First principles thinking
How to take a problem apart until you reach things that are actually true (physics, arithmetic, hard constraints), throw away what is merely customary, and rebuild a solution from those pieces. Covers where the idea comes from, when it beats reasoning by analogy and when it does not, a step-by-step method, the supporting tools (Socratic questions, Five Whys, Fermi estimates, unit economics, the idiot index, inversion, constraints), worked examples and the ways it goes wrong. It pairs with systems thinking, which puts the parts back together, and with the wider toolkit in general thinking models.
Definition and origin
A first principle is a proposition you cannot usefully derive from anything more basic within the problem at hand: a law of nature, a definition, an arithmetic identity, a measured fact. First principles thinking is decomposing a problem down to those propositions and reasoning upward from them, instead of starting from how things are usually done.
| Source | Idea | What it contributes |
|---|---|---|
| Aristotle, Metaphysics Book V (Δ), 1013a | a "beginning" (archē) is the first point from which a thing is, comes to be, or is known | the name: a starting point of knowledge that is not itself derived |
| Aristotle, Posterior Analytics I.2 | demonstrated knowledge rests on premises that are true, primary and immediate, better known than and prior to the conclusion (paraphrased) | reasoning is only as good as the premises it rests on |
| Descartes, Discourse on the Method (1637) | four rules: accept only what is clearly known, divide difficulties into parts, go from simple to complex, enumerate completely | the procedure: doubt, decompose, rebuild, check |
| Descartes, Meditations (1641) | method of doubt: suspend every belief that can be doubted to find what survives | the attitude: treat inherited beliefs as unproven |
| Physics | derive behavior from conservation laws, measured constants and units | quantification: fundamentals come with numbers |
| Musk (Wired, 2012; Kevin Rose interview, 2012) | "reason from first principles rather than by analogy" | the popular business framing, applied to cost |
Aristotle's wording in W. D. Ross's translation: "It is common, then, to all beginnings to be the first point from which a thing either is or comes to be or is known." The line often quoted online, "the first basis from which a thing is known", is a compressed rendering of the same passage (1013a), not Ross's text.
Descartes' first two rules, in John Veitch's translation of the Discourse: "never to accept anything for true which I did not clearly know to be such" and "to divide each of the difficulties under examination into as many parts as possible, and as might be necessary for its adequate solution." The third rule (ascend from the simplest objects to the complex) is the rebuild step; the fourth (enumerations "so complete" that nothing is omitted) is the check.
Musk's version, from his October 2012 interview with Chris Anderson in Wired: "I tend to approach things from a physics framework. And physics teaches you to reason from first principles rather than by analogy." In the Kevin Rose interview the same year he described it as a way to "boil things down to the most fundamental truths and then reason up from there".
First principles vs reasoning by analogy
Reasoning by analogy means copying what worked in a similar case, with small changes: "competitors charge $29, so we will", "every startup raises a seed round", "rockets cost what rockets have always cost". It is how most decisions are made, and usually rightly so.
| Reasoning by analogy | Reasoning from first principles | |
|---|---|---|
| Starts from | what others did in similar cases | what must be true (laws, constraints, measured facts) |
| Cost | cheap: minutes, borrows others' learning | expensive: hours to months, needs domain knowledge and data |
| Typical output | incremental improvement on the status quo | a different design, or confirmation that the status quo is near optimal |
| Error mode | inherits the reference class's mistakes and outdated assumptions | misses tacit knowledge, reinvents wheels, overconfident models |
| Best when | the domain is stable, the stakes are low, many others have iterated on it | conventions are wrong, the domain or its inputs are changing, stakes are high |
| Examples of good use | choosing a CRM, standard legal docs, office layout | cost-down of a core component, new product category, pricing a new kind of service |
| Signal that first principles will pay off | Why |
|---|---|
| the price or cost is far above the cost of inputs (high idiot index) | the gap is made of conventions and process, which can change |
| an input changed recently (cost of compute, a new material, a regulation) | industry habits were optimized for the old input |
| "that's how it's done" is the only justification anyone can give | no one has checked the reasoning recently |
| experts disagree on basic numbers | the reference class is not trustworthy |
| the decision is large and irreversible | the cost of the analysis is small relative to the stakes |
| incumbents are all doing the same thing for the same reason | shared assumptions are shared blind spots |
The practical rule: default to analogy, escalate to first principles for the few decisions where the gap between convention and physics looks large and the payoff from closing it matters.
