Generative AI and Foundation Models

Introduction and Context

Generative artificial intelligence (generative AI) refers to computational systems that learn patterns from data and use those patterns to produce new content, including text, images, audio, video, software code, molecular structures, and synthetic tabular data. Unlike conventional predictive machine-learning systems, which usually estimate a label, score, or numerical outcome, generative models attempt to represent aspects of the data-generating process itself.

Foundation models are large, generally pre-trained models that can be adapted to many downstream tasks. They are typically trained on broad and heterogeneous datasets using self-supervised objectives. Large language models (LLMs), vision-language models, speech models, and multimodal systems are important examples. Their generality comes from scale, transfer learning, and representations that encode statistical regularities across domains. However, generality should not be confused with truth, reasoning reliability, autonomy, or understanding.

Generative AI has become important in software engineering, scientific discovery, customer service, education, design, healthcare, finance, and public administration. Its adoption creates substantial opportunities but also raises questions concerning data provenance, copyright, privacy, bias, security, labor, energy consumption, accountability, and human oversight.

Learning Objectives

After completing this chapter, a student should be able to:

1. Explain the probabilistic foundations of generative modeling.

2. Compare autoregressive models, generative adversarial networks, variational autoencoders, and diffusion models.

3. Describe the architecture and training objectives of large language models.

4. Design effective prompts and explain embeddings and vector search.

5. Construct a retrieval-augmented generation pipeline.

6. Distinguish prompting, fine-tuning, parameter-efficient adaptation, and continued pre-training.

7. Explain multimodal learning and cross-modal alignment.

8. Evaluate generative systems for quality, factuality, safety, robustness, cost, and fairness.

9. Identify governance, ethical, legal, and operational risks.

1. Probabilistic Foundations of Generative Modeling

Let a dataset consist of observations x drawn from an unknown data distribution p_data(x). A generative model learns a parameterized distribution p_theta(x), where theta denotes trainable parameters. A common training objective is maximum likelihood:

L(theta) = sum over i of log p_theta(x_i).

Equivalently, minimizing the negative log-likelihood minimizes the cross-entropy between the empirical data distribution and the model distribution. This formulation is directly usable when p_theta(x) can be evaluated efficiently. In many modern models, likelihood is factorized or optimized indirectly.

For a sequence x = (x_1, x_2, ..., x_T), the chain rule gives:

p_theta(x) = product from t = 1 to T of p_theta(x_t | x_1, ..., x_{t-1}).

This identity is exact, not an approximation. It motivates autoregressive modeling: predict the next element conditioned on preceding elements. The elements may be words or tokens, image pixels, audio samples, or other ordered units.

A useful distinction is between density estimation and sampling. A model may assign high likelihood to data without producing particularly useful samples, and a model may generate visually convincing samples while lacking a tractable likelihood. Consequently, evaluation must match the application rather than rely on a single score.

2. Major Generative Model Families

2.1 Autoregressive Models

Autoregressive models generate one token or element at a time. For language, a decoder-only Transformer estimates the conditional probability of the next token given the preceding context. During training, teacher forcing supplies the true previous tokens. The loss for a sequence is commonly the average token cross-entropy:

L = -1/T sum from t = 1 to T of log p_theta(x_t | x_<t),

where x_<t denotes all tokens before position t. The logarithm is natural unless another base is specified. The loss is dimensionless because it is the logarithm of a probability.

At inference, the model produces a probability distribution for the next token. Greedy decoding selects the highest-probability token. Sampling introduces controlled variability. With temperature tau greater than zero, logits z_i may be transformed as:

p_i = exp(z_i/tau) divided by the sum over j of exp(z_j/tau).

Lower temperature concentrates probability on high-logit tokens; higher temperature produces more diversity. Temperature does not make a model more knowledgeable and can increase factual errors. Top-k sampling restricts sampling to the k highest-probability tokens, while nucleus or top-p sampling selects the smallest candidate set whose cumulative probability reaches p.

The Transformer uses self-attention to allow each position to interact with other positions. Given query, key, and value matrices Q, K, and V, scaled dot-product attention is:

Attention(Q,K,V) = softmax(QK^T divided by square root of d_k) V,

where d_k is the key-vector dimension. Scaling reduces excessively large dot products and helps stabilize gradients. Causal masking prevents a token from attending to future tokens during autoregressive training.

