Searching for Freight’s Move 37

Go is considered the most complex strategy game, even more so than chess. Success depends less on calculating every possibility than on strategic intuition honed over years of experience.

In 2016, Lee Sedol, one of the greatest Go players in history, competed in a five-match series against AlphaGo, an AI-powered player built by Google DeepMind. Conventional wisdom held that no machine could compete with the world’s best players for at least another decade. Sedol himself predicted an easy victory.

Then came Move 37. Early in the second game, AlphaGo placed a stone so unusual that commentators assumed it was a glitch. Sedol left the room for nearly 15 minutes before returning, visibly shaken.

AlphaGo’s Move 37 had a 1-in-10,000 chance of being played by a human. It wasn’t copied from human play or drawn from any database of historical games. It emerged after AlphaGo had played millions of games against itself, uncovering strategies humans hadn’t considered before. It went on to beat Sedol four out of the five games.

Today, most AI applications in logistics and freight transportation focus on automating workflows: performing existing tasks faster, cheaper and with less human input. Sorting packages more efficiently. Predicting delays more accurately. Processing unstructured emails in seconds.

Those are valuable improvements. But Move 37 illustrates that AI can do more than automate work. It can solve problems in ways humans haven’t thought of at all.

AI as a Discovery Engine

Of course, AI is not just one thing. It encompasses a wide range of methodologies and techniques. The tools most people know today are large language models (LLMs) like Claude, ChatGPT and Gemini. Reinforcement learning (RL), the methodology behind AlphaGo, is fundamentally different.

LLMs learn from enormous collections of human-created text. They excel at recognizing patterns in what people have already written, said or done.

RL agents take another approach. They interact with an environment, take actions and receive rewards or penalties. Through trial and error, they gradually discover strategies that maximize future outcomes.

This is happening in fields like protein folding and semiconductor design. Researchers are using RL models to produce solutions not obvious even to experts.

In cybersecurity, Anthropic’s Claude Mythos model has demonstrated the ability to identify software vulnerabilities at superhuman speed, uncovering flaws that had escaped decades of traditional security research. As writer Michael Morgenstern observed, Mythos represents a “Move 37” moment for computer security.

In each case, the pattern is similar. Freed from long-held assumptions and conventional wisdom, AI ventures into unexplored territory and returns with something genuinely new.

Researchers at the Massachusetts Institute of Technology’s (MIT) Center for Transportation & Logistics are beginning to apply RL to transportation. Two recent examples:

1) A master’s thesis, “Deep Reinforcement Learning for Tactical Vessel Planning in Large-Scale Maritime Logistics Networks” by Cuya Nizama and Eduardo Andre, used deep reinforcement learning for vessel planning in large-scale maritime networks. Given the complexity of tramp-ship routing and scheduling under spot-market uncertainty, the model produced high-quality decisions in milliseconds, and its performance held up as fleet size and cargo volume grew.

2) A doctoral dissertation, “From Zero-Sum to Threat-Adaptive Attacks: A Reinforcement Learning Methodology for Network Interdiction,” by Matthew Webb, explored how RL can identify non-intuitive ways to disrupt logistics networks. One surprising result: Agents on irregular networks ignored global structural bottlenecks and instead devoted their effort around a single target, a non-intuitive strategy that no human analyst had proposed.

The Untapped Potential of RL

Freight transportation is ripe for an RL approach. Supply chains are extraordinarily complex and hard to predict, with shippers, carriers, intermediaries, contracts, regulations, weather, capacity constraints and shifting demand all interacting in ways that are only partially understood.

For decades, the industry has made real progress with better forecasting models, routing algorithms and procurement tools — the equivalent of a Go master studying thousands of recorded games and applying those lessons with greater discipline and speed. But these applications still operate within the established vocabulary of how freight has always been managed. Our thinking is constrained by our own experience.

What would happen if an RL agent were given only the rules of the freight market and a reward function based on overall system performance? After simulating millions of procurement competitions against itself, what contract structures might it discover?

What carrier relationship models might emerge? What load-matching logic would it develop that humans haven’t tried — not because it’s wrong, but because it looks wrong from within the limits of our current assumptions?

Every freight network has its own Move 37 waiting to be discovered.

The industry is right to invest in AI for automation. The efficiency gains are real, and competitive pressures demand them. But the principal issue now is not how AI can help us do today’s work better. Rather, it’s what the freight system might look like if AI showed us approaches we never thought to try.

(Photo credit: Getty Images/FPM)