Arguing from First Principles
Reduce the problem to what is irreducibly true, then rebuild without borrowing anyone else's conclusions.
The Model
Almost all reasoning is analogy. This resembles that, that worked, so do this.
Analogy is fast, and the speed is real. It is also how you inherit the assumptions of whoever went first, at full price, without ever seeing the invoice.
A first principle is a fact that cannot be derived from anything more basic. Aristotle's word for it was archē — the starting point from which a thing is known. In practice it is a physical constraint, a contractual obligation, a measured cost, or a piece of arithmetic.
Everything above that layer is convention. Convention is not worthless; most of it encodes hard-won experience. But convention is the only part of a problem that can be changed by a decision, so it is worth knowing which part it is.
The method is one question, repeated until it stops producing new information. How do I know that?
The method is slow. Spend it where the inherited answer is expensive.
Why operators get this wrong
Two failures, and they pull in opposite directions.
The first is using the phrase as permission to ignore people who know things. A founder reasons from scratch about employment law, or a medical claim, or a customs code, and discovers that the constraint was real, the research was a forty-minute phone call, and the penalty is retroactive.
First principles means finding the irreducible facts. It does not mean the facts are all in your head.
The second failure is stopping one layer too early and calling it bedrock. "Our acquisition cost is too high" feels fundamental and is not — it is a ratio of four other numbers, at least two of which you control directly.
Pricing is where this costs the most money. A founder sets price at the competitor minus ten percent. That is pure analogy, and it silently accepts the competitor's cost structure, their segment, and their willingness to be cheap.
The first-principles question is different: what does the buyer's current alternative cost them per month, in money and in hours, and what fraction of that are you removing? That number sets the ceiling. It usually sits three times higher than the analogy price.
Applied
Body
You do not need a programme. You need the four facts a programme is made of.
Progressive overload, enough protein, enough sleep to adapt, and enough recovery to repeat the session next week. That is the whole substrate. Every split, brand, fasting window and supplement is decoration applied on top.
Here is what the analogy trap looks like in a gym. Someone spends six months optimising creatine timing, pre-workout dosing and the order of their accessory work, while sleeping five hours and forty minutes a night.
The variable that dominates the entire system is untouched, because it cannot be bought and does not feel like an intervention.
Run the reduction honestly and the plan gets shorter, cheaper and duller. That is the signature of having reached the bottom layer.
Business
Take the number you want to change and decompose it until every term is something a human can act on.
Revenue is not a lever. Traffic times conversion times price times repeat rate is four levers, and one of them is always further from its physical limit than the others.
A concrete case. An agency owner at four hundred thousand wants to double, and assumes the answer is headcount, because that is what every agency he knows did.
Decompose it: revenue equals billable hours times rate times utilisation. Hours have a hard ceiling set by human beings and the calendar. Utilisation is already at eighty percent and the last twenty is where the business breaks.
Rate has no physical ceiling at all. It is bounded only by what the work is worth to the buyer, which is a positioning question, not a hiring question. He was about to solve a pricing problem by recruiting eleven people.
AI Leverage
A language model is an analogy engine. It is trained, quite literally, to produce the most typical continuation of what came before.
Which means asking it what your onboarding should look like returns the average of every onboarding it has read. That answer is often fine, and it is never differentiated, because average is the mechanism.
Use it against its grain. Do not ask for the answer — ask it to enumerate the assumptions your plan is standing on, then sort them into three buckets: physical constraints, contractual obligations, and conventions.
The conventions column is where your opportunity is, and it is the one column you will never produce by yourself, because conventions are invisible from inside the industry that holds them.
A second use, unglamorous and high-return: hand it your pricing page, your contract or your unit economics and ask it to restate every claim as an arithmetic identity. Half the time a term you have been treating as fixed turns out to be an average of two things moving in opposite directions.
The Drill
Find one decision this week where you are copying somebody. It will not be hard. Most of an operating plan is inherited.
Write the claim down in a single sentence. Then ask how do I know that, in writing, five times, answering each one before you ask the next.
Stop when you hit something you could verify with a measurement, a contract, or a law of physics. That is your floor.
Now check whether the original decision still survives contact with the floor. Roughly half the time it does, and you have bought conviction cheaply. The other half you have found the assumption that was quietly costing you the quarter.
Do this on one decision, not five. The method is expensive by design and it is wasted on choices that do not matter.
Stoic parallel
Marcus practised something he never named but used constantly: taking a thing apart and describing it in the plainest possible terms, so that the word could no longer do the thinking for him.
Roasted meat is a dead animal. Fine wine is fermented grape juice. Imperial purple is wool dyed with shellfish blood.
The exercise was not cynicism. It was the removal of a borrowed judgment that had attached itself to the object, so he could see what was actually there and decide about that instead.
Epictetus taught the same discipline earlier in the chain, at the level of impressions. Something happens, the mind supplies a verdict instantly, and the verdict feels like part of the event.
It is not. It was added.
Separating the two is the oldest first-principles exercise on record, and it is still the one that pays.
One model per week.
Applied to training, business, and AI leverage. No fluff.
Related models
- InversionStop asking how to succeed. Ask what would guarantee failure, then refuse to do it.
- Circle of CompetenceThe size of the circle does not matter. Knowing exactly where its edge runs is the whole discipline.
- Product/Market FitA market that wants the product pulls it out of you. Everything before that is pushing.
Origin: Aristotle, whose Posterior Analytics defines the first principle as what cannot be derived from anything prior; revived as a working method by Descartes