Machine learning is a broad approach: systems that improve at a task by learning from data rather than following hand-written rules. Generative AI is a specific application of ML — using learned patterns to produce new content instead of predicting a label or a number.
Traditional (Non-Generative) ML — Real Examples
| Task | Typical ML Approach | Output |
|---|---|---|
| Predict house price | Linear regression, gradient-boosted trees | A number |
| Classify an image as cat/dog | CNN classifier | A label |
| Predict customer churn | Logistic regression, random forest | A probability |
None of these produce new content — they map an input to a fixed-shape prediction.
Generative ML — What Changes
Traditional ML: input → model → label / number / probability
Generative ML: input → model → new content (text, image, audio)Same Underlying Discipline, Different Objective
Both use the same core ML toolbox — training on data, optimizing a loss function, evaluating on held-out data. What differs is the objective the model is trained to minimize: a classifier is trained to predict the correct label; a language model is trained to predict the next token in a sequence as accurately as possible, across enormous amounts of text.
Practical Use Case
A retail company might use traditional ML for demand forecasting (regression) and inventory classification (classification), while using generative AI for writing product descriptions and answering customer support questions — genuinely different problems solved by the same broader field, using different techniques.
Common Mistakes
- Assuming generative AI replaces traditional ML — for structured prediction tasks (fraud scoring, churn prediction, demand forecasting), traditional ML is typically more accurate, interpretable, and far cheaper to run
- Using an LLM to solve a problem a simple classifier would handle better and far more cheaply — a common, costly mistake in real AI product decisions
Interview Relevance
Q: "When would you use traditional ML instead of an LLM?" A strong answer: when the task has a fixed, well-defined output space (a label, a score, a ranking) and you have labeled training data — traditional ML is usually cheaper, faster, more accurate, and easier to explain than reaching for an LLM.
Practice Question
You need to flag potentially fraudulent transactions in real time with sub-100ms latency. Would you reach for an LLM or traditional ML, and why?
