LLMs Deep Dive

Course Content

LLMs Deep Dive

10 sections · 40 lessons

What is the difference between discriminative AI and generative AI?

One shopping answer, four modelsIntentclassifierroutes the queryRanker scores3,000 phonesLLM writesthe comparisonChecker matchesprices to catalogueDiscriminative, discriminative, generative, discriminative.
The generative step sits between discriminative ones, and it was the final check that caught 'under Rs 20,000' on a Rs 21,499 phone.

What you need to know

How the engineering differs

Discriminative AIGenerative AI
OutputLabel, score, rankText, image, audio, code
Correct answerUsually oneMany acceptable answers
EvaluationAccuracy, precision, recall, AUCRubrics, human review, LLM-as-judge, task success
BehaviourDeterministic given the modelVaries with sampling
Main risksBias, drift, false positivesHallucination, unsafe content, prompt injection, data leakage
Cost per callMicro- to milliseconds, cheapHundreds of ms to seconds, per-token cost

Why generative outputs are harder to trust

A fraud classifier's worst case is a wrong label, which you can measure. A generative model can produce a fluent paragraph with one wrong number in it, a policy that does not exist, or text copied from a malicious document it read. So generative systems need:

  • Grounding — retrieval of the facts, with citations.
  • Output checks — schema validation, moderation classifiers, numeric checks against source data.
  • Input defences — treating retrieved text and user text as untrusted.
  • Human review for high-risk actions.

How they combine

  1. Route (discriminative) — an intent classifier decides what the user wants.
  2. Retrieve and rank (discriminative) — a ranker picks the best documents or products.
  3. Generate (generative) — an LLM writes the answer from the retrieved facts.
  4. Check (discriminative) — a safety or policy classifier and code-based checks approve the output.

A real-life example

An e-commerce search assistant answers "Best phone under Rs 20,000 for gaming?" A discriminative intent model classifies this as a product-recommendation query. A learned ranking model scores 3,000 phones and returns the top 10 by predicted purchase likelihood and match. An LLM (generative) writes a short comparison of the top 3 using their specs from the catalogue. Finally, a checker verifies that every price and spec in the answer matches the catalogue, and a moderation classifier checks the text.

The team measures each part differently: the ranker by click-through and NDCG, the LLM by a weekly human review of 200 answers against a rubric, and the checker by how many bad answers it catches. When the LLM once wrote "under Rs 20,000" next to a phone priced Rs 21,499, the discriminative-style numeric check caught it before the shopper saw it.

Follow-up questions to expect

  • "Can generative AI replace discriminative models?" — For many low-volume tasks, prompting an LLM is good enough. For high volume, strict latency or audited decisions like credit scoring, dedicated discriminative models remain better.
  • "How do you evaluate a generative system?" — A fixed test set with rubrics per case, automated checks where possible (format, facts against sources), model-based judges calibrated against human ratings, and online metrics such as task success.
  • "Is a recommender system generative?" — Traditionally discriminative (it scores items), though newer systems also use generative models to produce item IDs or explanations.