6 ideas
Using many models simultaneously produces better understanding than relying on one, because each model captures a different slice of reality and their errors are partly independent. Like an ensemble of diverse predictors, a collection of moderately accurate, low-correlation models outperforms even the single best model when their insights are combined.
Data are raw observations, information is structured data, knowledge is models and relationships explaining patterns, and wisdom is knowing which model to apply in which situation. Models occupy the knowledge tier, but the meta-skill of selecting the right model for a context is what constitutes wisdom.
Total predictive error equals average individual error minus the diversity of predictions (the diversity prediction theorem). This means making a crowd or model set more diverse reduces collective error just as much as making each member more accurate, so diversity is mathematically equal in value to individual skill.
When outcomes follow a power-law distribution rather than a normal one, it indicates underlying mechanisms like preferential attachment or self-organized criticality where success breeds success and large events are not anomalies but inherent. Treating such fat-tailed domains with normal-distribution intuitions catastrophically underestimates the frequency of extreme events.
Formalizing assumptions into a model forces internal consistency and exposes when a conclusion does not actually follow from its premises, defeating vague hand-waving. The discipline of building a model reveals the limits of your own knowledge, making you simultaneously more rigorous and more humble about what you can claim.
People and systems sort the continuous world into discrete categories, then treat everything within a category identically. This coarse-graining enables fast decisions but creates systematic blind spots, because relevant within-category variation is ignored and the choice of categories itself shapes which patterns become visible.