Intelligence is a Process
It's not a lump of knowledge
Gemini, Google’s version of Claudine Gay, became the Current Thing for the last half of February. I am not going to join the pile-on here.1 But I do want to meditate on the topic of “artificial intelligence.”
Usually, we define an artificial something in relation to its natural version. An artificial diamond is defined relative to a natural diamond. Artificial turf is defined relative to natural turf. Artificial insemination is defined relative to natural insemination.
In the case of artificial intelligence, we have a problem. There is no clear, settled definition of natural intelligence. If we are not sure what the natural thing is, how can we know what the artificial thing ought to be?
In fact, I want to claim that intelligence is not a thing at all. It is an ongoing process. It is like science. You should not think of science as a body of absolute truth. Instead, think of the scientific method as a way of pursuing truth.
One should resist the temptation to think of intelligence as a huge lump of knowledge that an entity possesses. Memorizing the encyclopedia does not constitute intelligence.
Human intelligence is collective. Pretty much everything I know I learned from other people, either directly or indirectly.
Human intelligence is not fixed. We are constantly trying out new ideas. Many of our beliefs are contestable. The science is not settled.
Human intelligence is evolutionary. We want to obtain true beliefs and to discard mistaken beliefs.
Institutions help to guide the evolutionary process. Free speech, open inquiry, and the scientific method for gathering and debating evidence are examples of such institutions. Jonathan Rauch coined the expression The Constitution of Knowledge.
I prefer to think of intelligence as residing in these collective institutions. It is the process by which good institutions guide our beliefs toward truth and away from error. The institutions are not just buildings with four walls or organizations to which we belong. They include social norms and prestige hierarchies.
Large language models are not the artificial equivalent of the process of improving knowledge. They are outstanding tools for communicating between humans and computers. But they do not perform the functions of the Constitution of Knowledge—trying new ideas, testing them, keeping what works, and discarding the rest.
Windows, icons, menus, and pointers were a big improvement in human-computer interface. Fortunately, nobody pinned the label “artificial intelligence” on WIMPs. Unfortunately, large language models and their relatives came with such a label, which puts a burden on them that I do not think they are able to bear.
As the public discovered with Google’s Gemini, human “trainers” strongly influence the behavior of the chatbots and their relatives. When the trainers are clumsy, so are the outputs of the models.
Intelligence, as I define it, will not come from larger datasets and more computing power. Intelligence is the process of improving knowledge. It requires treating beliefs as contestable. We need to be open to learning.
We may be able to use the latest models as tools within the process that I think of as intelligence. But they are not a substitute for that process.
I believe that we are at the early stage of learning how to use the new models. As we try to sort that out, I really wish that we would lose the expression “artificial intelligence.” It offers a misleading picture of what we can expect from the new models.
If you want more pile-on, Ben Thompson offers a useful analysis. Lee Bressler writes,
Put yourself in the shoes of the Google employee who worked on Gemini and knew that these problems existed but didn’t raise his hand. Do you think he didn’t speak up because he knew that if he did, the entire DEI apparatus of Google and the cultural weight of the firm would come crashing down on him? Would everyone complain to HR that he’s some sort of racist bad guy and then he would get fired just for putting up a flag? It’s quite possible.
The anthropologist Rob Boyd once began a seminar with this question: which species occupies the broadest range of habitats in the world? I was thinking some insect or bacterium but the answer is obvious, it's us. By a long way. He then noted that a young human child without access to adults would not be able to survive for more than a few days, if that. Not matter how individually capable mentally or physically. What we have managed to achieve as a species is entirely through our capacity to transmit and build on prior knowledge, which is as you say a form of collective intelligence. I was just reminded of this incident by your post.
Your definition of intelligence seems to be defining intelligence as not something individual humans possess more or less of, but something groups of humans have. I think you have applied the metaphor of group intelligence so hard that you have ceased to apply it to its original subject, leaving a gap where once we had a word for that capacity of humans to take a set of information and work out solutions to novel problems.