Machine Learning Basics

Modern AI rides on top of decades of classical machine learning. Even when the day-to-day work is just calling a foundation model, the failure modes — overfitting, distribution shift, label noise, bad evaluation — are inherited from classical ML. This page is the shared vocabulary I want any engineer joining the AI conversation to have.

Supervised, Unsupervised, and Reinforcement Learning

The three classic paradigms still cleanly describe most modern systems:

• Supervised — learn a mapping from inputs to known labels. Image classification, spam detection, most fine-tuning.
• Unsupervised — find structure without labels. Clustering, dimensionality reduction, much of pre-training.
• Reinforcement learning — learn a policy by interacting with an environment and receiving rewards. RLHF for LLMs, game-playing agents, robotics.

The Bias–Variance Tradeoff

Every model balances:

• Bias — the error from the model being too simple to capture the underlying signal.
• Variance — the error from the model being so flexible it captures noise.

Underfitting is high bias; overfitting is high variance. Regularization, more data, and simpler model classes pull toward bias; bigger models and more features pull toward variance.

Train / Validation / Test Discipline

The single most common ML mistake is contaminating evaluation:

• Train — what the model learns from.
• Validation — what you use to tune hyperparameters and pick model variants.
• Test — touched once at the end to estimate real-world performance.

If you tune on the test set, you no longer have an honest estimate. The same logic applies to LLM eval sets.

Common Failure Modes

• Distribution shift — production data drifts away from training data.
• Label noise — training labels themselves are wrong.
• Leakage — a feature secretly contains the answer.
• Spurious correlations — the model learns the wrong shortcut.

Why This Still Matters

Even when "the model" is an API call to a frontier LLM, you are still building a machine learning system: choosing data, defining evaluation, watching for distribution shift. The classical lens is what keeps modern AI work honest.