For years, the standard approach in machine learning was model-centric. You'd get a dataset, treat it as a fixed source of truth, and spend most of your time tweaking model architectures or tuning hyperparameters to squeeze out a bit more performance. The data was the constant; the model was the variable.

Data-centric AI flips this script. It recognizes that in many real-world applications, the biggest gains don't come from a slightly better model, but from better data. The new approach is to hold your model architecture constant and systematically improve the quality and coverage of your training data. This shift treats data not as a static asset, but as a dynamic, engineerable part of the system.

The bottleneck in most machine learning projects is the creation of a large, high-quality labeled training dataset. Manually labeling thousands or millions of data points is slow, expensive, and difficult to adapt. Once you hand-label a dataset, it's static. If your project requirements change, you often have to start over.

Data programming offers a solution: label data with code. Instead of labeling data points one by one, you write functions that express patterns, heuristics, or domain knowledge. These labeling functions (LFs) can label vast amounts of data almost instantly. This approach transforms data labeling from a manual task into a software development process. Your LFs can be versioned, tested, and updated just like any other piece of code.

Labeling functions are powerful, but they aren't perfect. They are a form of weak supervision—sources of labels that are cheaper and more scalable than manual annotation, but also noisier and less accurate. An LF might be right only 70% of the time. Two LFs might apply to the same data point and disagree on the label. Another LF might only apply to a tiny fraction of the data.

This is expected. The key is not to create perfect LFs, but to combine many diverse, imperfect signals. These signals can come from various sources:

• Heuristics: Simple rules and patterns, like searching for keywords.
• Distant Supervision: Existing knowledge bases or databases that can be used to label data.
• Legacy Models: An older machine learning model can be used as a noisy labeler.
• Subject Matter Expertise: The encoded knowledge of domain experts.

The challenge is to systematically resolve the conflicts and model the noise from these sources to produce a single, high-quality training label for each data point.

This denoising process is handled by a generative label model. Instead of looking at the data features, this model looks at the patterns of agreement and disagreement between the various labeling functions. By observing where different LFs overlap and conflict across the entire dataset, it can mathematically estimate the accuracy and correlation of each one—without ever seeing hand-labeled ground truth.

Based on these learned accuracies, the model then re-weights and combines the outputs of all the labeling functions for each data point. The result is not a single hard label, but a probabilistic or belief score (e.g., 85% chance this is 'SPAM', 15% 'NOT SPAM'). These probabilistic labels are then used to train a powerful, discriminative model (like a deep neural network), which learns to generalize beyond the original heuristics.

This data-centric workflow—using programmatic labeling and weak supervision, unified by a generative model—allows teams to build and adapt powerful AI applications far more efficiently than with manual labeling alone.