Strengths of autoregressive models include strong sequence generation, flexible conditioning, and a clear likelihood objective. Limitations include sequential generation latency, exposure bias between training and inference, context-window constraints, and the tendency to generate fluent but unsupported statements.

2.2 Generative Adversarial Networks

A GAN contains a generator G and discriminator D. The generator maps a random variable z, sampled from a simple prior such as a standard normal distribution, to a synthetic sample G(z). The discriminator estimates whether a sample is real or generated. The original minimax objective is:

min_G max_D V(D,G) = E_x~p_data[log D(x)] + E_z~p_z[log(1 - D(G(z)))].

Here D(x) is the discriminator's probability that x is real. The generator seeks to make generated samples indistinguishable from real samples, while the discriminator seeks to distinguish them. In practice, the non-saturating generator loss, -E_z[log D(G(z))], is often used because it provides stronger gradients early in training.

GANs can produce sharp images and have been influential in image synthesis, super-resolution, style transfer, and data augmentation. They are difficult to train because the two networks must remain appropriately balanced. Mode collapse occurs when the generator produces insufficiently diverse samples. Instability, sensitivity to architecture, and the absence of a simple likelihood also complicate evaluation.

2.3 Variational Autoencoders

A variational autoencoder (VAE) consists of an encoder q_phi(z|x) and decoder p_theta(x|z). The latent variable z is intended to capture a compact representation of the observation x. Since the exact posterior p_theta(z|x) is often intractable, the encoder approximates it.

The evidence lower bound (ELBO) for one observation is:

ELBO = E_q_phi(z|x)[log p_theta(x|z)] - KL(q_phi(z|x) || p(z)).

The first term is a reconstruction or data-fit term. The second term is the Kullback–Leibler divergence between the approximate posterior and the prior p(z). Maximizing the ELBO provides a lower bound on log p_theta(x). The KL divergence is non-negative and equals zero only when its two distributions are equal almost everywhere.

For a Gaussian encoder, the reparameterization trick expresses a latent sample as z = mu_phi(x) + sigma_phi(x) elementwise multiplied by epsilon, where epsilon follows a standard normal distribution. This permits gradient-based optimization through random sampling.

VAEs provide structured latent spaces, probabilistic inference, and stable training. Their samples may be less sharp than GAN outputs, particularly when the decoder uses a likelihood that averages over plausible details. They are useful for representation learning, anomaly detection, controllable generation, and scientific latent-variable modeling.

2.4 Diffusion Models

Diffusion models define a gradual forward noising process and learn to reverse it. In a common formulation, a clean sample x_0 is transformed through a sequence:

q(x_t|x_{t-1}) = Normal(sqrt(1-beta_t) x_{t-1}, beta_t I),

where beta_t lies between 0 and 1, I is the identity covariance matrix, and t indexes the noise step. Under a suitable noise schedule, x_T approaches an isotropic Gaussian distribution.

A neural network learns to estimate the noise or an equivalent denoising quantity. A widely used simplified objective is:

E_{t,x_0,epsilon}[||epsilon - epsilon_theta(x_t,t)||^2],

where epsilon is standard Gaussian noise and x_t is a noisy version of x_0. Generation begins with noise and applies learned reverse transitions to obtain a sample.

Diffusion models have achieved high-quality image, audio, and video generation. Conditioning may use text embeddings, class labels, spatial controls, or reference images. Classifier-free guidance combines conditional and unconditional predictions to strengthen conditioning, although excessive guidance can reduce diversity or introduce artifacts. Their principal disadvantage is computational cost: generation may require many denoising steps, although accelerated samplers and distillation reduce latency.

2.5 Comparative View

Autoregressive models are particularly effective for discrete sequences and flexible conditioning. GANs can generate sharp samples quickly after training but may suffer mode collapse. VAEs offer useful probabilistic latent spaces and relatively stable optimization, with a possible sharpness–likelihood trade-off. Diffusion models offer strong sample quality and controllability but often require substantial computation.

No family is universally superior. Selection depends on data modality, latency requirements, controllability, interpretability, training resources, and evaluation objectives.

3. Large Language Models and Foundation Models

3.1 Tokenization and Representation

Language models operate on tokens rather than necessarily on complete words. A tokenizer may represent frequent words as single tokens and rare words as subword units. Tokenization affects context length, computational cost, multilingual performance, and susceptibility to attacks involving unusual character sequences.

