What does Metcalfe's Law claim about network value, and what is the chapter's caveat?	Value grows as the square of connected nodes; the caveat is that this overstates for mature networks where most node pairs never actually communicate.
Name the two genuine compounding loops the chapter identifies in the AI stack.	The data flywheel and the agent flywheel.
What is the data flywheel loop, step by step?	Deploy model → users interact → interactions produce signal → signal fine-tunes the model → next version is better → usage grows → more signal.
What is the agent flywheel loop, step by step?	Agent accesses a tool → accomplishes task → success is logged → tool interface is refined → next agent does better → new tools are exposed → capability grows.
Does a strong data flywheel let a weaker model catch a stronger one?	No — the flywheel amplifies an existing lead; it does not invert a quality gap.
What is 'model collapse' and who named it?	The recursive narrowing of training data on dominant-model biases when models train on other models' outputs; named by Shumailov et al. (2023).
What is 'tool-bloat collapse' in agent systems?	When the tool ecosystem grows past a certain size, models lose track of which tool to use and performance degrades — several agent frameworks have hit this wall.
How many years of headstart does the chapter estimate current AI incumbents have if they execute well?	~3–5 years.
Why does the chapter say AI moats are less durable than pre-AI internet platform moats?	Because AI capability depends partly on model architecture, which improves rapidly across the whole industry through open releases — a late entrant with a better model can catch up faster than in social networks or marketplaces.
By roughly how much did the deployed agent population grow from 2024 to 2026?	~10× — roughly an order of magnitude.