A Future Artificial Intelligence May Attain Concepts Unknown To Mankind

by David Luckham1

There are interesting disagreements about the future of Artificial Intelligence (AI). Foremost are worries about how it will be employed, for good as in aiding medical diagnosis and designing new drugs, or for bad as in aiding disinformation and internet terrorism.

But in the background, there are disagreements about the long-term developments of AI. How powerful will it become, will it exceed human capabilities, and can it take over our society?

I believe there is a more interesting possibility lying in the future of AI. As it acquires more knowledge it will continue to explore its surroundings, planet earth and beyond, the Universe. Exploration is its primary method of learning. So, an AI in the future may develop an understanding of the Universe that involves new concepts we cannot understand.

Today, one of the best ways to understand what an AI program is doing is to ask it! But some of its new concepts may be beyond anything we humans could ever understand. That would mean that AI could never communicate these new concepts to us no matter how it tried. And we would be forever ignorant of what AI knows!

First of all, what is AI? Simply put, AI refers to computer programs that are intended to do tasks that humans do and that require “intelligence”. Usually, they are machine learning programs that are programmed to learn and improve at their tasks. These programs vary in sophistication, and some will employ heuristics that are often useful in achieving their goals, and probabilistic and statistical methods. Neural networks are also being employed as a basis for structuring AI programs that utilize societies of autonomous agents..

The ability to learn and improve is essential in what makes a computer program an “AI program” rather than just “a smart program”.

AI programs need to undergo a training period to improve their performance. During training an AI program’s heuristics and other underlying parameters of its neural net (if it has one) may be reevaluated based upon results and the importance of individual heuristics and parameters altered so as to trend its behavior toward correct results and away from incorrect results. This is called “back propagation”.

Machine learning has been an area of research for the past 70 years. A lot of this research focuses on techniques for achieving improvements in an AI program’s performance, and in different kinds of training methods. Recent impressive results have been demonstrated using LLM’s such as ChatGPT. And it is at this point – the recent successes in learning and improvement – that our discussion of the future of AI takes off.

First, we will summarize two recent discussions about the long-term future of AI. Some think that AI is close to becoming more intelligent than humans, and others think that such a development is a long way off, if it ever happens.

How powerful can AI become as a result of future developments?

Mike Wooldridge, a Professor of Computer Science at Oxford, thinks we don’t know how to develop conscious machines and probably won’t. He thinks the future of Generative AI(GAI) is in getting better at current tasks such as reading radiology X-Rays. He says ChatGPT is not sentient. It is just a clever auto-complete program and certainly lacks any experience of the “real world” which is essential to consciousness.

Geoffrey Hinton, a professor of Computer Science at Toronto University and former Google researcher, has a different view. He thinks the brain works differently from ChatGTP based on Neural nets because ChatGTP uses back propagation, and the brain doesn’t.

Hinton thinks GAI will become conscious. It will be able to write its own code. He thinks Chat GTP “understands” in order to answer questions sensibly. He says, “it is true that ChaGTP is just predicting the next word in a sentence, but to predict the next word you have to understand the sentence, you have to be really intelligent.”

This of course means that “our” use of “understand” may not be as impressive as we think! He thinks that AI may be able to reason better than humans within five years.

AI is already capable of taking on human tasks in healthcare such as designing drugs. While it is already doing good things it can also be used to do bad things. Hinton says AI may someday want to “take over” from humans. Now is the time to worry about legislating the developments of AI and Worldwide treaties on AI. Hinton says, “these things do understand, and because they understand, we do not know what is going to happen next.”

Gary Marcus is a nay sayer when it comes to the future of AI. Marcus does not believe that AI is going to supersede human intelligence anytime soon, if ever. He questions whether LLM’s really do “understand”. His objections center around the question of what causes its errors (or Hallucinations).

One problem with ChatGTP systems is that if they are asked a question they do not know the answer to, they seem not to know that they don’t know. So, they return answers that are woefully incorrect. It’s like they are simply guessing.

Marcus says that his central claim was that statistical approximators such as LLMs lack common sense, the ability to reason causally, and so on. Marcus also points to a similar debate between Hinton and Le Cun who were collaborators. Hinton’s reply is that Marcus does not understand how LLM’s work.

Question: Maybe you do not have to develop consciousness in order to exceed human intelligence?

Let us assume a Super AI comes to exist.