An embedding maps a discrete token or object to a vector in a continuous space. Learned positional information represents token order. Similar meanings may be represented by nearby vectors, although geometric similarity is dependent on training data and objective and should not be interpreted as a complete semantic theory.

3.2 Transformer Architecture

A decoder-only LLM generally contains repeated blocks of masked self-attention, feed-forward transformations, normalization, residual connections, and output projection. During pre-training, the model learns to predict missing or next tokens from large corpora. This self-supervised task supplies labels from the data itself.

Encoder-only Transformers are commonly used for representation and classification, whereas encoder–decoder architectures support conditional generation such as translation and summarization. Decoder-only systems are well suited to open-ended continuation and instruction following after suitable adaptation.

3.3 Scaling and Emergent Capabilities

Increasing parameters, data, and compute can improve performance, but gains depend on data quality, optimization, architecture, and evaluation. Apparent emergent capabilities may reflect nonlinear measurement thresholds, task composition, or prompting effects rather than a sudden change in internal cognition. A large model can still fail on simple tasks, reproduce biases, or provide fabricated citations.

A foundation model should therefore be understood as a reusable statistical component, not as an autonomous expert. Domain deployment requires evaluation on representative tasks, users, languages, and failure conditions.

4. Prompting, Embeddings, and Retrieval-Augmented Generation

4.1 Prompt Engineering

A prompt specifies the task, context, constraints, and desired output format. Effective prompts are explicit about the audience, source boundaries, units, evaluation criteria, and uncertainty requirements. Few-shot prompting provides examples; zero-shot prompting provides instructions without examples. Chain-of-thought-style internal reasoning may improve some tasks, but systems should generally expose concise answers and verifiable evidence rather than treating generated reasoning as guaranteed proof.

A robust prompt may include:

• role and task definition;
• authoritative context;
• input delimiters;
• output schema;
• constraints and prohibited assumptions;
• a request to identify uncertainty;
• validation instructions.

Prompt injection occurs when untrusted content attempts to override system or application instructions. Delimiting retrieved documents, applying privilege separation, validating outputs, and limiting tool permissions are important defenses.

4.2 Embeddings and Vector Search

An embedding model maps an object x to a vector f(x). Similarity can be measured using cosine similarity:

cosine(u,v) = (u dot v) divided by (||u|| ||v||),

for nonzero vectors u and v. The result lies between -1 and 1. Cosine similarity is often useful when vector direction is more important than magnitude, but its suitability must be empirically tested.

In a retrieval system, documents are divided into chunks, embedded, indexed, and retrieved in response to an embedded query. Chunk size must balance semantic completeness against retrieval precision. Metadata filters, hybrid lexical–semantic search, reranking, and deduplication frequently improve results.

4.3 Retrieval-Augmented Generation

Retrieval-augmented generation (RAG) combines an information-retrieval system with a generator. A typical pipeline is:

1. Ingest and clean documents.

2. Segment documents while preserving titles, tables, and provenance.

3. Compute embeddings and build an index.

4. Retrieve candidate passages for a query.

5. Rerank or filter passages.

6. Place selected evidence into a controlled prompt.

7. Generate an answer with citations or source references.

8. Evaluate retrieval and generation separately.

RAG reduces dependence on model parametric memory and can incorporate current or private information without updating all model parameters. It does not guarantee factuality: retrieval may be incomplete, documents may be wrong, and the generator may ignore or distort evidence.

Evaluation should distinguish retrieval recall, precision, ranking quality, answer faithfulness, answer relevance, citation correctness, latency, and cost. A system that produces fluent answers from irrelevant passages is not a successful RAG system.

5. Adaptation and Fine-Tuning

Prompting changes inputs but not model parameters. Fine-tuning updates parameters using task- or domain-specific data. Supervised fine-tuning uses input–output examples. Continued pre-training exposes a model to additional domain text using its original self-supervised objective. Preference optimization uses judgments or preference pairs to encourage desired behavior, but preferences may encode annotator and institutional biases.

Parameter-efficient fine-tuning (PEFT) updates a small number of additional parameters or low-rank transformations. Low-rank adaptation (LoRA), for example, represents a weight update as a product of two lower-rank matrices rather than updating the full weight matrix. This reduces memory and storage requirements, though the appropriate rank and training data remain application-dependent.