The method
State the goal as an outcome, not a solution
"Get 5 people from A to B daily for under $X" rather than "buy a van". "Store energy at under $Y/kWh" rather than "buy cheaper battery packs". A solution-shaped goal smuggles in assumptions before you start.
List every assumption
Write down what you, the industry and your team believe about the problem: costs, constraints, customer behavior, required features, regulations, timelines. Include the ones that feel too obvious to write.
Classify each assumption
Sort into law of nature, hard constraint, convention, or opinion (table below). Most of the value is found here: conventions masquerading as laws.
Decompose to fundamentals
Break the cost, time or performance into its components: materials, energy, labor hours, machine time, information flows, approvals. Keep going until each piece is something you can measure or look up.
Quantify
Put a number and a unit on each component, with a range. Fermi estimates are fine at this stage; you need orders of magnitude, not precision. Check units add up.
Rebuild from the fundamentals
Ask: given only the laws and hard constraints, what is the cheapest, fastest or simplest design that meets the goal? Then add back only the conventions that earn their place.
Test against reality
Build the smallest experiment that could prove the rebuilt design wrong: a prototype, a quote from a supplier, a price test, a pilot. Update the model with what breaks. Theory that has not met a supplier or a customer is not done.
Classifying assumptions
| Class | Test | Can you change it? | Examples |
|---|---|---|---|
| Law of nature | would violating it break physics, math or logic? | no | energy conservation, speed of light, a lithium-ion cell's theoretical energy density limit, compounding arithmetic |
| Hard constraint | is it enforced by something outside your control, at real cost if breached? | rarely, slowly, at high cost | law and regulation, contracts in force, available capital, human biology, current supplier capacity |
| Convention | did people choose it, and could a competitor do it differently tomorrow? | yes | pricing norms, org structure, batch sizes, approval steps, standard part choices, "enterprise needs a sales team" |
| Opinion | is it a prediction or preference without data behind it? | yes, by testing | "customers won't pay for that", "this market is too small", "users hate onboarding flows" |
Two useful probes: "Who decided this, and when?" (conventions have authors and dates; laws do not) and "What would have to be true for this to be false?" (if the answer is "physics would have to change", it is a law).
Tools
Socratic questioning
Disciplined questioning to expose assumptions and test reasoning, named after the method Plato shows Socrates using. The six categories below are the standard classification taught in critical-thinking courses (see the Wikipedia entry in References).
| Category | Purpose | Questions to ask |
|---|---|---|
| Clarification | make the claim precise | What exactly do you mean by "expensive"? Compared with what? Can you give an example? |
| Probing assumptions | surface what is taken for granted | What are we assuming here? Is that always true? Why do we think it holds in this case? |
| Probing reasons and evidence | test the support | How do we know? What is the source? What would change our mind? Is there reason to doubt this evidence? |
| Viewpoints and perspectives | find alternatives | How would a competitor, a customer, a regulator see this? What is the strongest counter-argument? |
| Implications and consequences | follow it through | If this is true, what else must be true? What happens next? Who is affected? |
| Questions about the question | check you are solving the right problem | Why does this question matter? Is there a better question? What would answering it let us do? |
Five Whys
Ask "why?" repeatedly until you reach a cause you can act on structurally. Taiichi Ohno, architect of the Toyota Production System, described it as "the basis of Toyota's scientific approach by repeating why five times the nature of the problem as well as its solution becomes clear" (Toyota Production System, English edition 1988).
Ohno's own example, condensed:
| Why? | Answer |
|---|---|
| 1. Why did the machine stop? | there was an overload and the fuse blew |
| 2. Why was there an overload? | the bearing was not sufficiently lubricated |
| 3. Why was it not lubricated? | the lubrication pump was not pumping sufficiently |
| 4. Why was it not pumping sufficiently? | the pump shaft was worn and rattling |
| 5. Why was the shaft worn? | there was no strainer, so metal scrap got in |
Replacing the fuse fixes the symptom; fitting a strainer stops the recurrence.
