fundamental truths and reason up from there, or you can reason by analogy, which he described as doing what others do with slight variations. He was not sneering at analogy. He said most of life has to run on it. Nobody derives breakfast from physics.

The battery is the famous case. Ask what a battery pack costs and analogy answers: what battery packs have always cost. Ask what a battery is and the answer is a list. Cobalt, nickel, aluminum, carbon, some polymer, a steel can. Price the list and the analogy turns out to have been carrying a long story about how batteries had always been made.

The rocket teaches the deeper lesson. Add up what a rocket is made of and the materials come to a small fraction of the price. The expensive thing was not in the parts at all. It was an assumption about use: you fly a rocket once and throw it away. Question that, land it, fly it again, and the cost collapses. Sometimes the custom hides in the materials. Sometimes it hides in how the thing is used.

A good recent explainer from the First Principles channel walks this road from Aristotle through Socrates and Descartes, and gives the classic recipe in four steps: name the problem, ask of every assumption how you actually know it is true, keep peeling until you reach bedrock, then rebuild from that alone. Its most useful habit is the plainest one. Separate facts from conventions, because we confuse the two constantly. That recipe was written for people reasoning on their own. When the one doing the reasoning is an AI, two of its steps need to change, and this lesson is about which two and why.

Analogy is not a mistake. It is a loan. You get the answer cheaply, and you inherit every assumption the lender made.

Why your AI defaults to analogy

A language model was built by reading an enormous share of what people have written and learning, at each step, what usually comes next. That is not a flaw in the design. It is the design. A model’s first principle is precedent. Its default answer is the nearest case, adjusted: reasoning by analogy, at a scale no person has ever reached.

The lean has been measured. Researchers at MIT and Boston University gave models familiar tasks and then counterfactual twins of them: the same skill with one default assumption changed, such as doing addition in a base other than ten. The models stayed above chance on the twins, and their performance dropped substantially and consistently. They were partly reasoning and partly reciting the familiar version.

So the obvious fix, typing “think from first principles,” does not work the way you hope. First principles is also a phrase with precedent. The model has read thousands of essays that use it. Ask for it by name and the reasoning gets better, sometimes genuinely better, but the answer still lands where the noun points. In Part 2, a question about pricing with “think from first principles” in front reasoned carefully for six paragraphs and then arrived at the same three-step ladder as the plain question. First-principles in style. Analogy in destination.

The analogy does not live in your instructions. It lives in the noun. Say pricing and you summon every pricing page. Say restaurant website and you summon the menu PDF and the stock photo of a plated scallop. The category word is a pointer to the nearest precedent, and the model follows pointers faithfully.

This is the first step of the classic recipe that has to change. For a person, naming the problem is how you aim. For a model, the name is the aim, and it points backward, at everything already done under that name.

Name the outcome. Leave the thing unnamed.

The other half of the method has evidence behind it too. Google DeepMind’s step-back prompting asks a model to name the principles that govern a question before it answers the question itself, and accuracy rose on hard physics, chemistry and multi-step questions. Abstraction first, the specific case second. That is Aristotle’s order, run on a machine.

The outcome is the only part you have to write

Once the name is gone, something has to take its place, and this is where most of the skill lives. A first-principles method can derive almost everything on its own. Given a target, it can work out the gap, the cost, the time, and how you would check. The one thing it cannot derive is the target itself. The outcome is the only free variable. Everything the method builds is built from it.

That means the answer can only be as specific as the outcome you give it. Give it a single word like success and a good method will not collapse; it will take the word apart, explain what success is made of, and ask you for a target. True, and not yet useful. Give it a sentence with checkable clauses and every clause becomes a design decision.

Here is a model outcome. “A local owner hands me a job they’ve been avoiding, gets it back finished within a week, pays without negotiating, and it takes me under six hours.” It has a person, a checkable result, two limits and an ending. It never says service, consulting, agency, or AI. In Part 2 I run the prompt below with this sentence, with the word success, and with an ordinary named question, so you can see the difference for yourself.

The method: five moves

The method keeps the old instinct, separate what is true from what is merely usual, and changes the two places where a machine differs from a person: how you start, and what you do with the analogy at the end.

1. Write the outcome, not the name

   Describe what has to happen without the category word, using the four parts above. If the noun sneaks back in, the precedent comes with it.

2. Ask for parts, not examples

   What must exist, happen, or be paid for for this outcome to occur? Materials, steps, people, minutes, dollars, rules. Hold off on “how others do it.” The battery became a list before it became cheap.

