Skip to content
GenAI Learn/From AI Basics to Generative AI
Browsing as a guest. Sign in to save your progress and earn XP as you complete chapters.

What Makes Generative AI Different

4 min read

You'll learn to

  • -Explain why evaluation is much harder for generated text than for a predicted label
  • -See hallucination arise from fluent recombination in a model small enough to inspect
  • -Preview the engineering concerns Phase 2 covers: prompting, grounding, tools and production

Everything in Phase 1, such as loss functions, overfitting, evaluation metrics and bias, still applies to generative AI underneath. But generating content as a task brings genuinely new engineering challenges on top of that foundation. That is why this course gives it a dedicated phase.

An Analogy: Marking Multiple Choice vs. Marking Essays

A multiple-choice test has a key. Each answer is right or wrong and a machine can mark it. An essay has countless good answers, and marking one means judging quality, accuracy and style. A discriminative model is a multiple-choice test. A generative model writes the essay.

Evaluation Gets Much Harder

A discriminative model has a clean check: does the predicted label match the true one? A generative model that writes a paragraph has no single correct output to compare with. There are countless reasonable ways to answer the same question well. The last lines of the demo show the problem: two correct paraphrases score zero under exact matching. Phase 2 spends real time on this, because "how do you know if the output is good?" is one of the hardest open questions in the field.

The Interface Changes Entirely

Classical ML systems are usually built and used as an API that returns a structured value: a category, a score, a number. Generative systems are frequently built and used through natural language itself. You tell them what you want mostly through the prompt you write, not through a rigid, predefined input schema. That moves real engineering weight onto prompt design, a skill with little counterpart in classical ML.

New Failure Modes Emerge

Generative models bring failure modes that do not exist for discriminative ones. The best known is hallucination: the model produces fluent, confident text that is nevertheless wrong or entirely made up. A spam classifier cannot hallucinate in this sense. It can only misclassify. The demo is a model small enough to inspect. It has seen exactly two true facts. Every sentence it can produce is listed.

Press Run: from two true facts, the fluent model can also say two false ones

Hallucination is not a bug that will simply be patched away. It is a structural consequence of how these models generate text: predicting plausible content, not looking up verified facts. Understanding this is why techniques like retrieval-augmented generation exist. Phase 2 covers it in depth.

What Phase 2 Actually Covers

  • -How large language models process and generate text: tokenization, embeddings, the transformer architecture previewed earlier.
  • -How to get useful behavior out of them reliably: prompting, sampling controls, model selection.
  • -How to ground them in real, verifiable information: retrieval-augmented generation.
  • -How to let them take real-world actions: agents and tool use.
  • -How to run all of this reliably, safely and affordably at scale: production and LLMOps.

Interview Signal is part of Pro

See a real weak answer next to a real strong one for this exact topic.

Quiz is part of Pro

Test what you just read with a short quiz, and bank the XP.

Discussion0

Join the Discussion

Sign in to leave comments, reply to others, or like insights.

Sign In to ScaleDojo

No comments yet. Be the first to start the thread!