Classical ML vs Generative AI

Two hammers, different nails

• Classical ML (this track): learns from labeled tables — churn, fraud, prices, demand. Models are small, fast (milliseconds), cheap (pennies per million predictions), and explainable ("tenure drove this decision").
• GenAI (AI track): handles open language — summarize, chat, draft, reason over messy text. Flexible and fluent, but slower, pricier, and occasionally confidently wrong.

Neither replaces the other. The skill is reaching for the right hammer without ideology.

The decision rule (with examples)

Structured rows + a clear right answer → train a small model. Fraud detection on transaction tables, demand forecasting from sales history, churn from usage logs — a random forest beats a prompted LLM on accuracy, cost, and latency here, while running on a laptop.

Messy text + judgment → prompt an LLM. Summarizing support threads, drafting replies, extracting terms from contracts — no labeled table exists, and writing rules is hopeless. This is the home turf you learned in the AI track.

The power combo

Many real products use both, each where it wins: a classical model scores (fraud probability: 0.87) and an LLM explains ("flagged for review: three out-of-state charges in an hour"). Cheap filter first, expensive reasoning only where judgment is needed — the same triage pattern as model routing in production AI.

Remember this

• Tables + labels + budgets → classical ML. Messy language → LLMs.
• Best systems combine: score cheaply, explain richly.