This is the first in a series in which I think about intellectual creativity, the relationship between humans and machines in intellectual activity, and the limitations of machines. Do these limitations reflect only the limitations of current architectures or are they inherent in the nature of artificial intelligence? The idea is to refine my thinking about the problem so as to better specify that nature of those possible limitations. Perhaps when they are more closely specified we can see what would be required to overcome them.
* * * * *
Back in 1990 David Hays and I published an article, The Evolution of Cognition, in which we asserted:
A game of chess between a computer program and a human master is just as profoundly silly as a race between a horse-drawn stagecoach and a train. But the silliness is hard to see at the time. At the time it seems necessary to establish a purpose for humankind by asserting that we have capacities that it does not. It is truly difficult to give up the notion that one has to add “because ... “ to the assertion “I’m important.” But the evolution of technology will eventually invalidate any claim that follows “because.” Sooner or later we will create a technology capable of doing what, heretofore, only we could.
Just what did we have in mind? It seemed obvious at the time, at least I think it seemed obvious to us. Take that final sentence. Was it predicting what is now sometimes called AGI (artificial general intelligence) or only that sooner or later some computer would perform any (intellectual) task as well as (the best?) a human? That’s not clear.
That was a long time ago. It would be seven years before Deep Blue would beat Gary Kasparov at chess to become the first computer to beat a reigning human champion. The linguistic fluency of LLM-based chatbots was decades in the future. Back then the issue seemed distant and so clarity and specificity were not needed. These days the situation is quite different. Some are telling us that AGI will happen any day now, then then super-intelligence will not be far behind. The fact that neither AGI nor super-intelligence are well-defined is no deterrence to such predictions. Are these people only asserting what Hays and I had said 25 years ago?
I don’t know. But I want to take another crack at the problem. Current systems are, for the most part, benchmarked using bounded problems. The problems may be difficult to very difficult. The best current LLM-based systems do very well on such benchmarks, leaning many in the industry to believe that AGI is just around the corner. Those who are skeptical about that nonetheless believe that AGI will be accomplished when the current architectures are sufficiently scaled.
I’m not impressed. That’s quite different from real intellectual activity, which is often unbounded. In that situation the first problem is simply to identify a bounded problem. Once that has been done, one can proceed to solve it. LLM systems can’t do this because they operate only when given a prompt, and the prompt thereby serves as a boundary. So that’s one thing I want to look at, unbounded problems.
Here’s how I’m framing it:
All consequential intellectual activity takes place in a network of interconnected agents. These agents provide various intellectual services to one another. Some or all of the agents are human; some or all are computers. I assume that there are at least some, perhaps many, activities where a completely computerized network is more effective than any network where there is at least one human agent. The job of that human agent is to place a boundary in the space so that there is a solvable problem. Are there any activities such that a network with at least one (highly trained) human agent is more effective than any completely computerized network?
I am going to think through this issue by considering two specific cases: 1) interpreting Spielberg’s Jaws using the ideas of René Girard, 2) analyzing the population of Xanadu ‘memes’ on the web. I’ve chosen those two problems because: 1) I have done both of them myself and so can talk about how I bounded the problem space, and 2) I have directed ChatGPT to perform them. That will give us the next two posts in this series.
In the fourth post I want to consider the problem of creating a cognitive network diagram that expresses some of the semantics underlying Shakespeare’s sonnet 129. I have done this, and ChatGPT has failed to do it. It is not at all clear to me whether or not Whether or not that is so, it involves establishing a close relationship between a linguistic object, the sonnet, and a visual object, the diagram. I also want to discuss chess in this context as it has both a linguistic aspect, the rules of the game, and a visual aspect, the disposition of pieces on the game board.
In the fifth post I want to take up the issue of a problem solution that requires creating a new paradigm, in the sense Thoms Kuhn used the term in The Structure of Scientific Revolutions (1962). Solving such problems requires access to the external world so that one can make observations and, I believe, it also requires discussions with other epistemic agents, agents as powerful as the one proposing a solution. At the moment humans are the only such agents. Will there ever be a computer that is such an agent? That’s the question, isn’t it? Access to the external world seems to be a solvable problem. I’m not sure about epistemic power.
In the sixth final post I’ll discuss memory in humans and machines, for that’s what I think the issue is, the nature of memory. The by now conventional assumption is that more is better, the more memory the better. When we get enough memory, along with computer and model parameters, BINGO! we’ll hit AGI, and super-intelligence will not be far behind. This assumption fails to take account of the fact that the machine must be able to access the contents of its memory both efficiently and accurately. That’s what I’ll be discussing the sixth and final post.
No comments:
Post a Comment