3. Sort law from custom

   Have it label every part: required by physics, money, time, regulation, or how human attention works, or merely how it is usually done. Include how the thing is used, not only what it is made of; the rocket’s costliest custom was flying it once. Custom is not wrong. It is optional, and you need to see which is which.

4. Rebuild from the law alone

   Ask for the simplest arrangement that satisfies only the parts marked law. It will feel naked. That is the point. It is the first version of the answer that is yours.

5. Then let the analogy back in

   Now ask how this is usually done, and make it explain every difference between the usual way and the rebuilt one. Each gap is a custom worth keeping, a custom worth dropping, or a law you missed. This is the second step that changes: the classic advice is to ignore how it has always been done. Don’t. Analogy is a poor architect and an excellent auditor.

I am going to describe a problem without naming what kind of thing it is.
Do not guess the category. Do not tell me how others usually do it yet.

1. List the irreducible parts: everything that must exist, happen,
   or be paid for for this outcome to occur.
2. Mark each part LAW (physics, money, time, regulation, human
   attention) or CUSTOM (it is usually done this way). Include
   assumptions about how the thing is used, not just what it is.
3. Build the simplest arrangement that satisfies only the LAW parts.
4. Only now: describe how this is normally done, and explain every
   difference between the normal way and yours.

What has to happen: [a person, a result you could check, the limits,
and how it ends. No names.]

What I expect

1. Sameness becomes visible

   Pricing pages, menus, proposals and websites drafted with AI will converge, because every assistant starts from the same nearest case. You will learn to spot the default by sight, the way you can spot a template.

2. The outcome sentence becomes the core skill

   The category word becomes the most expensive word in a prompt, and the outcome sentence the most valuable. Skilled users will leave the noun out and put their effort into one checkable sentence. They will get answers no competitor’s assistant would have produced, from the same model.

3. Sorting law from custom becomes the human job

   A model lists parts faster than any person. Knowing which part is a real constraint and which is a habit wearing a badge is local, earned knowledge: the contractor who knows the “rule” is a manufacturer’s preference, the owner who knows which customers never read the fine print.

4. Assistants will mark custom by default

   Answers will increasingly separate what is required from what is merely conventional, because the people paying for them will ask, and because that separation is the part they cannot get anywhere else.

5. Analogy keeps its job, as the auditor

   The best practice will be parts first, precedent second. The order flips. The tool stays.

Where I live in this

The nearest analogy for an AI business is a software company or a training company. Charleston AI is neither. We deliver finished work to hospitality businesses and local owners: the thing done, not the tool to do it or the class on how. Take the words “AI company” off and ask an owner what has to happen, and the parts list is short. The work, finished, in their hands. Most of the rest of the usual list is custom.

Prediction leans toward the familiar continuation, which I wrote about in Most Likely, Least Likely. That is a feature when you want something dependable and a trap when you want something that is yours. And when everyone leans on the same few models, the mistakes stop being independent, which is the argument of The Safety Will Be Spent. Leaving the thing unnamed is the cheapest defense I know against both.

Already true

Language models lean on familiar defaults: change one assumption in a familiar task and performance drops. Asking a model to name the governing principles before it answers improves accuracy on hard questions. And in the worked example in Part 2, a named question returned the familiar three tiers, the same question with “think from first principles” returned the same shape by a better argument, and the same model, given a checkable outcome and no names, built a design in which every choice traces to a clause.

What has to happen

People have to notice that the noun carries the precedent, build the habit of asking for parts before examples, and write the outcome as a test rather than a wish. Habits form slowly. Most prompting advice is read once and forgotten.

Where I am probably wrong

The models may learn to do this unprompted. Reasoning models already step back on their own some of the time, and if a plain question starts returning the parts-first answer by default, the prompt becomes unnecessary; the outcome sentence will not, because no model can write it for you. And the law column can lie too: plenty of what passes for a rule is a custom with a badge. The method moves the error. It does not abolish it.

Ask again

Back to Tuesday, 11:40. Ask again, and do not say pricing. Say what has to happen: forty owners a year want this, each one takes me six hours, they decide on one call, and I would rather turn ten away than discount. Ask what has to be true. Ask what is merely usual. Then ask what everyone else does, and why.

It will not hand you good, better, best. It will hand you something with your fingerprints on it.

The parts were always there. The name was in the way.