Prediction Machines (AI as Cheap Prediction)

Definition

From Prediction Machines, the economic lens on AI: economists don’t judge technology by how impressive it is, but by price. The current AI wave is not “intelligence” but a prediction technology becoming radically cheap. Prediction is the process of filling in missing information — using data you have to generate information you don’t. Just as lighting costs collapsing ~400× in the 19th century put electric light everywhere, cheap prediction becomes near-free infrastructure that lights up domains previously too costly to enter.

Core Ideas

Decision = Prediction + Judgment

Any decision decomposes into two steps:

• Prediction (machine) — tells you what will happen.
• Judgment (human) — tells you what to do about it. Judgment requires values, weighing consequences, emotional connection, and taking responsibility — which AI cannot supply.

Substitutes vs Complements

When the price of prediction falls toward zero:

• Devalued (substitutes) — work that is only prediction: routine logic, standardized data analysis, basic translation, entry-level support/copy. These get replaced.
• Appreciated (complements) — the complements of prediction become scarce and expensive: human judgment (setting goals and decision frameworks), unique/proprietary data, and the execution to turn a prediction into action.

Three layers of AI application

The gap between winners and everyone else is whether you move up the layers:

1. Tool layer (efficiency) — AI optimizes existing tasks (drafting email, summarizing, tidying data); the work is unchanged, just faster.
2. Decision layer (optimization) — precise prediction enables higher-quality, previously impossible decisions (e.g. precisely targeting churn risk).
3. Strategy layer (reinvention) — rebuild the business model entirely (e.g. Amazon’s “ship-then-buy” anticipatory model, impossible without prediction).

Data as strategic moat

Algorithms are copyable; data ecosystems are not (Intel’s Mobileye acquisition bought self-driving data collection, not code). Protect your own data and go deep in a vertical where the data and judgment you hold are hardest to replicate.

Counterfactual thinking to train judgment

Don’t passively accept an AI prediction. Interrogate it:

1. If this prediction were 20% too optimistic or pessimistic, how would my decision change?
2. Which direction of error (over- vs under-preparing) is the one I can least afford?
3. If everyone can get this same prediction from AI, does my unique value still hold?

Don’t compete with the machine — collaborate with it. Give prediction to the machine; keep judgment for yourself.

Relationships

• Economic Thinking — the substitute/complement and price reasoning here are applied microeconomics
• Measurement Dysfunction — counterfactual interrogation guards against blindly optimizing a predicted metric
• Machine Learning — the underlying prediction technology whose cost is collapsing

References

• AI极简经济学（Prediction Machines）核心精髓