Generative AI sits inside a stack of older, broader ideas, and a leader who cannot keep those ideas straight will be sold the wrong product. This section walks down the stack from artificial intelligence at the top to specific model families at the bottom, then explains the two intervention surfaces a business actually touches: prompt engineering and prompt tuning. The framing matters because procurement contracts, vendor pitches, and internal build decisions all hinge on which level of the stack the problem actually lives at. Asking a foundation model to do what a classical machine learning model does is wasteful; asking a classical machine learning model to do what only a generative model can do is impossible.
The conceptual hierarchy is nested. Artificial intelligence is the outermost field. Machine learning is the dominant technique inside that field. Deep learning is the dominant family of machine learning techniques. Generative AI is a class of deep learning models specialised for producing new content. Foundation models, multimodal foundation models, diffusion models, and large language models are specific shapes inside generative AI. Most enterprise gen AI conversations are happening at the foundation-model level, even when the participants call it something else.
Artificial Intelligence
Artificial intelligence (AI) is the broad discipline of building systems that perform tasks that would otherwise require human intelligence: perception, reasoning, planning, learning, and language use. AI as a label is older than machine learning and covers a wide span of techniques, including rule-based expert systems (decision trees written by humans), search algorithms (the chess engines of the 1990s), symbolic reasoning, robotics, computer vision, and natural language processing. For a leader, the key point is that AI is a category, not a product. When a vendor says "our solution uses AI", the relevant follow-up question is which technique, on what data, against what benchmark.
The business framing of AI has tightened with the emergence of generative models. AI was historically positioned as automation of narrow tasks (fraud detection, image classification, recommendation ranking). Generative AI has widened that framing to include content production, conversation, and tool-using agents. Both framings remain valid; they sit on the same continuum and use overlapping infrastructure.
Machine Learning
Machine learning (ML) is the subfield of AI that builds systems by learning patterns from data rather than from rules a human wrote. Instead of an analyst encoding "flag the transaction if amount > 10,000 and country differs from card country", an ML system is shown millions of labelled transactions and learns the pattern itself. The output is a model: a set of parameters (weights) that, given a new input, produces a prediction.
Three learning paradigms cover almost every business case the exam tests:
- Supervised learning uses labelled data. Each training example carries the right answer, and the model learns the mapping from input to label. Spam detection, credit scoring, and image classification are supervised problems.
- Unsupervised learning uses unlabelled data. The model discovers structure without being told the right answer. Customer segmentation, anomaly detection in network traffic, and embedding-based search rely on unsupervised methods.
- Reinforcement learning uses a reward signal. An agent takes actions in an environment and learns the policy that maximises cumulative reward. Recommendation systems, robotics, and the alignment phase of large language models all use reinforcement learning.
Generative AI uses all three at different stages of a model's life. Foundation models are pretrained with a self-supervised objective (a flavour of unsupervised learning), then aligned with reinforcement learning from human feedback (RLHF), and then, optionally, fine-tuned with supervised examples for a specific task.
💡 Exam Trap: Supervised learning is not the same as labelled production data. A model can be supervised-trained once, then deployed and used on unlabelled production data. The labels are needed during training, not at inference.
Generative AI
Generative AI is the class of machine learning models whose primary output is new content rather than a classification or a numeric prediction. The content can be text, image, audio, video, code, structured data, or any combination of those. The defining characteristic is that the model produces a sample from a learned distribution, not a label drawn from a fixed list.
The business significance is that generative AI shifts what computers can do unsupervised. A classical ML system can decide whether an email is spam; a generative system can write the reply. A classical ML system can score the risk of a contract; a generative system can summarise the contract, draft a redline, and propose alternative clauses. That shift opens use cases (content creation, summarisation, conversation) that were previously dependent on human labour at every step.
Foundation Models
A foundation model is a large model pretrained on a broad dataset using self-supervised learning, designed to be adapted to many downstream tasks. The term was coined to capture the observation that one large pretrained model now serves as the starting point for many specific applications, instead of each application training its own model from scratch. Gemini, Claude, GPT, Llama, and Imagen are foundation models. So is Google's Chirp speech model and the Veo video model.
Foundation models matter to leaders for three reasons. First, they invert the cost curve: pretraining is expensive and centralised, while adaptation is cheap and decentralised. Second, they consolidate the supply chain: a small number of providers train the base models, and everyone else builds on top. Third, they change the procurement question from "which model do we train?" to "which model do we license, and how do we adapt it?".
Multimodal Foundation Models
A multimodal foundation model accepts more than one input modality (text, image, audio, video) and can also produce more than one output modality. Gemini is the canonical Google Cloud example: a single model that reads a PDF, a chart, a voice memo, and a video clip in the same prompt, and returns a written analysis with embedded image references.
Multimodality matters because business data is rarely a single modality. A field service ticket comes with a photo, a description, and a voice note. A clinical record has text, scanned images, and audio dictation. A retail listing has text, images, and sometimes a short video. A multimodal foundation model removes the engineering plumbing that older systems needed to glue together a text model, an image model, and an audio model.