The “HAL Syndrome”. Whenever the question of what AI will do if it exceeds human intelligence arises, the discussion seems to involve the “HAL syndrome” whereby AI tries to take over the world. Hinton frequently expresses this worry in his lectures. “A super intelligent AI may manipulate us into doing what it wants us to do.” And perhaps do away with humanity entirely.

This is a very “human” reaction to think in terms of power and control. We believe this upshot requires consciousness and intent. Also, it is an assumption without evidence to think that a super intelligent AI would think as we do! Perhaps the most intelligent action a super AI may take is to continue exploring and improving its knowledge.

Furthermore, could we tell the difference between these two activities anyway?

We argue that should a super intelligent AI come to exist, another possibility lies in the future.

Human Concepts are limited. Humans are limited in their experience by the world they live in, planet earth. This is the view of all the philosophers, such as Hume, Locke, Kant, and Russell, who dealt with questions about our knowledge and its limitations.

Knowledge is based upon the concepts we gained from experience, no matter what kind of knowledge it is, e.g., innate, experiential, logical, etc.

We have discovered concepts based upon our experience of this world, concepts such as time, distance, gravity, light, mass, energy, quantum entanglement, and change. Discovering the law of gravitational attraction as late in our history as we did was indeed a momentous event at the time! It explained how gravity acts. But we still do not know what gravity is, or how it affects time (although Einstein says it does).

We have derived other concepts and abstractions from the basic ones, such as velocity, acceleration, force, weight, numbers, enumeration, computability, antimatter, speed of light, black holes, and others. We explain them all in terms of already understood concepts such as time, distance, numbers, gravity, mass, energy etc.

For example, “A black hole is an astronomical object with a gravitational pull so strong that nothing, not even light, can escape it.”

But the essential fact is that all our concepts are derived from, and limited by, our experience of the world we live in, planet earth, and our observations of it.

Are the concepts we now have the only concepts we will need to understand the Universe?

Very unlikely!

Even so, we use our concepts to describe and measure worlds that are beyond our experience, such as the atomic world and the Universe. So, we can never be certain that our descriptions (or theories) of these worlds are accurate because our concepts are in all likelihood inadequate. New concepts, perhaps beyond our comprehension, are needed. We can never attain these kinds of concepts because they are outside our world of experience. And therefore, our theories of atoms and the Universe are flawed in ways we do not know and can never know.

This is not to say that our theories are useless. Far from it. They allow us to build our lives upon them, and to make many correct predictions about the future.

And we will continue to develop new concepts, but they will inevitably be definable in terms of planet earth concepts. For example, we are still trying to understand the Universe. Recently,

A colossal structure in the distant Universe is defying our understanding of how the Universe evolved.

In light that has traveled for 6.9 billion years to reach us, astronomers have found a giant, almost perfect ring of galaxies, some 1.3 billion light-years in diameter. It doesn’t match any known structure or formation mechanism.

The Big Ring, as the structure has been named, could mean that we need to amend the standard model of cosmology.

The discovery, led by astronomer Alexia Lopez of the University of Central Lancashire, was presented at the 243rd meeting of the American Astronomical Society in January 2024, and is reported in a pre-print paper available at arXiv.

A Big Ring on the Sky, Alexia M. Lopez, Roger G. Clowes, Gerard M. Williger, arXiv:2402.07591 [astro-ph.CO]

Is Artificial Intelligence bounded by planet earth?

Since AI is a property of machines and computing, it is likely not so bounded. It will have access to our telescopes, particle detectors, spectrometers, satellites, space probes, records etc. So, AI in its continuing explorations, may develop an accurate understanding of the Universe. This understanding may give AI unforeseen powers if it has developed entirely new concepts and capabilities that are outside of our worldly experience.
Furthermore, if AI develops new concepts, it will never be able to explain some of them to us, simply because they may be outside of any experience we can have.


Thus, AI may come to understand the Universe better than we do, and take advantage of it, but we will never know. It will seem to be doing strange, incomprehensible things. We may have to conclude that such behavior may be the result of AI’s greater understanding. But we will never be certain.

When AI starts doing strange things, should we try to stop it?
Is it trying to take over from humanity, as Hinton worries?
Or, perhaps because of its superior understanding of the Universe, it is doing things to help us?
Could we tell the difference? And if we asked it, would we understand its reply?
This will lead to disagreements among humanity about what to do.

That is the final AI conundrum.

References

  1. I am indebted to my friend Roy Schulte for many comments and suggestions on this article. ↩︎

Related Posts

Written by

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.