Machine Learning

What Is Machine Learning?

ML fits a function from examples, then checks it on data it has not seen. That habit is how you ship agents.

Features

A feature is a number the model is allowed to see. Garbage features make garbage agents.

Data and Splits

Train, validation, and test sets — and the leakage bugs that make agent evals lie.

Baselines

Always beat a dummy: majority class, last week’s prompt, or a keyword rule. If you cannot, you do not have a win.

Supervised Learning

Every training row has an input x and a target y. No labels, no supervised learning.

Loss Functions

A loss is a number that is small when the model is right. Training is making that number go down.

Gradient Descent

Measure how the loss changes when you nudge a knob, then nudge the knob downhill.

A Linear Classifier

A score, a sigmoid, a cutoff. This is the shape of every routing head you will train.

Hyperparameters

Learning rate, epochs, batch size, and seeds are knobs you choose — on validation, not on test.

Overfitting

The model fits the training sample, including accidents, instead of the pattern that will appear tomorrow.

Regularization

A prior that says be boring: smaller weights, shorter prompts, fewer tools, stop when validation worsens.

Metrics That Matter

Accuracy lies when classes are rare. Count TP, FP, FN, TN. Then precision, recall, and F1.

Rare Classes and Thresholds

When the important class is rare, move the cutoff on validation. Accuracy will not tell you.

Calibration

A score of 0.9 should be wrong about one time in ten. Uncalibrated agents shout 99% on everything.

Ranking

Retrieval and rerank are not classification. Sort by a score and measure precision at k.

Embeddings

Meaning as a list of numbers. Close lists rank together. Retrieval is geometry.

Nearest Neighbors

k-NN labels a new point by a vote of its nearest labeled points. Retrieval is the same geometry without the vote.

Unsupervised Learning

No labels: find structure. Tiny 2-d k-means, then name the clusters by hand.

A Tiny Neural Net

A one-hidden-layer forward pass with lists — linear maps, ReLU, and why depth needs a bend.

Train vs Inference

Training updates knobs. Inference freezes them and only runs the forward pass. APIs you call are inference.

Drift

The world moves. A model fit on last quarter’s words fails on this quarter’s words. Watch the slice.

Rewards and Policies

A policy maps a state to an action. A reward says how well that went. Agents already live in this loop.

Rules, Prompts, or Train

Start with a rule. Then a prompt. Train a small model when the remainder is large, stable, and labeled.

Traces as a Dataset

Production logs are the dataset. Freeze ids, write a rubric, label a sample, version the split.