Five laws for what lasts as AI gets cheaper
Each one is a question to ask before you trust a system, adopt a tool or back a bet, plus two principles for when several bind at once.
01 · Five laws
Each is written so it can lose: its page says what would break it, and how it changed when a test did.
- Law I Bottleneck Migration When a layer gets cheap, value moves to the most stubborn layer next to it: usually up, and down to compute, power or regulatory access when the scarcity is physical. AskIf this layer becomes abundant, where does scarcity move next?
- Law II Difficulty Is Load-Bearing A hard part carries value when it gives you a second, independent way to check the answer; effort along the same route can go. It covers knowing, not training. AskDoes this hard part give you a second, independent way to check the answer, or only more effort along the same route?
- Law III Architecture Outlives Content In a system that runs a process, the structure lasts while the content turns over; in one that preserves something, the content lasts and the structure can be rebuilt from it. AskDoes this system run a process or preserve something? Then ask what survives when the pieces change.
- Law IV Instruments Over Theory Hidden structure stays hidden until you build the instrument that measures it, in fields where things are built before they are understood; in fundamental physics and pure maths, theory usually leads. AskWhat would make the hidden structure visible?
- Law V The Targeting Problem Misaimed capability always wastes the effort, but harms the system only when the miss is hard to spot or correct, so a target that looks wrong is a hypothesis to test. AskIs this aimed at the right target, and how cheaply would you find out if it is not?
02 · Two principles built from them
What goes wrong when several of the laws bind at once.
- Principle A · from Laws II, IV, V Goodhart Corollary Optimise a stand-in hard enough and it parts company with what it stands for, after tracking it at first, unless the number fully captures the thing, which you usually cannot know in advance. AskIf this number keeps going up, at what point would it stop meaning what you want, and what separate check would tell you?
- Principle B · from Laws IV, V Regime Problem Work out which situation a system is in before you choose a method: the right method in the wrong situation gives a confident, wrong answer, with no error to warn you. AskWhich situation is this system in, and would this method tell you if you guessed wrong?
The newsletter · Free
Subscribe to get each new piece in your inbox, with something to use in every one.
Several essays a week, plus Notes through the week.
Signup completes on Substack. Unsubscribe from any email.