Understanding AI (2/4): How AI really learns ?

Training on labeled data where each example includes the “correct answer.” Example: thousands of cat vs dog images so the model can recognize them later.

Learning without labels: the model finds hidden structure in the data. Example: grouping customers by purchasing behavior without predefined categories.

Learning by trial and error with rewards. Example: an agent learning to walk, or an AI becoming a chess champion.

A technique used for systems like ChatGPT: humans rate outputs so the model learns responses that are more helpful, natural, and context-aware.

Reusing knowledge learned from one task to accelerate learning on another. Example: a vision model trained on animals can be adapted to recognize vehicles.

The model creates its own “labels” from raw data. This approach is central to large language models like GPT.

Training happens on users’ devices rather than centralizing all data. The goal is collaborative learning while preserving privacy.

The model keeps learning after deployment without losing what it already knows—e.g., integrating new cases while maintaining accuracy.

Making a model “forget” certain data without retraining from scratch—useful for privacy requirements and the right to be forgotten.