| Limit | What to do instead or as well |
|---|---|
| "Five" is arbitrary; the root may be at 3 or 9 | stop when you reach a cause you can change structurally, not at a count |
| produces one causal chain; real failures usually have several contributing causes | branch the tree at each level; use a fishbone (Ishikawa) diagram or a causal loop diagram |
| investigators stop at symptoms or at "human error" | ban "someone made a mistake" as an answer; ask why the system allowed the mistake |
| results depend heavily on who is asking and what they already believe | do it as a group with people from each part of the process; check each link with data |
| critics: Teruyuki Minoura (ex-Toyota) called it too shallow in practice; Alan Card (BMJ Quality & Safety, 2017) argued it should be abandoned for serious incident analysis | treat it as a quick triage tool, not a method for complex incidents |
Fermi estimation
Estimate an unknown quantity by multiplying rough, independently estimated factors. Named after Enrico Fermi, who was known for good order-of-magnitude answers from little data. Errors in the factors tend partly to cancel, so the product is often within a factor of a few of the truth, provided no single factor is badly biased.
Worked example 1: piano tuners in Chicago (the classic)
| Factor | Estimate | Reasoning |
|---|---|---|
| population | 2.7 million | 2020 US census, rounded |
| people per household | ~2.5 | typical US figure |
| households | ~1.1 million | 2.7 M ÷ 2.5 |
| share of households with a piano | ~1 in 20 | guess; the weakest factor |
| household pianos | ~55,000 | plus schools, churches, venues: call it ~60,000 |
| tunings per piano per year | ~1 | recommended once a year; many are tuned less often |
| tunings a tuner can do per year | ~1,000 | ~2 hours each with travel → 4 a day × 250 working days |
| tuners | ~60 | 60,000 ÷ 1,000 |
The answer is "tens, not thousands". If a directory showed thousands, the estimate tells you which factor to question.
Worked example 2: how many servers for 1 million daily active users?
| Factor | Estimate |
|---|---|
| requests per user per day | 50 (assumption; measure it) |
| requests per day | 50 million |
| average requests per second | 50 M ÷ 86,400 s ≈ 580 |
| peak-to-average ratio | 3× (assumption) → ~1,750 req/s |
| throughput per instance | 500 req/s (from a load test) |
| instances at peak | 1,750 ÷ 500 = 3.5 → 4, plus 50% headroom → 6 |
Dimensional analysis and unit economics
Dimensional analysis: every term in an equation must have the same units, and units multiply and cancel like algebra. It catches most spreadsheet errors ("users × $/month = $/month", not "$").
Unit economics is first principles applied to a business: decompose the P&L into what one unit (a customer, an order, a ride) earns and costs.
| Quantity | Formula | Example |
|---|---|---|
| gross margin per customer per month | $50 × 80% = $40 | |
| expected customer lifetime (constant monthly churn c) | 1 ÷ c months | 1 ÷ 0.03 ≈ 33 months |
| lifetime value (LTV) | $40 ÷ 0.03 ≈ $1,333 | |
| CAC payback | $400 ÷ $40 = 10 months | |
| LTV : CAC | LTV ÷ CAC | $1,333 ÷ $400 ≈ 3.3 |
The same decomposition works for anything: cost per delivered parcel, cost per inference, cost per hire, cost per kilogram to orbit.
The idiot index
The ratio of what a finished part costs to what its raw materials cost. Walter Isaacson's 2023 biography Elon Musk describes Musk using it at SpaceX and Tesla: a component with a high index is a sign that the design is too complex or the manufacturing process too inefficient.
| Index (rough heuristic) | Reading |
|---|---|
| ~1–3 | commodity-like; little room except in materials |
| ~10 | typical of machined or assembled parts; look at process, tolerances, supplier margin |
| ~50+ | the cost is almost all process, overhead, low volume or margin: first principles territory |
Caveats: it ignores the value of design, qualification, testing, certification and low volume, which are real costs. A high index is a question ("where does the money go?"), not a verdict.
