Driving Intelligence examines artificial intelligence through the lens of autonomous driving, using driving as a real-world test case for intelligence. At a time when expectations of artificial general intelligence are rising, the book challenges how intelligence is defined and measured, arguing that current benchmarks saturate, leak, and can be gamed.
The book develops a framework for understanding intelligence through the concept of real-world agency, with autonomous driving as its central case study. It argues that driving uniquely combines scale, open-ended environmental complexity, and continuous multi-agent interaction under uncertainty. Unlike most AI benchmarks, which are static, isolated, or susceptible to saturation and gaming, autonomous driving unfolds in a dynamic physical and social world that cannot be exhaustively specified in advance. It requires learning, abstraction, and transfer. The book uses this setting to examine what current AI systems can and cannot do in practice. It analyses failure modes in perception, prediction, planning, and interaction, showing how these limitations reveal a deeper gap between statistical learning from large-scale data and genuine agency. Across these analyses, the book connects technical constraints in autonomous driving to broader questions about the nature of intelligence and the adequacy of current AI paradigms.
The book’s central contribution is to reframe autonomous driving as a benchmark for intelligence grounded in real-world agency. It identifies systematic limitations in current AI approaches and clarifies what is missing for genuine autonomy. In doing so, it provides a concrete framework for evaluating progress in Real-World AI and Physical AI, offering both a diagnostic tool for researchers and a reference point for assessing future systems.