What Makes Generative AI Different

Everything in Phase 1, loss functions, overfitting, evaluation metrics, bias, still applies to generative AI underneath. But generation as a task introduces genuinely new engineering challenges on top of that foundation, which is exactly why this course treats it as its own dedicated phase rather than one more chapter of Phase 1.

A discriminative model has a comparatively clean way to check correctness: does the predicted label match the true label? A generative model producing a paragraph of text has no single correct output to compare against. There are countless different reasonable ways to answer the same question well. Phase 2 dedicates real attention to this problem, since "how do you know if the output is actually good" turns out to be one of the hardest open questions in the entire field.

Classical ML systems are typically built and consumed as an API returning a structured value: a category, a score, a number. Generative AI systems are frequently built and consumed through natural language itself. You communicate what you want largely through the prompt you write, not through a rigid, predefined input schema. This shifts real engineering weight onto prompt design, a skill Phase 2 covers extensively, that has comparatively little equivalent in classical ML.

Generative models introduce failure modes that simply do not exist for discriminative ones, most notably hallucination, where a model generates fluent, confident-sounding text that is nevertheless factually wrong or entirely fabricated. A discriminative spam classifier cannot hallucinate a wrong answer in this specific sense. It can only misclassify. A generative model can produce something that sounds entirely correct while being completely invented.

• -How large language models actually process and generate text: tokenization, embeddings, the transformer architecture previewed earlier.
• -How to reliably get useful behavior out of them: 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 real scale: production and LLMOps.