Inversion
Solve the problem backwards: instead of "how do we succeed?", ask "what would guarantee failure?" and avoid those things; instead of "how do we make this faster?", ask "what makes it slow?". The mathematician Carl Jacobi is said to have told students to "invert, always invert" (man muss immer umkehren); Charlie Munger popularised it as a thinking tool. In first principles work it finds the constraints you forgot to list.
| Forward question | Inverted question |
|---|---|
| How do we reduce churn? | What would make a happy customer leave next month? |
| How do we ship faster? | What stops a finished change reaching users today? |
| What should this product do? | What must it never do? |
| How can this plan succeed? | It is a year from now and the plan failed: why? (a pre-mortem) |
Constraints analysis: physics vs policy
For every limit on performance, ask whether it is physics (a law or hard constraint) or policy (a rule, habit, or decision someone made). Most bottlenecks in organizations are policy.
Eliyahu Goldratt's theory of constraints (The Goal, 1984) adds the discipline of working on one constraint at a time: the throughput of a system is limited by its tightest constraint, so improving anything else does not help.
| Step (Goldratt's five focusing steps) | In practice |
|---|---|
| 1. Identify the constraint | where does work queue up? what is at 100% utilization? |
| 2. Exploit it | get the most out of the constraint as it is: no idle time, no low-value work on it |
| 3. Subordinate everything else | pace the rest of the system to the constraint |
| 4. Elevate it | invest to increase its capacity (hire, buy, redesign) |
| 5. Repeat | the constraint has moved; do not let inertia become the new constraint |
| Constraint | Physics or policy? | First-principles response |
|---|---|---|
| "Releases happen fortnightly" | policy | ask what risk the cadence controls; control it directly |
| "A London server cannot answer a Sydney user in under ~170 ms" | physics: ~17,000 km each way at ~200,000 km/s in fiber | move the data or the compute closer, not the code |
| "Enterprise deals need 6 months" | mostly policy (buyer procurement), partly hard constraint | find which steps are legal requirements; remove or parallelize the rest |
| "We can't hire senior engineers" | usually policy (pay bands, location, process) | test each: which would a competitor change? |
Musk's "algorithm" (as reported)
Isaacson's biography reports a five-step process Musk repeated to his teams, in this order (paraphrased):
| Step | Idea |
|---|---|
| 1. Question every requirement | each requirement should come with the name of the person who made it, not a department |
| 2. Delete any part or process step you can | if you are not adding some back later, you did not delete enough |
| 3. Simplify and optimize | only after deleting; the common mistake is optimizing something that should not exist |
| 4. Accelerate cycle time | speed up only what survived steps 1–3 |
| 5. Automate | last, not first |
It is first principles thinking turned into an operating routine: steps 1–2 are "classify assumptions", step 3 is "rebuild".
Worked examples
Battery packs (Musk, 2012)
In a September 2012 episode of Kevin Rose's Foundation interview series, Musk used battery packs to illustrate the method. The common assumption he described was that packs had historically cost "$600 per kilowatt hour" and always would. His first-principles version: list the material constituents (cobalt, nickel, aluminum, carbon, polymers for separation, a steel can), price each at London Metal Exchange rates, and add them up: "It's like $80 per kilowatt hour."
| Step | Content |
|---|---|
| goal | cheap stored energy for cars, in $/kWh |
| analogy answer | ~$600/kWh, "historically", so electric cars stay expensive |
| decomposition | cell chemistry → materials by mass per kWh → commodity prices |
| fundamental floor | ~$80/kWh in materials (Musk's 2012 figure) |
| conclusion | the ~$520 gap is manufacturing, design and scale, which can be engineered down |
| what happened | BloombergNEF's December 2025 survey put the average lithium-ion pack price at $108/kWh, and battery-electric-vehicle packs at $99/kWh |
Caveats: the materials floor is not an achievable price (processing, cell manufacturing, packs, margins are real costs), commodity prices move (lithium prices spiked in 2022, for example), and the fall in pack prices came from the whole industry's scale and chemistry changes, not one company's reasoning. The method's value was showing that the high price was not a law.
Rocket materials (Wired, 2012)
Musk told Chris Anderson (Wired, October 2012) that he asked what a rocket is made of ("aerospace-grade aluminum alloys, plus some titanium, copper, and carbon fiber") and what those materials cost on the commodity market. His answer: "the materials cost of a rocket was around 2 percent of the typical price", against "probably 20 to 25 percent" for a Tesla car. An idiot index near 50 said the cost was almost all process, overhead and low volume. SpaceX's response over the following decade (vertical integration, in-house manufacturing, then reusable boosters) attacked that gap.
