Foundations

AI vs. Machine Learning vs. Deep Learning

These three terms get used interchangeably, but they nest inside each other. Artificial Intelligence is the broad goal — software that performs tasks we'd call intelligent. Machine Learning is the dominant method of getting there: instead of writing explicit rules, you show a system many examples and it derives the patterns itself. Deep Learning is a subset of machine learning that uses many-layered neural networks, which is what made the last decade's leaps in vision, speech, and language possible.

• AI — the goal: systems that perform tasks requiring intelligence
• Machine Learning — the method: learn patterns from data instead of hand-coded rules
• Deep Learning — the technique: layered neural networks that learn their own features

The Three Learning Types

How a model learns depends on what kind of feedback it gets. Almost every practical business system you'll encounter is supervised learning or a generative model built on top of self-supervised pretraining.

• Supervised — trained on labeled examples (this email is spam, this scan shows a fracture). Best when you have historical outcomes.
• Unsupervised — finds structure in unlabeled data: clustering customers, spotting anomalies, compressing dimensions.
• Reinforcement — learns by trial and reward over time. Powers robotics, game-playing, and the preference-tuning stage of modern chat models.

Training vs. Inference

Training is the expensive, one-time (or periodic) process of teaching a model — it consumes enormous compute and data. Inference is what happens every time you actually use the model: you send input, it produces an output. Your ongoing operating cost is almost entirely inference. This distinction drives most architecture and budget decisions: you rarely need to train anything; you need to use an existing model well and control what each call costs.

Tokens, Parameters, and Weights

Language models don't read words — they read tokens, which are word fragments (roughly four characters each). Pricing, context limits, and speed are all measured in tokens. Parameters are the internal numbers the model learned during training; weights are those numbers' values. Parameter count is a rough proxy for capability and cost, but it's not the whole story — smaller, well-tuned models routinely beat larger ones on narrow business tasks.

• Tokens — the billing and context unit; ~750 words per 1,000 tokens
• Context window — how many tokens the model can consider at once
• Parameters/weights — the learned internals; more is not automatically better