Gen AI Leader
Deep Dive
The basic vocabulary the GENAI-L exam assumes you already master: how artificial intelligence, machine learning, natural language processing, and generative AI relate to each other, what a foundation model is, and a classic exam distinction — prompt tuning versus prompt engineering.
Contents
These three terms are often used as synonyms, but they form a hierarchy of nested concepts. The artificial intelligence (AI) is the broadest field: any technique that allows a computer system to mimic human cognitive abilities, such as reasoning, learning, or decision-making.
Machine learning (ML) is a subfield of AI: instead of programming explicit rules, the system learns patterns from training data and uses them to make predictions or classifications on new data it hasn't seen before.
Natural language processing (NLP) is, in turn, the application of AI/ML techniques specifically to human language: tasks like machine translation, sentiment analysis, text classification, or entity extraction are classic examples of NLP, many of them predating the explosion of generative AI.
How they nest
AI ⊃ Machine Learning ⊃ Deep Learning ⊃ Foundation Models ⊃ Generative AI. Each layer is a more specific subset of the previous one: every generative model is a deep learning model, every deep learning model is a machine learning model, and all machine learning is a form of AI — but not all AI uses machine learning (for example, a fixed "if-then" rule-based system is AI but not ML).
Generative AI (gen AI) is a branch of machine learning focused on creating new, original content — text, images, audio, video, or code — rather than only classifying, predicting, or detecting patterns in existing data. The key difference from "traditional" ML is that the output is generated content, not a label or a number.
A foundation model is a large-scale model, pre-trained on huge volumes of data, designed to be general enough to adapt to a wide variety of downstream tasks (text generation, summarization, classification, code generation, etc.) without needing to be trained from scratch for each one.
Multimodal foundation models
A multimodal foundation model can process and/or generate more than one type of data at once — for example, receiving an image and text together, and responding with text — instead of being limited to a single input or output modality. Gemini is Google's flagship example in this category.
Diffusion models
Diffusion models are an architecture used mainly for image and video generation: they learn to gradually reverse a "noise" process added to a piece of training data, until they produce new, coherent content from random noise guided by a prompt. Google's Imagen and Veo are based on this approach.
A large language model (LLM) is a specific type of foundation model specialized in text: trained on massive volumes of text to predict, token by token, the most likely continuation of a given sequence. Every LLM is a foundation model, but not every foundation model is an LLM — Imagen, for example, is a foundation model specialized in images, not an LLM.
LLM ↔ foundation model relationship
Think of "foundation model" as the general category (any modality) and "LLM" as the subset focused exclusively on language. Gemini is a multimodal foundation model that includes LLM capabilities, in addition to being able to process image, audio, and video.
Under the hood, an LLM is a deep neural network: data flows through layers of neurons that learn increasingly abstract representations, from individual tokens to meaning. The diagram below shows that flow (forward pass) in a generic neural network:
This distinction appears frequently on the exam because the two terms sound alike but describe technically very different processes.
Prompt engineering
This is the manual, iterative design of the prompt's text — how you word the instructions, which examples you include, in what order you structure the information — to get better responses from an already-trained model. It involves no additional training or change to the model's parameters: it's purely work on the input text. Techniques like few-shot prompting or chain-of-thought prompting are forms of prompt engineering.
Prompt tuning
This is a technical, automated lightweight fine-tuning technique: instead of editing the prompt text by hand, a small set of "prompt embeddings" (additional vectors, not readable as text) is trained and prepended to the model's input, using an example dataset. It does require a training process, though much cheaper and faster than retraining the entire model.
The typical exam trap
If the question describes someone rewriting or reordering the text of their instructions with no training process involved, it's prompt engineering. If it describes a process that "trains" or "tunes" additional vectors using an example dataset, without touching the base model's weights, it's prompt tuning —a lightweight form of fine-tuning.
The exam expects you to distinguish the three main learning paradigms, regardless of whether generative AI is used or not.
| Approach | How it learns | Example |
|---|---|---|
| Supervised | With labeled data: each training example includes the expected correct answer. | Classifying emails as spam/not spam based on thousands of already-labeled emails. |
| Unsupervised | With unlabeled data: the model discovers patterns or groupings on its own, without predefined answers. | Grouping customers into segments with similar purchasing behavior without predefined categories. |
| Reinforcement | Through trial and error: an "agent" receives rewards or penalties based on its actions, and adjusts its behavior to maximize cumulative reward. | Tuning an LLM's responses with RLHF (reinforcement learning from human feedback) based on human preferences. |
Did you get this topic?
Test what you just learned
A product team wants to improve the responses of a corporate chatbot by training a small set of additional vectors that get prepended to each input, using a dataset of example conversations, without manually rewriting the instruction text or modifying the base model's weights. Which technique does this process best describe?
Put what you learned to the test with original Gen AI Leader exam questions.
Start Gen AI Leader mock examUnlimited mock exams with the Pro plan
Log in to track your progress.