The Wright brothers' lift data
After disappointing glider flights in 1900–1901, the Wrights stopped trusting the published aerodynamic data they had inherited (including the long-accepted Smeaton coefficient, which turned out to be too high) and built their own wind tunnel in late 1901 to measure lift and drag on many small wing shapes. The accepted numbers were a convention that everyone had copied; the measurements were the fundamentals. Their 1902 glider, designed from their own data, worked.
Pricing a product from its cost structure (illustrative numbers)
A B2B scheduling tool. Analogy says: "competitors charge $29 per seat, so charge $25."
| Component | Monthly cost per customer | Note |
|---|---|---|
| compute and storage | $4 | from the cloud bill ÷ active customers |
| support | $6 | ~20 minutes per customer per month at a loaded $18/hour |
| payment fees | ~3% of price | card processing |
| onboarding (amortized) | $2 | $48 of setup time over 24 months |
| unit cost | ~$12 + 3% |
| Anchor | Calculation | Result |
|---|---|---|
| price floor at 80% gross margin | $12 ÷ (1 − 0.80 − 0.03) | ~$71/month |
| value ceiling | saves 5 staff hours/month × $40/hour | $200/month |
| sensible range | between floor and roughly a third to a half of value | ~$70–100/month per location |
The first-principles answer prices per location, not per seat (the value scales with appointments, not logins), at roughly 3× the analogy price. Then test it (step 7): a price test is cheaper than a debate.
Redesigning a process: the weekly change board
"Every production change needs sign-off from the Thursday change advisory board." Classified: convention.
| Question | Answer |
|---|---|
| What is the goal? | few customer-visible failures, fast recovery when they happen |
| What are the fundamentals? | failure probability per change, blast radius, time to detect, time to roll back |
| What does the board do to each? | little to failure probability; batches changes weekly, which increases blast radius and makes causes harder to find |
| Rebuilt design | small changes, automated tests, canary releases, feature flags, one-click rollback, peer review in the pull request |
| Evidence | DORA's 2019 State of DevOps report found no evidence that formal external review was associated with lower change fail rates |
Personal: do I need a car? (illustrative numbers)
Analogy says: "adults own cars." First principles says: the goal is trips, not a car.
| Fundamental | How to get it |
|---|---|
| trips per month by type (commute, shopping, weekend, emergency) | a week of logging, scaled up |
| full annual cost of ownership | depreciation + insurance + fuel or charging + maintenance + parking + tax, from your own quotes |
| cost of the alternative mix | public transport pass + taxis for the awkward trips + rentals for weekends away |
| time cost | door-to-door minutes each way × your value of an hour |
| hard constraints | disability, children, rural location, shift work at night |
If ownership costs, say, $8,000 a year and the alternative mix for your actual trips is $3,500 plus 80 hours of extra travel time, the question becomes "is an hour of my time worth more than $56?", which is answerable. Put in your own numbers; the structure is the point.
Worked template
FIRST-PRINCIPLES BREAKDOWN: <problem>
1. GOAL (outcome, with a number and a unit)
e.g. "cut cost per delivered order from 9.40 to under 5 USD"
2. CURRENT ANSWER BY ANALOGY
what the industry / we currently do, and its cost/performance
3. ASSUMPTIONS CLASS EVIDENCE
a. ............................... law/hard/conv/opinion ...
b. ............................... ... ...
c. ............................... ... ...
4. DECOMPOSITION (tree down to measurable pieces)
total = component A + component B + ...
A = (units) x (price per unit) low / likely / high
B = ...
5. FUNDAMENTAL FLOOR
cost/time/performance if only laws + hard constraints applied
gap between current and floor = ........ (the opportunity)
6. WHERE THE GAP LIVES
component current floor gap convention causing it
7. REBUILT DESIGN
the simplest design that meets the goal using the floor;
conventions added back ONLY with a stated reason
8. RISKS AND CHESTERTON'S FENCES
conventions we are removing - do we know why they exist?
9. CHEAPEST TEST THAT COULD PROVE THIS WRONG
experiment, owner, date, pass/fail threshold
10. DECISION AND REVIEW DATEChecklist
- The goal is an outcome with a number and a unit, not a solution.
- Every assumption is written down and classified; "law" is used only for physics, math or logic.
