Nature First, Human Rules Second

Nature First, Human Rules Second | Generated by AI

Home 2026.08

Question: Is studying natural/mathematical/computational laws more essential than studying human-made (corporate) rules?

Answer: Yes, with a caveat worth being precise about.

Why nature/math/code wins:

Invariant. F=ma doesn’t get amended in a board meeting. A proof either holds or it doesn’t. Code either compiles or it doesn’t.
Verifiable by you alone. No committee needed to check if your gradient descent converges.
Compounding. Master linear algebra once, use it for 20 years. Master your company’s Jira workflow, and it’s obsolete in 18 months when someone reorgs.
Adversarial-noise-free. Corp rules are frequently optimized for legal defensibility, liability allocation, or someone’s ego — not for truth. Reading them as if they encode signal is often a category error.

Where the claim overreaches: Corporate bureaucracy isn’t arbitrary the way it feels — it’s a lawful emergent system too. Conway’s Law (your org chart becomes your system architecture), Goodhart’s Law (any metric you optimize stops measuring what you wanted), principal-agent problems — these are as “natural” as thermodynamics, just running on a messier substrate (humans with incentives) instead of particles. DeepSeek under Liang Wenfeng and early OpenAI under Brockman weren’t successful because they ignored organizational design — they were successful because they treated org design itself as an optimization problem and drove the free-parameter count toward zero. Flat structure, small teams, minimal process — that’s not “no rules,” that’s rules engineered to get out of math’s way.

So the actual lesson isn’t “ignore human rules, study nature only.” It’s:

Spend ~90% of your study budget on things with unforgiving feedback loops: math, physics, ML internals, systems/distributed computing, the actual weights and gradients.
Spend ~10% on the minimum human-rule literacy needed to not get blindsided running a company — contracts, IP, export control (relevant to you specifically, given cross-border AI work), tax. This isn’t “essence,” it’s insurance. Skipping it entirely isn’t a badge of purity, it’s a landmine.
Treat org/process rules you do encounter (yours or a client’s) as a system to reverse-engineer for its incentive structure, the same way you’d read someone else’s codebase — not as ground truth to memorize.

The instinct to prioritize is right. The instinct to treat human systems as categorically beneath study is where I’d push back — Goodhart and Conway are as real as anything in your notes-en, just running on wetware.

Back Donate