Fine-tuning is appropriate when consistent behavior, terminology, style, or task performance is required and representative labeled data exist. RAG is usually preferable when the central problem is access to changing factual content. Fine-tuning should not be used as a substitute for correcting unsafe data, and it does not reliably eliminate hallucinations.

A sound adaptation process includes data governance, train–validation–test separation, leakage checks, baseline comparison, early stopping or regularization where appropriate, robustness testing, and post-deployment monitoring. Evaluation must test both desired capabilities and regressions in general behavior.

6. Multimodal Learning

Multimodal models process two or more modalities, such as text and images. Fusion may occur early, by combining inputs; late, by combining modality-specific predictions; or through cross-attention, in which one modality attends to representations of another. Contrastive learning often trains paired examples so that matching text and image embeddings are close while nonmatching pairs are separated.

A multimodal system may answer questions about an image, generate images from text, summarize video, or use speech as an interface. Challenges include alignment errors, missing modality information, OCR mistakes, visual hallucination, temporal reasoning, accessibility, and unequal performance across demographic or environmental conditions. Modality-specific evaluation is essential; a fluent text response does not demonstrate accurate visual perception.

7. Evaluation and Reliability Engineering

Generative quality is multidimensional. Useful evaluation categories include:

• Fidelity: Does the output resemble the target distribution or satisfy the task?
• Diversity: Does the system avoid producing only a narrow set of outputs?
• Factuality and faithfulness: Are claims supported by evidence?
• Safety: Does the output avoid harmful, illegal, or discriminatory content?
• Robustness: Does performance persist under perturbations and adversarial inputs?
• Utility: Does the output improve a real workflow?
• Efficiency: What are latency, memory, energy, and monetary costs?
• Fairness: Are error rates and benefits acceptably distributed across groups?

Perplexity is derived from average next-token log loss and can be expressed as exp(L) when L is the average negative natural-log likelihood. Lower perplexity indicates better predictive performance on the evaluation distribution, but it does not directly measure factuality, usefulness, or safety. Human evaluation should use clear rubrics and multiple raters where feasible. Automated evaluators can scale assessment but may share the same biases as the systems they judge.

Reliability engineering should include red-team testing, adversarial evaluation, incident logging, fallback behavior, confidence or uncertainty communication, access controls, and human review for consequential decisions. Generated content should not be treated as verified merely because it is grammatically polished.

8. Ethical, Legal, and Governance Considerations

Training data may contain personal information, copyrighted works, confidential material, or socially biased representations. Organizations must establish lawful data acquisition, retention, access, and deletion practices. Privacy-preserving techniques can reduce risk but do not automatically guarantee privacy. Model inversion, membership inference, prompt leakage, and insecure tool use are relevant threat classes.

Copyright and licensing questions vary by jurisdiction and use case. Organizations should document data sources, model licenses, output policies, and responsibilities for downstream users. Provenance mechanisms, content credentials, watermarking, and audit logs can improve accountability, although none is a complete solution.

High-impact uses require stronger controls. A model assisting a clinician, lender, employer, or public authority should be subject to domain validation, human accountability, traceable evidence, and procedures for contesting decisions. Automation bias can cause users to accept incorrect outputs, while excessive restrictions can prevent beneficial access. Governance should therefore be risk-based rather than based solely on whether a system is labeled “AI.”

Environmental costs arise from training and inference energy, hardware production, and data-center cooling. Model selection should consider whether a smaller model, caching, quantization, retrieval, or limited context can meet requirements. Responsible deployment includes accessibility, multilingual support, clear disclosure, and monitoring for changing data distributions.

9. Illustrative Case Study: RAG for Engineering Maintenance

Background

A large manufacturing organization maintains thousands of pages of equipment manuals, maintenance bulletins, safety procedures, and historical work orders. Technicians currently search separate repositories, causing delays and inconsistent procedures.

Problem

The organization proposes a conversational assistant that answers maintenance questions. An unsupported answer could damage equipment or endanger personnel, so fluent generation alone is unacceptable.

Analysis

A general LLM has broad language ability but may not contain the organization's current manuals. Fine-tuning on all documents would be expensive, could memorize sensitive information, and would not automatically reflect revised procedures. RAG is therefore selected for document access, combined with deterministic safety rules.

