Transformers: The Definitive Guide

Chapter 12. Where to Go Next: From Models to Intelligent Systems

This is the last chapter of the book, and by now you’ve learned how diverse and adaptable the transformer architecture is. You started from first principles and saw how a single architecture can power models for time series, vision, image and video generation, audio, reinforcement learning, planning, and coding. You also saw how test-time compute, such as tree search or reranking, and reinforcement learning can enhance performance and steer behavior. And you saw how agentic systems create loops of planning, action, feedback, and self-correction.

This was never meant to be a collection of isolated techniques. It was meant to show that transformers have become a general interface between data, decisions, and tools. Treat this book as your foundation to move from individual models to larger systems of specialist components working together to solve real tasks. By now, you understand the architecture, the patterns, and the reasoning that make transformers work across domains. Your next step is not learning another model or chasing another benchmark. The next step is learning how to wire these components into something coherent, reliable, maintainable, and capable of improvement through experience.

The shift happening in the field reflects this. Progress is no longer defined by a single checkpoint or a bigger context window. It’s defined by systems that integrate multiple models, allocate compute at test-time, coordinate ...