The Human-AI Divergence

Computers Are Poor Human Analogs

The Human-AI Divergence

As humans, we do a lot of learning by observation, but we are also capable of learning by experience. In fact, we are not very good at learning strictly from observation, and without the experience of physically manipulating real objects, what we are capable of learning is quite limited. Let’s call this “hands-on experience,” although it doesn’t necessarily involve only our hands.

I’m sure we’ve all had the experience of hearing someone explain how to perform some task—or even of watching a demonstration of how it’s done—and having the deep, profound sensation that we’ve grasped it. What happens next quite often is the discovery that we retained almost nothing from the instructions or demonstrations when we try doing the task ourselves. Research into human learning and education have described this phenomenon: the illusion of understanding that comes from lacking the hands-on experience.

Experiential learning is fundamentally different from how computer programs learn things. And how large-language models that are so often lumped together under the label “artificial intelligence” are fundamentally distinct from a lot of human learning. The LLM algorithms are “trained” (to use the industry jargon word) on massive quantities of data, and given the task of finding relationships between words based on statistical probabilities as to which words are most likely to happen next, given the previous sequence of words.

But computers have no means of exploring the world as humans do—much less as human babies and children do. They can’t touch and feel. They can’t explore and find out what things give emotional rewards. They can’t try and fail, either. As animal creatures, we do a lot of trial and error in the first person—although we can also learn by watching others try and fail.

These were the thoughts I had when I first heard this dialog between Dwarkesh and the computer scientist Richard Sutton. Dwarkesh is something of an AI (LLM) enthusiast, but Sutton is more of a skeptic. Their conversation particularly caught my attention for challenging the optimistic hype about the technology. And if you don’t have time for the whole discussion, you can skip to the section marker around 14 minutes in to get to the meat of the dispute.

The debate was very worthwhile, and Dwarkesh did a good job sticking to his position, while respecting Sutton’s points. I got the feeling that Dwarkesh missed a few details that might dawn on him sometime later. It was genuine debate, and not the political punch-up where each party tries to make the other cry “Uncle” and admit defeat. Which, of course, is exactly what prevents such a resolution, or any other, for that matter.