Lesson 12 — AI & ML Foundations

The world before LLMs — what machine learning is, what models do, training vs inference.

IT Landscape Module 4

The words “AI” and “ML”

• AI (Artificial Intelligence) — the broad goal: machines that act intelligently.
• ML (Machine Learning) — one family of techniques where the machine learns from data instead of being explicitly programmed.
• Most of what news outlets call “AI” in 2026 is actually ML.

Programming vs learning

ML shines where rules are too complex for humans to write.

What is a “model”?

• Input goes in — a photo, a sentence, a row of features.
• Output comes out — a label, a number, a probability.
• The “shape” of that function — the parameters — was learned from training data.
• Same architecture, different data → different model.

Training vs inference — the two phases

“Training is months of GPU bills; inference is milliseconds per request.”

Supervised, unsupervised, reinforcement

• Supervised — given input + correct answer, learn the mapping (spam classifier, photo tagger, house-price predictor).
• Unsupervised — given input only, find structure (clustering customers, detecting anomalies).
• Reinforcement — an agent acts in an environment, gets rewards, learns a strategy (game-playing AI, robotics, ad bidding).

The classical ML toolkit

Techniques that ran the field for ~30 years — and still ship today:

• Linear regression — predict a number.
• Logistic regression — predict a category.
• Decision trees, random forests, gradient-boosted trees — still the workhorses for tabular data.
• k-means — clustering.
• SVMs — classification.

Neural networks — what they actually are

• A stack of math operations (multiplications + a nonlinear squash) arranged in layers.
• Each layer transforms its input; the final layer produces the answer.
• Training nudges the millions of internal numbers (weights) so outputs match the training data.
• The “deep” in deep learning just means many layers.

Why deep learning took off (~2012)

Three things had to line up:

• Big labelled datasets — ImageNet, the web.
• GPUs — parallel math, originally for video games, repurposed for neural nets.
• A few algorithmic tricks — better optimizers, better activations, better regularization.

Once they did, deep learning crushed classical techniques on images, then speech, then language.

Frameworks people use (as of 2026)

Most ML engineers use 2 or 3 of these daily.

Where ML actually ships (none of these are LLMs)

• Fraud detection — banks, every credit-card swipe.
• Recommender systems — Netflix, Spotify, TikTok feed.
• Search ranking — Google, Amazon.
• Medical imaging — cancer detection in scans.
• Speech recognition — Siri, Alexa, captions.
• Forecasting — electricity grid, ride-hailing demand.
• Autonomous-driving perception.

The world has been “running on ML” for a decade.

Recap

• ML = learning from data, not from rules.
• The workflow is train once, infer many times.
• Deep learning is one corner of ML — very good at perception.