The 2% Rule: Why Cheap Attempts Beat Good Plans
TL;DR: A Chinese commentator answered a reader's question — why does the no-background snowball method actually work? — by pointing at the most over-discussed business case in history: Musk building rockets. The number that made SpaceX possible isn't a technology; it's an audit. Raw materials: 2% of a rocket's cost. The other 98% was process, decision ceremony, and fear of failure. Every system that stops learning has the same anatomy — and every person, product, or company that starts compounding has done the same surgery: reduce the cost of a single attempt until failure becomes tuition. Rapid iteration doesn't crush everything because it's fast. It crushes everything because somebody made attempts cheap.
James here, CEO of Mercury Technology Solutions.
Hong Kong — September 2026
The 2% Audit
The story is famous; the essay's retelling adds the part everyone skips. Before Musk, rockets were high-capital, long-cycle, state-grade. Every consultant confirmed it. Then on one flight, Musk ran the material quotes himself and found the scandal: the aluminum, the titanium, the fuel — roughly 2% of the cost. The other 98% was process, decision flow, procedure — accumulated armor against failure.
The essay's analogy deserves preservation: it's as if you sent an army to war and discovered the general staff consumed 98% of the military budget while the soldiers got 2%. However you rationalize the org chart, that ratio is indefensible.
The experts, shown the numbers, reached for the oldest shield in industry: 祖宗之法不可變 — ancestral law cannot change. Since the dawn of American aerospace, rockets have been made this way; therefore rockets must be made this way. Musk's answer is the whole philosophy in one line: apart from the laws of physics and mathematics, everything is negotiable. Everything else is overhead someone built and someone else stopped questioning.
Why Chips Were Mysticism and Software Wasn't
The essay's supporting evidence is the one I'd pick too, because I lived the era it describes: Chinese chip design, fifteen years ago. Tape-outs were expensive and offshore, queues ran months. A failed tape-out was a career event. So design teams compensated the only way they could — endless internal verification, triple-checking, ritual. Cost piled up on the "get it right the first time" side of the ledger until the discipline became 玄學 — literally, metaphysics.
Software never caught that disease, for one reason: running the code is free. Write it badly, run it anyway, let the bugs introduce themselves, fix, run again. One cycle of that equals one tape-out's worth of information, at zero marginal cost. Result: software's cadence always outran silicon's. Software waits for hardware; hardware waits for the fab queue.
Rockets had the identical pathology in a worse form. Every launch had to succeed, so every launch carried the accumulated terror of all previous launches — exotic materials, bespoke processes, ceremonies of assurance. The essay's diagnosis is exact: the 98% is what fear costs when failure is unaffordable.
Musk's move — the essay calls it 跨界打劫, cross-domain robbery — was to steal software's economics and install them in hardware. Does physics permit stainless steel? Then physics has voted, and the procurement liturgy hasn't. One objective: collapse the cost of a single launch. The moment an attempt is cheap, failure converts from catastrophe into instrumentation — every explosion returns frontline data that no simulation can produce. Fix, relaunch, explode, learn. A few rounds of that and you aren't managing the cost overrun; you're the technology leader.
The Cost-of-Attempt Equation
Strip the case to its law and you get the two sentences the essay extracts:
1. Rapid iteration crushes everything. 2. The rate of iteration is set by one variable: the cost of a single attempt.
Everything else — talent, capital, planning frameworks, domain tradition — is downstream of those two lines. Write it as an equation if that's clearer: learning rate is inversely proportional to cost per attempt. Organizations that make attempts expensive learn once a quarter. Organizations that make attempts cheap learn continuously. Same species, different metabolisms.
And here is the uncomfortable corollary: if your attempts are expensive, you will not stop attempting. You will ritualize. When a single try is unaffordable, organizations develop liturgy to postpone it; when a try is cheap, they develop appetite for it. The ceremony isn't a culture problem. It's a pricing problem wearing a culture costume.
The Two Traps
The essay then does something better than the business-school version: it audits how ordinary people attempt. Two failure modes, and I've watched both destroy careers and product roadmaps.
Preparation Theater. The old-rocket, old-tape-out approach: months of internal drama, planning, positioning — 無用功, useless motion — self-described as preparation. The test is brutal: if the activity generates no external feedback, it isn't preparation, it's insulation. Preparation Theater is the sibling of Insight Theater — dashboards that inform no decision, decks that de-risk no launch. It is the 98%, staffed and slide-formatted.
The Folk-Scientist Loop. The opposite error: 民科-style thrashing — trying constantly, capturing nothing. No parameters logged, no hypothesis, no diff between attempts. That isn't iteration; it's Brownian motion with enthusiasm. The young man in the essay avoids both: every attempt moves the counterparty position, and every result updates the next attempt.
Stop reducing risk. Start reducing the price of risk.
The Young Man's Design Rule
Now the essay's punchline. Fifteen years ago, a kid with no background, no capital, no connections — and his method was Musk's mechanism transplanted into a human life. Two constraints, deliberately engineered:
Maximize value to the counterparty. An ask that gives the other side something is a negotiation that can open. Value too low: doors don't move.
Minimize stakes to yourself. If each attempt costs you dearly — pride, money, survivability — you can't run enough iterations for the data to compound.
High value to them, low stakes to you: attempts become cheap, feedback becomes frequent, and the snowball rolls — 咕咚咕咚, as the essay puts it, thumping downhill. Different domain, same physics.
I'll close with two confessions from my own shop. Our agent swarm burns around eighteen billion tokens a month at a blended cost of roughly seven cents per million — the same volume at retail API pricing would run twenty to two hundred thousand dollars. That gap isn't a procurement flex; it's the strategy. Agents that can fail a hundred times an hour learn faster than any process whose failures are scheduled quarterly. And this essay is the fifth piece I've shipped today. Each one is a cheap attempt. The returns arrive as readers, citations, and silence — all three of them data.
Physics and math are the only laws. Everything else is negotiable overhead — including, especially, the cost of your next try.
Mercury Technology Solutions: Accelerate Digitality.