AI in transportation research has gone through four eras: pre-DNN models, deep neural networks, large language models, and now AI agents. Each one changed what AI could do. Each one also changed the role of our theory. In the pre-DNN era, the analyst wrote the model: the variables, the equations, the form of the utility function. Data estimated a relatively small number of parameters. In the DNN era, the model learned its own features and relationships from data; the analyst chose the data, the architecture and what to predict. In the LLM era, one model pretrained on a vast body of text and code could take on many tasks; the analyst described the task in plain language, often with a few examples, and supplied the data. In the agentic AI era, AI plans and carries out multi-step work: it writes the code, runs the models and reports the results. The analyst sets the goal, reviews the steps and judges the answer. Each era also changed the role of our theory. In the pre-DNN era, it was the center of our work. Deep learning and language models pushed it to the side. Now it is coming back. The AI methods that work best build established theory into the model, such as discrete choice, network equilibrium and conservation of traffic flow, and use data to learn what the theory misses. They tend to hold up better than pure AI on rare events, shocks and new policies. Theory is both the structure of these models and the check on their answers. Theory gives them structure. It also lets us check their answers. Seven frontiers are worth watching. Six are specific to transportation: Generative world models for mobility systems. AI for behavioral, social, and multi-agent simulation. Multimodal foundation models for transportation. AI for decision, control, and large-scale optimization. Trustworthy and verifiable AI. AI for transportation science. One is general. 7. Recursive self-improvement. AI that improves its own code and training could speed up every frontier above. Three of the most innovative young scholars in the field will take them on and join me for a conversation. Featuring... Shenhao Wang, University of Florida, Lijun Sun, McGill University Sharon Di, Columbia University