📊 Limit: Gemini 2.5 models published by Google Cloud support a context window of up to 1,048,576 input tokens for Gemini 2.5 Pro and Gemini 2.5 Flash, with 65,536 output tokens; the exact ceilings are version-specific and listed in the model card for each release.
Diffusion Models
A diffusion model is a generative model architecture that learns to produce content by reversing a noise process. During training, the model is shown an image (or audio sample, or video frame) and progressively more corrupted versions of it. The model learns the denoising step: given a noisy sample, predict a slightly cleaner one. At inference, the model starts from pure noise and iterates the denoising step until it produces a coherent output.
Diffusion is the dominant architecture for image and video generation. Google's Imagen (image) and Veo (video) are diffusion-based. The architecture matters to leaders because diffusion models have different cost, latency, and safety characteristics from text-only LLMs: each generation is many forward passes through the model, so latency is higher and compute cost per output is higher than a single-pass LLM. Safety filtering (against deepfakes, copyright bleed-through, and disallowed content) is also more involved for image and video than for text.
Large Language Models
A large language model (LLM) is a foundation model trained on a very large text corpus to predict the next token in a sequence. "Large" is by parameter count (typically billions to trillions of parameters) and by training data volume (typically trillions of tokens). Gemini, PaLM 2 (legacy), and the Gemma open-source family are Google's LLMs.
For business framing, LLMs are useful precisely because the next-token-prediction objective, scaled up enough, produces a model that can summarise, translate, classify, extract, draft, and converse with no task-specific training. A single LLM endpoint replaces what used to be a stack of separate NLP services.
⚠️ Anti-Pattern: Treating an LLM as a database. LLMs do not store facts in a queryable way. They store statistical patterns about tokens. If the business answer must be authoritative (regulatory reporting, financial disclosure, clinical guidance), the LLM must be grounded against an external source of truth, not relied on as a memory.
Prompt Engineering
Prompt engineering is the practice of designing the input text given to a foundation model to make its output more useful, accurate, or safe. It is the cheapest, fastest, and most reversible adaptation lever, and it requires no access to model weights. A prompt engineer writes instructions ("You are a senior compliance analyst. Read the following policy. List every clause that conflicts with GDPR Article 6."), provides examples ("Here are three examples of correctly extracted clauses..."), and constrains the output format ("Return JSON with fields clause_id, risk_level, rationale.").
For business leaders, prompt engineering is the right starting tool for almost every gen AI initiative. It is fast, low-cost, and produces measurable lift before any platform spend. The risk is that a prompt designed by one team often does not transfer cleanly to another team's data; prompts are brittle and need version control and evaluation.
Prompt Tuning
Prompt tuning is a parameter-efficient adaptation technique that learns a small set of additional model parameters (a "soft prompt") on top of a frozen foundation model. The original model weights are not changed; only the prepended vector is learned from a small labelled dataset. The result is a customised model behaviour without the cost and risk of full fine-tuning.
Prompt tuning sits between prompt engineering (free, fast, brittle) and full fine-tuning (expensive, slow, durable). It is the right choice when prompt engineering hits a quality ceiling but the dataset is too small or sensitive to fine-tune the whole model. Vertex AI offers prompt tuning as a managed feature on selected models.
💡 Exam Trap: Prompt engineering changes the input to the model. Prompt tuning changes a small set of learned parameters that are prepended to every input. Fine-tuning changes the model weights themselves. The exam tests this distinction directly.
Architecture Overview
The relationships between these terms are clearest as a containment hierarchy.
Concept Hierarchy: From AI to Specific Model Families
🎯 Scenario: A retail bank wants to summarise customer call transcripts and extract complaint topics. The engineering team proposes training a custom classifier on six months of labelled calls. The product lead suggests prompting Gemini with a few examples and a strict output schema. The correct first move is prompt engineering. If quality is insufficient after a structured evaluation, prompt tuning on a small labelled set is the next step. Full fine-tuning is the last resort and is rarely justified for an extraction task that an LLM handles natively.
Decision Anchor
Choose prompt engineering when the task is well described in natural language, no proprietary data needs to be memorised by the model, and time to value is the constraint. Choose prompt tuning when prompt engineering plateaus, a small labelled dataset exists, and inference cost must stay close to the base model. Choose full fine-tuning when the task requires a persistent change in behaviour across many prompts, the labelled dataset is large enough (typically thousands to tens of thousands of examples), and the model card permits it.
💡 Exam Trap: Not every generative model is an LLM. Imagen and Veo are foundation models but not LLMs because their primary modality is not language. The exam phrases questions to test whether the candidate confuses "foundation model" with "LLM".
⚠️ Anti-Pattern: Buying a fine-tuned model when prompt engineering has not been measured. Fine-tuning a foundation model is a long-running commitment (custom artefact, custom evaluation, custom drift monitoring). Skipping prompt engineering and jumping straight to fine-tuning is one of the most expensive mistakes a gen AI programme can make.
The takeaway for a leader is to read every gen AI pitch through this stack. "AI platform" almost always means foundation models plus orchestration. "Custom model" almost always means prompt tuning or fine-tuning of a base foundation model, not a model trained from scratch. "Multimodal" means the model accepts more than one input type, and the cost and safety profile is different from a text-only LLM. Getting these labels right is the first defence against overspend.