What Is Generative AI? How It Works, Key Architectures & Why It Matters

How generative AI works

Learning from data

A generative model is trained on a large dataset — millions of text documents, billions of images, or vast collections of code. During training, the model learns the statistical relationships within the data: which words tend to follow which, how pixel values relate to each other in images, how musical notes sequence into melodies. The result is a mathematical representation — a probability distribution — that captures the ‘shape’ of the data.

Generating new content

When given a prompt (a text instruction, an image description, or a starting seed), the model samples from its learned distribution to produce a new output. For text, this means predicting the next most likely token in a sequence. For images, it means iteratively refining noise into a coherent picture. The output is new — it does not exist in the training data — but it follows the patterns the model has learned.

Generative vs. discriminative models

The key distinction: discriminative models learn to classify or label existing data (is this image a cat or a dog?), while generative models learn to produce new data (generate an image of a cat that looks realistic). Most of the AI systems that power everyday tools — spam filters, recommendation engines, image classifiers — are discriminative. Generative AI is the subset that creates.

Key architectures

Transformers (2017–present)

The Transformer architecture, introduced by Vaswani et al. at Google Brain in 2017, uses a mechanism called self-attention to process every element in a sequence in relation to every other element simultaneously. This replaced earlier recurrent architectures (RNNs, LSTMs) that processed sequences one element at a time. Transformers are the foundation of modern large language models (LLMs) — including GPT, Claude, Gemini, and Llama — and dominate text generation, code generation, and increasingly multimodal tasks [1].

Diffusion models (2020–present)

Diffusion models, formalized by Ho, Jain, and Abbeel in 2020, generate content through a two-phase process. During training, they systematically add random noise to data until it becomes pure noise. During generation, they learn to reverse this process — denoising step by step — to reconstruct or create new content from a text prompt. Diffusion models power most modern image and video generation systems, including Stable Diffusion, DALL-E, and Midjourney [2].

Generative Adversarial Networks (2014)

GANs, introduced by Goodfellow et al. in 2014, pit two neural networks against each other: a generator that creates fake data and a discriminator that tries to distinguish real data from fakes. As the discriminator improves, the generator must produce increasingly realistic outputs to fool it. GANs were the dominant image generation architecture before diffusion models and remain influential in specialized applications like face generation and style transfer [3].

Variational Autoencoders (2013)

VAEs, introduced by Kingma and Welling in 2013, compress data into a compact latent representation and then reconstruct it, with the ability to sample new points in that latent space to generate novel outputs. VAEs are foundational to generative modeling theory and remain important in applications requiring structured latent spaces, such as drug discovery and anomaly detection [4].

Multimodal models

Modern frontier models increasingly combine multiple modalities — processing and generating text, images, audio, and video within a single system. Models like GPT-4o, Gemini, and Claude can analyze images and generate text about them, or produce images from text descriptions. This multimodal capability is a defining trend of the foundation model era, moving beyond single-modality generation toward systems that operate across the full spectrum of human communication [9].

How generative models are shaped after training

RLHF and alignment

Raw pretrained models are powerful but unaligned — they predict likely next tokens without regard for helpfulness, safety, or human preferences. Reinforcement Learning from Human Feedback (RLHF), introduced by Christiano et al. in 2017 and scaled to language models by Ouyang et al. in InstructGPT (2022), shapes model behavior after pretraining. Human raters compare model outputs, a reward model learns from those preferences, and the base model is fine-tuned to maximize the learned reward. The key finding: a 1.3-billion-parameter model aligned via RLHF was preferred by human raters over the 175-billion-parameter unaligned GPT-3 [6].

Constitutional AI and DPO

Anthropic’s Constitutional AI (Bai et al., 2022) replaces human raters with AI-based self-critique against a set of written principles, enabling scalable alignment without human labeling bottlenecks. Direct Preference Optimization (DPO) simplifies RLHF by training directly on preference data without a separate reward model. These methods shape the difference between a raw next-token predictor and a helpful, harmless assistant [7].

Open-source vs. closed models

The generative AI landscape is split between open-weight models (Llama, Mistral, DeepSeek, Qwen) that anyone can download, fine-tune, and self-host, and closed models (GPT, Claude, Gemini) available only through API. As of 2026, open-weight models have closed the capability gap on many benchmarks — DeepSeek R1 matches or exceeds GPT-5 on math and reasoning — while closed models retain advantages in enterprise tooling, multimodal consistency, and managed infrastructure. This split matters for the agentic economy: open models enable self-hosted agents with full data control, while closed models offer faster time-to-production and stronger safety guarantees [8].

Why frontier models are so expensive

The cost of training generative models follows scaling laws — empirical relationships between compute, data, and model performance. Kaplan et al. (2020) established that model performance improves predictably with more parameters and data. Hoffmann et al. (2022, the ‘Chinchilla’ paper) refined this: for a fixed compute budget, models should be trained on roughly 20 tokens per parameter. This means the largest models require not just more parameters but proportionally more training data — driving costs into the hundreds of millions of dollars. The Chinchilla finding also showed that smaller models trained on more data can outperform much larger undertrained models, reshaping how the industry allocates compute [10].

Why generative AI matters for AI agents

Generative AI is the reasoning engine that makes autonomous AI agents possible. Interpreting goals, planning actions, generating tool calls, and producing responses all rely on generative models — though the tools themselves may execute deterministically. The transformer architecture that powers modern LLMs is the cognitive foundation of the agentic AI systems described across this library. When an agent negotiates a contract, writes code, or navigates a multi-step workflow, it is generating tokens one at a time, drawing on the same learned distributions that power a chatbot. The quality, reliability, and alignment of generative models directly determine how capable, trustworthy, and safe autonomous agents can be [11].