- Each convention has an owner or origin you can name, or is marked "origin unknown: investigate".
- The decomposition bottoms out in things you can measure or look up, with units that check.
- Each number has a range and a source; the weakest estimate is flagged.
- The floor (physics-only answer) is computed and compared with the current answer.
- You know why each removed convention existed (Chesterton's fence) and have spoken to someone who operates it.
- An expert in the domain has tried to break the model.
- There is a cheap test with a pass/fail threshold before any large commitment.
- You have checked whether the analogy answer was, in fact, close to optimal.
Failure modes
| Failure mode | What it looks like | Guard |
|---|---|---|
| Convention mistaken for law | "enterprise software needs a 6-month sales cycle", treated as fixed | ask "who decided this and when?"; look for anyone who does it differently |
| Law mistaken for convention | "latency is just an engineering problem" when it is the speed of light; "we can grow 20% a month forever" | check against physics, math and base rates before rebuilding |
| Reinventing the wheel | a startup designs its own database or payments stack because "we reasoned from scratch" | first principles is for the core problem; buy or copy everything else |
| Infinite regress | decomposing to quarks when the question is office rent | stop at the level where the numbers become stable and measurable |
| Ignoring tacit knowledge | removing a rule, step or margin whose purpose is invisible on paper | Chesterton's fence: find out why it is there before you remove it |
| Materials floor treated as a target | "the parts cost $80, so we can sell at $100" | the floor excludes processing, labor, testing, warranty, overhead, margin |
| Overconfidence in the model | a neat spreadsheet replaces contact with suppliers and customers | step 7: every model gets a real-world test |
| Selective scepticism | doubting others' assumptions but not your own favorite | classify your own assumptions with the same table |
| Cost of analysis ignored | a week of first-principles work on a $200 decision | use analogy for small, reversible decisions |
| Using it as rhetoric | "I'm thinking from first principles" as a way to dismiss experience | show the decomposition and numbers, or it is just an opinion |
Chesterton's fence
From G. K. Chesterton's The Thing (1929): a reformer comes across a fence across a road and says, "I don't see the use of this; let us clear it away." The wiser reformer answers: "If you don't see the use of it, I certainly won't let you clear it away. Go away and think. Then, when you can come back and tell me that you do see the use of it, I may allow you to destroy it." The point is not "never remove fences"; it is "understand the reason before deciding the reason no longer applies". In first principles work, that means every convention you delete gets a line in step 8 of the template.
Critiques and when not to use it
| Critique | Substance |
|---|---|
| It is not a distinct method | it is ordinary scientific and engineering reasoning; the name adds branding, not technique |
| Survivorship bias | famous examples (batteries, rockets) are remembered because they worked; many "first principles" ventures failed on things the model left out |
| Fundamentals are not always knowable | in markets, organizations and people, the "laws" are statistical and change; decomposition can give false precision |
| Analogy encodes accumulated learning | conventions often exist because many people tried alternatives and they failed; discarding them wastes that knowledge |
| Decomposition misses interactions | reducing a system to parts can miss feedback, delays and emergent behavior (this is what systems thinking is for) |
| It can justify arrogance | "reasoning from scratch" is sometimes a license to ignore experts, regulators or safety margins |
Do not use it when:
- the decision is small, cheap and reversible: copy a good default and move on;
- the domain is well understood and stable, and others have iterated on it for decades (tax filing, accounting standards, most HR processes);
- the "conventions" are safety-critical rules written after accidents (aviation, medicine, electrical codes) unless you have the expertise and the regulator's agreement;
- you do not have the domain knowledge to tell a law from a convention, and cannot get someone who does;
- speed matters more than optimality, and the analogy answer is good enough.
Related thinking tools
| Tool | Relationship |
|---|---|
| Systems thinking | first principles takes a problem apart; systems thinking studies how the parts interact over time (feedback, delays, stocks). Use both: decompose to find the fundamentals, then map the loops before changing anything. |
| General thinking models | inversion, second-order thinking, map vs territory and circle of competence are the companion checks on a first-principles model |
| Decision-making models | reversibility tells you how much first-principles effort a decision deserves |
| Probability and risk | ranges, base rates and expected value turn estimates into decisions |
| Cognitive biases | anchoring on the current price and status quo bias are what analogy reasoning is most exposed to |
| Microeconomics | cost curves, marginal cost and opportunity cost are first principles for business questions |
Practice exercises
- Your biggest line item. Take your company's (or household's) largest monthly cost. Decompose it to units × price per unit, compute a floor, and list which conventions sit in the gap.