Approach and Solution

Documents are versioned and classified by equipment type, date, and approval status. Tables and warning boxes are preserved during extraction. Chunks include document identifiers and page references. A hybrid search retrieves candidates using both lexical terms, important for part numbers, and embeddings, useful for paraphrased questions. A reranker prioritizes approved and current documents.

The generation prompt requires the assistant to answer only from retrieved evidence, cite document and page identifiers, state when evidence is insufficient, and direct technicians to emergency procedures when specified hazards appear. The application prevents the model from directly issuing commands to machinery. A human technician remains responsible for execution.

Evaluation uses a test set of authentic questions annotated for relevant documents, required safety warnings, and acceptable answers. Metrics include retrieval recall, citation accuracy, unsupported-claim rate, response latency, and escalation compliance. Red-team tests include obsolete manuals, malicious text inserted into documents, ambiguous part numbers, and questions outside the knowledge base.

Outcome

In this illustrative case, the system improves document discovery and reduces unsupported answers by refusing questions when approved evidence is absent. It does not replace maintenance expertise. Its success depends as much on document governance, access control, interface design, and evaluation as on the language model.

Learning Points

1. The best architecture is determined by information requirements, not model fashion.

2. Retrieval quality and document lifecycle management are central to answer quality.

3. Citations are useful only when they are correct and traceable.

4. Safety-critical systems require constrained actions and human accountability.

5. Deployment evaluation must measure workflow outcomes, not only linguistic quality.

10. Best Practices for Professional Development

A professional generative-AI project should begin with a clearly defined user problem and risk classification. Establish a baseline using non-generative methods where appropriate. Select the smallest model that satisfies quality and latency requirements. Separate trusted instructions from untrusted data, minimize privileges, and validate structured outputs. Maintain versioned datasets, prompts, models, indexes, and evaluation sets. Test representative edge cases and monitor production drift. Provide users with clear limitations and escalation paths. Finally, conduct periodic review involving technical, legal, security, domain, and affected-user perspectives.

Chapter Summary

Generative models learn distributions or conditional distributions from data and generate new content. Autoregressive models factorize sequences and predict one element at a time; GANs learn through competition between generators and discriminators; VAEs optimize a reconstruction-and-regularization lower bound; and diffusion models learn to reverse a progressive noising process. LLMs use Transformer architectures and self-supervised token prediction to form reusable foundation models.

Prompting controls model inputs, embeddings support semantic comparison, and RAG connects generation to external evidence. Fine-tuning changes model behavior through additional training, while parameter-efficient methods reduce adaptation cost. Multimodal systems align information across text, images, audio, and other modalities. Reliable deployment requires evaluation of factuality, faithfulness, safety, robustness, fairness, efficiency, and real-world utility. Governance must address privacy, copyright, bias, security, accountability, environmental cost, and human oversight. Generative AI is most effective when integrated into carefully engineered workflows rather than treated as an infallible autonomous expert.

Key Terms

Foundation model: A broadly pre-trained model adaptable to multiple tasks.

Generative model: A model that learns to produce new samples or outputs.

Autoregressive model: A model that predicts each element conditioned on preceding elements.

Transformer: A neural architecture based substantially on attention mechanisms.

Self-attention: A mechanism that computes interactions among elements within a sequence.

Tokenization: Conversion of text or other inputs into model-processing units.

GAN: A generative adversarial network containing a generator and discriminator.

Mode collapse: A GAN failure in which generated samples lack diversity.

VAE: A variational autoencoder with probabilistic encoder and decoder components.

ELBO: Evidence lower bound used to train variational latent-variable models.

Diffusion model: A model trained to reverse a gradual noising process.

Embedding: A vector representation of an object or token.

Cosine similarity: A similarity measure based on the angle between two nonzero vectors.

RAG: Retrieval-augmented generation using external retrieved evidence.

Prompt injection: An attempt by input content to override intended instructions.

Fine-tuning: Additional parameter training for a task or domain.

PEFT: Parameter-efficient fine-tuning that updates a limited parameter set.

Multimodal learning: Learning across two or more data modalities.

Hallucination: An unsupported or fabricated model output.

Model drift: Degradation caused by changing data, users, or operating conditions.

