The traffic light that would not change

I was ten, in New York City, and a traffic light was holding a red for an empty road.

The question I had was not sophisticated and I want to keep it that way, because the unsophisticated version is the one that turned out to matter: why is this thing not looking? It has a street in front of it. There is information everywhere. And it is running a timetable somebody wrote before any of us arrived.

I did nothing with the question. I was ten. It went wherever questions go when you have no means of acting on them.

Four years

In October 2025 we were driving back toward Ahmedabad after a trip through Rajasthan, and we were in a minor accident. Nobody was badly hurt.

Here is where the tidy version of this story would arrive, and I am not going to write it. The tidy version says: a traffic system failed and I resolved to fix traffic systems. That would be a good story and I have no evidence for it. I cannot tell you a signal caused that accident. I cannot tell you a better one would have prevented it. I have never seen the intersection data and I would not be able to interpret it if I had.

What actually happened is smaller and, I think, more honest: something I had been able to treat as an abstraction stopped being one. The question came back with urgency attached that it had not earned on the evidence.

I have thought about whether that is a good reason to build something. I have decided it is a fine reason to start and a terrible reason to stop checking your work. Motivation and justification are different objects. Mine was motivated by an accident and has to be justified by everything else.

What the question was actually about

The traffic version is the easy one. Fixed-cycle signal timing is not stupid — it is predictable, auditable and very hard to make worse, which are real virtues when the failure mode involves cars. Adaptive timing is not obviously better; it is better on average and worse in ways that are harder to anticipate. Anyone who tells you the answer is simply "make it adaptive" has not thought about who explains the pile-up.

The general version is what kept me. A signal that reads the street is a system that has been granted authority to act on a changing environment. Once you grant that, you own three problems at once: it has to act well, it has to be able to explain what it did, and someone has to be able to overrule it. Most systems manage two.

It took me another year to work out that the third problem is the hard one, and that the answer is not a better model but a structure the model sits inside. That is what VECTOR turned into.

Why I am writing this down now

Because in ten years I will have a cleaner account of why I started, and it will be partly false. Memory does that. It removes the coincidence and the four-year gap and the accident that proved nothing, and leaves a narrative where a child saw a problem and set out to solve it.

So: I saw a traffic light. I forgot about it. Something frightening happened that was probably unrelated. I remembered. That is the actual sequence, and I would like it on the record before I improve it.