The artificially intelligent dog

Suppose you woke up one morning and your family dog was communicating in English. You could say, “Tell me when you’re ready to go out for a walk.” And instead of scratching at the door, your dog would say, “I would like to go outside now.”

I think of the Large Language Models as primarily a way to interface with a computer using English. Recall my essay.1

In my thought experiment, the dog does not have a big brain, capable of conceptual understanding . It still has only some basic instincts and a few learned behaviors. It just happens to be able to express those characteristics in English.

You would not ask the dog to give you advice on fixing your bicycle. You would not ask its opinion on immigration policy. But the dog’s ability to communicate in English might affect your feelings about the dog. You would tend to anthropomorphize it, You would be tempted to treat it like a human (many dog owners already do that).

What is different about a computer? Physically, it has fewer capabilities than a dog. But it has much more skill at manipulating information.

Is manipulating information the same thing as what we call thinking? Is it the same thing as what we call intelligence?

The assumption that manipulating information is the same thing as thinking or intelligence leads people to call these models “artificial intelligence.” If this is a valid assumption, then the discussions people are having about the models (what jobs can they replace? should we treat their answers as authoritative?) are appropriate.

Here I am questioning the assumption that manipulation of information is the same thing as intelligence. It is a valuable skill, to be sure. But it is not comparable to human intelligence.

Our intelligence is collective. Read your Joseph Henrich. Or, as Leonard Read put it, none of knows how to make a pencil. We have to collaborate to make one.

Each of the models is trying to achieve individual intelligence. Compared to an individual human, each model can exhibit remarkable knowledge. But I claim that our collective intelligence is still greater than that of a model. And the models will not necessarily be able to catch up.

When faced with a question for which different humans give different answers, a model will try to give a consensus response. Or, if it is a subject that its human trainers have determined to be controversial, it will decline to answer.

Try not to anthropomorphize the computer. Even if it can now talk to you in English.

Share