Visualizing Generative AI

Preface

Generative AI has captured the world’s imagination like no other technology. Within months of ChatGPT’s release in November 2022, everyone, from CEOs to college students, was experimenting with AI-powered tools that could write, code, create images, and solve complex problems. The enthusiasm is infectious—and justified. But as the initial excitement settles, a sobering reality emerges: building production-ready applications with generative AI is fundamentally different from using ChatGPT for quick tasks.

We wrote this book because we’ve seen countless teams struggle to bridge the gap between the promise of generative AI and the reality of deploying reliable, responsible applications at scale. The technology is powerful, but it’s also nondeterministic and prone to hallucination, and it requires careful architectural decisions to balance creativity, risk, cost, and latency. Traditional software engineering practices need to evolve to accommodate these unique challenges.

What Makes GenAI Applications Different

Generative AI applications aren’t just traditional software with an AI component bolted on. They require a fundamentally different approach to development, testing, and deployment. Unlike deterministic software, GenAI applications produce different outputs for the same inputs. Unlike traditional AI models trained for specific tasks, foundation models are general-purpose tools that can be applied to countless problems—but also carry new risks around bias, toxicity, ...