data flywheel
In plain wordsA cycle in which a product's users generate data that improves its AI, the better AI attracts more users, and those users generate still more data.
A self-reinforcing loop in which a deployed AI system produces data — user interactions, corrections, ratings, failure cases, sensor logs — that is collected, labeled or filtered and used to retrain the system, whose improvement attracts more use and therefore more data. Engineering versions of the loop mine production traffic for errors and hard cases, route them to labelersdata labeling, n.: The work of attaching correct answers to examples — often done by large paid workforces, sometimes by experts — so that AI systems can learn from them or be tested against them. or automated graders, retrain and redeploy, as in the fleet-learning 'data engine' described for driver-assistance systems. In language-model products, feedback buttons, edited responses and agent trajectories feed post-trainingpost-training, n.: Everything done to an AI model after its initial bulk training on huge amounts of text: the extra teaching that turns a raw text predictor into a helpful, safe assistant or a careful reasoner.. The business claim that such loops create durable advantages is disputed: the value of additional data often diminishes, data from one product may not transfer to another, and competitors can turn to public data, synthetic datasynthetic data, n.: Data created by computer programs or AI models rather than collected from the real world — such as practice math problems written by a model, or simulated street scenes for self-driving cars. or distillationknowledge distillation, n.: Training a smaller or cheaper 'student' AI to copy the behavior of a larger 'teacher' AI, so the student keeps much of the teacher's skill while being faster to run. instead.
Every correction a radiologist makes goes back into our data flywheel for the next model release.
Origin
Applies the flywheel image of management writing — Jim Collins's Good to Great (2001) and Amazon's 'virtuous cycle' — to machine learning, where it became common in the 2010s; Andrej Karpathy's descriptions of Tesla's 'data engine' from 2019 are a well-known engineering example.
Usage note
Skeptics, notably a 2019 Andreessen Horowitz essay on 'the empty promise of data moats', argue that data advantages erode faster than flywheel stories assume. Loops can also amplify errors and bias when a system's outputs shape the data it later learns from.
Related
- data-centric AIdata-centric AI, n.: An approach to building AI that focuses on improving the data — fixing labels, adding missing cases, cleaning inconsistencies — rather than endlessly tweaking the model.
- human-in-the-loophuman-in-the-loop, adj.: Describes AI systems in which a person checks, corrects or approves the machine's work at key points, rather than letting it act entirely on its own.
- active learningactive learning, n.: A way to train a model with fewer labeled examples by letting it choose which unlabeled examples a person should label next, usually the ones it is most unsure about.
- moatmoat, n.: A lasting business advantage that keeps competitors out, like a castle's moat; in AI, the question of whether any company can stay ahead when rivals and free models catch up quickly.
- post-trainingpost-training, n.: Everything done to an AI model after its initial bulk training on huge amounts of text: the extra teaching that turns a raw text predictor into a helpful, safe assistant or a careful reasoner.
- model monitoringmodel monitoring, n.: Continuously checking an AI model after it goes live — what data it receives, what it predicts and whether it is still right — so that problems are caught before they cause harm.
Mentioned in
data-centric AIdata-centric AI, n.: An approach to building AI that focuses on improving the data — fixing labels, adding missing cases, cleaning inconsistencies — rather than endlessly tweaking the model.