- Fermi warm-ups. Estimate, with ranges: liters of coffee drunk in your city per day; the number of software engineers in your country; the energy in a fully charged phone battery, in joules.
- Classify ten assumptions. Write ten beliefs your team holds about your market. Classify each; for every "convention", name who decided it.
- Idiot index hunt. Pick three products you buy (a charger cable, a sofa, a SaaS subscription). Estimate raw input cost and compute the index. Where does the gap go?
- Invert your roadmap. List five ways next quarter could fail. Which one is not addressed by any current plan?
- Five Whys, with branches. Take a recent incident. Do five whys, but allow two answers at each level. Compare the tree with the single chain.
Worked answer: energy in a phone battery
A typical phone battery is around 4,000–5,000 mAh at a nominal ~3.8 V (read the numbers printed on yours).
About 60 kJ: roughly the energy needed to lift a 60 kg person 100 m. Dimensional analysis (Ah × V = Wh) is what makes the conversion safe.
Worked answer: classifying "we need an office to build culture"
Opinion, dressed as a hard constraint. The fundamentals are: how often people need synchronous, high-bandwidth contact (design reviews, onboarding, conflict), how much that costs remotely vs in person, and what rent costs per employee per year. A first-principles design might be a remote team with paid quarterly offsites, or an office used two fixed days a week; either can be tested for a quarter against retention and delivery metrics.
References
- Aristotle, Metaphysics Book V, trans. W. D. Ross (MIT Internet Classics Archive) (opens in a new tab): the definition of "beginning" (archē), 1013a
- Aristotle, Metaphysics Book V (Perseus Digital Library) (opens in a new tab): parallel text with the Greek
- René Descartes, Discourse on the Method, trans. John Veitch (Project Gutenberg) (opens in a new tab): the four rules, Part II
- René Descartes, Meditations on First Philosophy (1641): the method of doubt
- Chris Anderson, "Elon Musk's Mission to Mars", Wired (21 October 2012) (opens in a new tab): the rocket materials-cost quote (~2% of price; 20–25% for a Tesla)
- Kevin Rose, "Elon Musk and Kevin Rose" (Foundation, 2012), YouTube (opens in a new tab): the battery-pack example ($600 vs ~$80 per kWh)
- CNBC, "Why Elon Musk wants his employees to use an ancient mental strategy called 'first principles'" (2018) (opens in a new tab): transcript excerpts of the Rose interview
- BloombergNEF, "Lithium-Ion Battery Pack Prices Fall to $108 Per Kilowatt-Hour, Despite Rising Metal Prices" (December 2025) (opens in a new tab): 2025 average and BEV pack prices
- Walter Isaacson, Elon Musk (Simon & Schuster, 2023): the idiot index and the five-step "algorithm"
- Taiichi Ohno, Toyota Production System: Beyond Large-Scale Production (Productivity Press, 1988; Japanese original 1978): Five Whys and the machine-stoppage example
- Alan J. Card, "The problem with '5 whys'", BMJ Quality & Safety (2017) (opens in a new tab): critique of Five Whys for incident analysis
- Wikipedia: Five whys (opens in a new tab): history and criticisms (including Minoura)
- Wikipedia: Socratic questioning (opens in a new tab): the six question categories with examples
- Wikipedia: Fermi problem (opens in a new tab): background, why estimates work, the piano-tuner example
- Eliyahu M. Goldratt and Jeff Cox, The Goal (North River Press, 1984): theory of constraints and the five focusing steps
- Wikipedia: Carl Gustav Jacob Jacobi (opens in a new tab): "invert, always invert"
- DORA: Streamlining change approval (opens in a new tab): 2019 State of DevOps finding on change advisory boards
- Wikiquote: G. K. Chesterton, The Thing (1929) (opens in a new tab): the full "fence" passage, Ch. IV "The Drift From Domesticity"
- James Clear, "First Principles: Elon Musk on the Power of Thinking for Yourself" (opens in a new tab): popular overview with further examples