Review Questions

Short Answer Questions

1. What is the principal difference between a predictive classifier and a generative model?

2. Why is causal masking used in decoder-only language models?

3. What is mode collapse in a GAN?

4. How does RAG differ from fine-tuning for incorporating current organizational information?

5. Why does lower perplexity not necessarily imply greater factuality?

Long Answer Questions

1. Compare autoregressive models, GANs, VAEs, and diffusion models in terms of objective, strengths, limitations, and suitable applications.

2. Explain the Transformer self-attention mechanism and its role in large language models.

3. Derive the conceptual components of the VAE ELBO and explain the role of the KL-divergence term.

4. Describe a complete RAG pipeline, including document preparation, retrieval, generation, and evaluation.

5. Discuss the ethical, legal, security, and governance challenges of deploying foundation models in high-impact domains.

Analytical / Application Questions

1. An organization needs an assistant that answers questions from frequently changing regulatory documents and must cite sources. Should it primarily use prompting, fine-tuning, RAG, or a combination? Justify the architecture and evaluation metrics.

2. A diffusion image system produces highly detailed but nearly identical images for different prompts. Analyze possible causes and propose technical and evaluation-based remedies.

3. Design a risk-control plan for a multimodal maintenance assistant that interprets photographs and manuals but must never directly control industrial machinery. Include data, model, interface, human-oversight, and monitoring controls.

Multiple Choice Questions with Answers

Question 1

Which objective is directly associated with a decoder-only autoregressive language model?

A. Predicting the next token conditioned on preceding tokens

B. Classifying every input into a fixed set of image labels

C. Maximizing discriminator accuracy without a generator

D. Adding Gaussian noise without learning a reverse process

Correct Answer: A. Predicting the next token conditioned on preceding tokens

Question 2

What is the primary purpose of causal masking in a decoder-only Transformer?

A. To reduce the vocabulary size

B. To prevent access to future tokens during prediction

C. To force all tokens to have identical embeddings

D. To replace the attention mechanism with convolution

Correct Answer: B. To prevent access to future tokens during prediction

Question 3

Which problem is especially associated with GAN training?

A. Mode collapse

B. Tokenization overflow only

C. Inability to sample any output

D. Mandatory use of labeled classes

Correct Answer: A. Mode collapse

Question 4

In a VAE ELBO, what does the KL-divergence term generally encourage?

A. The approximate posterior to remain related to the chosen prior

B. The discriminator to classify real samples correctly

C. The diffusion schedule to use zero noise

D. The tokenizer to create longer words

Correct Answer: A. The approximate posterior to remain related to the chosen prior

Question 5

What is the usual direction of the generation process in a diffusion model?

A. From clean data to progressively noisier data only

B. From random noise through learned denoising steps

C. From labels to discriminator decisions

D. From embeddings to a fixed database index without decoding

Correct Answer: B. From random noise through learned denoising steps

Question 6

Which statement best describes retrieval-augmented generation?

A. It updates every model parameter whenever a user asks a question

B. It retrieves external information and supplies it to a generator as context

C. It removes the need for document quality controls

D. It guarantees that every generated claim is true

Correct Answer: B. It retrieves external information and supplies it to a generator as context

Question 7

For two nonzero vectors, cosine similarity is based primarily on their:

A. Vector angle

B. Number of database records

C. Token frequency alone

D. Training duration

Correct Answer: A. Vector angle

Question 8

Which approach changes model parameters using task- or domain-specific examples?

A. Fine-tuning

B. Greedy decoding only

C. Cosine normalization only

D. Document chunking without training

Correct Answer: A. Fine-tuning

Question 9

Why can lower perplexity fail to demonstrate factual reliability?

A. Perplexity measures predictive likelihood, not necessarily evidence support or truth

B. Perplexity is calculated only for images

C. Lower perplexity always means more hallucinations

D. Perplexity does not use model probabilities

Correct Answer: A. Perplexity measures predictive likelihood, not necessarily evidence support or truth

Question 10

Which control is most appropriate for a safety-critical maintenance assistant?

A. Allowing unrestricted direct control of machinery

B. Removing citations to make responses shorter

C. Using approved evidence, escalation rules, and human responsibility

D. Increasing temperature so the model explores more answers

Correct Answer: C. Using approved evidence, escalation rules, and human responsibility