A data flywheel, sometimes called a data engine, is a self-reinforcing loop in which a deployed model continuously improves by feeding on the data its own use generates. A model in production encounters situations it handles poorly; those hard or ambiguous cases are detected, collected, labeled, and added to the training set; the model is retrained and redeployed; and the improved product attracts more users and more data, which surfaces the next batch of edge cases. Each turn of the loop makes the model better and the data richer.
Tesla’s autonomy work is a much-cited example. In his June 2021 CVPR Workshop on Autonomous Driving keynote, Tesla’s then senior director of AI, Andrej Karpathy, described how the company’s fleet is used to identify rare driving scenarios that the system should learn from, gather examples of them, and iterate the model around them, spinning this loop repeatedly until accuracy is high enough. The idea connects closely to active learning, where the model itself helps choose which examples deserve human labeling.
The flywheel framing explains why early scale can compound into a durable advantage: more deployment yields more data, which yields a better model, which yields more deployment.
For a business reader, the data flywheel is the strategic reason data-rich incumbents are hard to catch, and the reason getting a product into real use early can matter more than a marginally better initial model.