# Data Flywheels & Cold Starts

**Track:** Building the AI-Native Product — AI for Entrepreneurship — complete (29)
**Framework / surface:** venture strategy
**Level:** Intermediate
**Prerequisites:** Evals Are the Product Spec

**In one line:** Usage that makes the product better — engineering the loop and surviving the empty start.

## Theory, aesthetics & inspiration

A data flywheel is the loop where using the product generates data that makes the product better, which attracts more use — the compounding engine behind the strongest AI-native businesses, and the modern home of Andrew Ng's data-centric argument that improving your data beats improving your architecture. The loop must be engineered, not assumed: capture the correction, not just the click; structure feedback so it becomes evals and fine-tuning corpora rather than a sentiment dashboard; close the loop visibly so users see their corrections take effect. Andrew Chen's "The Cold Start Problem" (2021) names the hard part — every flywheel begins stationary — and the AI-era answers are concierge phases that manufacture early data, synthetic data to rough in coverage, and wedge markets small enough that modest data still clears the local quality bar. One caution the exponential adds: generic capability improvements accrue to everyone renting the same model, so the flywheel only defends you when the data it captures is workflow-specific — the corrections only your customers, in your product, could have produced.

**Founder question:** What does your product learn from each use that only your product could learn?