Is Artificial Intelligence Stuck in the Language Bubble?

"Artifical Intelligence" raises more questions far outside the field of technology.

As one who has spent a full career in "Information Technology", I can testify that "IT" is directly or indirectly about language. New languages and new development environments come and go, but they all boil down to the interpretation of text.

As Mitchell's book teaches us, even "deep learning" models that "recognize" faces ultimately depend on masses of text - even "languages" whose vocabulary is nothing but ones and zeros.

This touches on a problem that has been recognized for thousands of years, most notably by Buddha and Lao Tsu (founders of Buddhism and Taoism respectively) and recently by Ludwig Wittgenstein, that "language", no matter how clever or deep, only stands for reality. It is about reality but it is not reality itself. You may say that this points to a fundamental problem that we may share with machines - "software" can never be more than a "description" of reality. However, unlike machines, our knowledge of the world relies on the fact that we are flesh and blood participants in the world. Billions of years of evolution have made us objects in the world that react and learn in ways that go far beyond knowledge. To cite a simple example, our hearts beat perfectly well for millions of years before we discovered the circulation of blood. The same goes for our "understanding" of the world, which seems to mostly take place "under the surface".

Mitchell gives us a snapshot history of the futile effort to make a list of everything a human "knows" by "common sense". The effort mostly shows the sad fact that computer programmers should read more philosophy. To mention just one example, Hofstader's "Surfaces and Essences" sketches out the way that virtually infinite bits of "knowledge" can be spun out from the sparse base of human experience by analogy - the engine of thought. Computers need to be "taught" to recognize analogies, such as whether there is a car in the picture. They are terrible at it. And what you get, in the end, is a clunky attempt of the computer to "learn" from thousands of people to "recognize" categories that it is utterly unable to recognize on its own.

It's even worse than that. As philosophers endlessly point out, human categories are somewhat arbitrary - suited for the unique situation that humans find themselves in the world. A chair is only a chair if you are a human that knows about sitting. 

Artificial Intelligence: A Guide For Thinking Humans

I'm an avid reader, often getting through two or three books a week. This partly due to my enviable status as a retired armchair philosopher and partly due to a thirst for knowledge - asking impertinent questions - that goes back as far as I can remember.

A persistent feeling I get with all this reading is: (a) there is so much to know and (b) almost every great idea I thought I had has already been written about by somebody smarter than me. At length.

Today, I read a special kind of book, namely Melanie Mitchell's "Artificial Intelligence: A Guide For Thinking Humans". Melanie's mentor was Douglas Hofstadter, someone we have both admired for a good part of 50 years. She "gets" Hofstader in a way few do. Inspired by Hofstader's alarm at the way the fiel
d of AI is heading, she provides us a deep, knowledgeable and informative overview of how "AI" actually works and what still remains as its fundamental challenges.

As it happens, her very readable book ties together a number of threads that I have been following in my main blog."Man and Superman". As with any great book, she provides powerful insights while straying maddeningly close to ideas I have been thinking about almost all my life. My own career started in "AI" in 1969 in a silo that would now be called "Good Old Fashioned AI" - GOFAI. The questions and methods that Melanie describes in her book were all very much alive back then. In fact, in spite of all the hype, most of the original questions remain unanswered. Indeed, it is quite reasonable to say that the fundamental question: "Can machines think?" will never be answered if, by that, you mean "think" in a way that is in any way similar to what humans do all the time, every day.

Why not? Mitchell provides an exhaustive survey of the major lines of research over the last 1/2 century that aim to create "thinking machines". It is not surprising to find her referring back to her mentor's most famous work (GEB-Goedel Escher Bach) which uncovers several of the key problems. For one thing, "understanding" seems to be a fundamentally human kind of thing. All our efforts since then have done little more than shine a bright light on what we mean by "understanding" while creating computer programs that stray further and further from true "understanding". These machines have no more idea of what they are doing than a book would "know" what it's "about". What is confusingly referred to as "Artificial Intelligence" is really an extremely indirect and convoluted form of human intelligence, "mashing up" input from thousands or hundreds of thousands of human beings to produce a "consensus" view that can sometimes be amazing and sometimes laughably wrong.

In this blog, I will often return to the example of my treasured books. Each one is full of wonderful ideas. None are strictly original, in the sense that the author claims to have come up with something "out of the blue". In that sense, each book is a "mash-up" of the ideas of others - sometimes thousands of others. What the author does is provide a "mash-up" that is perhaps original, educational, relevant and interesting. In this blog, I try to do no more than that.

The point is that the book itself (a technology as old as civilization) doesn't "know" anything at all. It is simply a means to report the knowledge of humans.

This is true of the mass of knowledge we refer to as "Science". We use the jaw-dropping technology of the Hubble Telescope to discover (and "know") all kinds of things about the Universe. But the telescope itself has no idea what it is doing or why. Stepping on to the moon, Niel Armstrong acknowledged this fact by claiming his act as a "great leap for mankind".

Between the pages of Mitchell's book is a revolutionary but un-stated perspective. There is really nothing like "Artificial Intelligence" in the way most people understand it. Many people have reified this idea, making "AI" a monster that threatens to destroy us. It makes good movies, such as the Terminator. But the danger doesn't come from thinking machines. It comes, as it always has, by the misuse of technology. The danger in our time comes from stunningly powerful machines owned and operated by stupid people. My thesis is that "artificial intelligence" should refer to our misguided assumptions that humans are intelligent. Our own intelligence is what is "artificial". Our technologies, whether they be bows and arrows or IBM's WATSON, do nothing more than pool the hard-won discoveries of the human race to create tools that can save us or destroy us.