Value the goose, crack every egg

What the paper argues

This note concerns AI systems that write and ship code, read through an investor’s metaphor. On stage in Tokyo, Masayoshi Son stood in front of a goose laying golden eggs, with slides reading “eggs do not lay eggs”. His argument is that the market makes a mistake by pricing only the visible eggs and ignoring the thing that produces them. On the investment point, he may well be right.

Carried across into software, the framing inverts. The egg is the output you can inspect: the pull request, the config change, the generated function. The goose is the process that laid it: the model, the agent, the capability. Son’s pitch is the same pitch made for every AI system that writes and ships code: trust the mechanism, it is smarter than you, stop checking each egg. The discipline of shipping safe software runs the other way. You keep a competent check at the point where the egg leaves the factory, and you make sure that check can halt the line.

Key points

• Every pitch for an autonomous system carries Son’s line in disguise: the author is more capable than you, so stop inspecting the output. That is capability being offered as a substitute for a check. It is not one.
• Capability is the most measured variable in AI and the least decisive. Knight Capital did not lose 440 million dollars because the code was written by something insufficiently clever, but because nothing at the point of action could stop what was already running.
• There are two separate trust questions. Is the mechanism sound, and is this specific output sound? A capable goose raises confidence on the first and tells you nothing about the second.
• The check belongs where the egg leaves the factory, not three steps downstream, and it earns its place only if it can halt the line. A check that cannot say no is decoration.
• A more productive goose makes the check more important, never less. Scale the checking with the laying.