David Deutsch's framework, cited by Cohen as a personal lens: progress compounds without limit once you have a good explanation — a causal account of why something happens, not just a correlation or a pattern that works. Good explanations are hard to vary without breaking, which makes them a stable foundation to iterate on; correlational or trial-and-error knowledge is brittle and doesn't compound the same way.
Cohen applies this to building: understand root causes before layering iteration on top of them, rather than optimizing a black box. Consistent with his critique of blanket-access AI agents that pattern-match instead of reasoning from curated ground truth (see Golden-Examples Curation Beats Blanket Knowledge-Base Access).
Из тем: Unsorted