When “Because” Isn't a Comfortable Answer

The uneasy shift from deterministic logic to probabilistic AI

“And you may ask yourself, well… how did I get here?”

Talking Heads, “Once in a Lifetime”

When I was a kid, I wanted to be like my dad. He was a very smart man, a mechanical engineer who made things, like rockets for NASA. Years later, when I was quoted in the press as saying he worked for NASA, he corrected me. He worked for a contractor to NASA. But like I said, I was a kid. As far as I was concerned, my dad supervised Mission Control.

So I, too, decided to be an engineer, although a chemical versus a mechanical one. I would get to use math, physics, and chemistry — all things I enjoyed learning about — and I would get paid for it! Why chemical engineering? In 1979, I read that its entry-level jobs paid more than those in other engineering disciplines. As long as I was going to get paid to do cool stuff, I might as well get paid well — simple, high school logic.

I liked how engineering could explain things, even put them into equations so that you could predict, optimize, and control the world around us. It is incredible how deterministic things can be; if you know the dew point of a gas and compress and cool it to the right conditions, it will condense very predictably, over and over. In my first job out of engineering school at Dow Chemical, I focused on writing code to automate plant operations. Here, I could integrate the deterministic patterns of materials, the ability to measure inputs and control operations (e.g., pumps and valves), and software to control results. I know this sounds dorky. And a lot of this was a grind. But it was also exhilarating, for me at least. What’s more, and I know this had to be great cocktail conversation, I could explain why!

Many years later, in 2017, having just left Google and pondering what I would do next, I got a call from Joe Averkamp, an old colleague and friend, who asked if I would be on a panel in Charlotte for a meeting of the IBTTA. I had to look it up — International Bridge, Tunnel and Turnpike Association. One of the many areas where Joe had developed expertise was in intelligent transportation.

Why me? The intelligent transportation industry wanted to discuss how they could utilize big data for applications like traffic management, and the fact that I had just left Google, not that I had worked with big data, was a draw. To be clear, I had rubbed shoulders with some of the talent at Google (I even met Demis Hassabis, a founder of DeepMind, a story I can share in another post). But I was clearly no big data guru.

The words machine learning and AI were starting to seep into conversations around big data. At the time, I had heard of neural networks but could not explain what they were. But I have never been scared to feel my way through how new technologies might impact the world, and less afraid to talk about it, even if I was not an expert. So I came into this panel with some ideas and opinions about how AI would apply.

So I shared, “As engineers, we like to think that we can predict and explain things. When your algorithm, which you wrote, instructs you to turn the entrance ramp stoplight green or adjust the fee for entering the express lane by 73 cents, you are accustomed to being able to point to that carefully constructed algorithm to explain 'why'. With AI, you will have a system that tells you to do just those types of things. But it may not tell you why. It will tell you just to do it. For things like traffic control, the question is, will we have the confidence to just do it??” I was anticipating a reckoning that engineers and humans will confront as AI advances. Now I better understand that what I was really brushing against was the shift from deterministic systems we can fully explain, to probabilistic ones we can't.

I did not use the words deterministic, probabilistic, or explainability directly in that conversation. But this morning as I was reading through some articles, I saw this thoughtful post from Gian Segato at Anthropic. It touches more broadly on how this shift in mindset could change how we engineer, and more broadly, how we problem solve.

This is going to be exhilarating, again.