But is it intelligence?

I hope you welcomed the new year in style. While right now many have hit Pause, this newsletter brings you some thoughts on just one topic (after a short preview of another): two concepts of intelligence.

Blog article: writing good technical definitions

I have just republished as a (longish) single article, with some updates, a discussion released earlier in three successive part on the Communications of the ACM blog. It is a summary of reflections, developed from the examination of (and struggle with) numerous definitions, some good and many bad.

Starting from some of the later — including pretty pathetic ones from such prestigious sources as the IEEE — the article develops a set of principles for effective technical definitions. I was surprised to see that they boil down to just one rule. I will come back to the topic in a future newsletter but here is the current article: Criteria and recipes for good technical definitions.

And now for the main course.

Two concepts of intelligence

“But is it intelligence?”

The rise and spread of AI brings along endless discussions of what constitutes intelligence. A debate which, even if we limit ourselves to computer science, goes back not just to Weizenbaum's Eliza but to Turing and von Neumann.

As an example of one viewpoint, a committee of the French National Assembly recently held some hearings on Artificial Intelligence and invited a philosopher of science, Olivier Rey. (The video is here on the official site. It shows an attendance of four, probably including the secretary, suggesting that French deputies do not find AI a topic worth their time.) Throughout his speech, he explains that “artificial intelligence is not intelligence” because the program “does not understand”. I find this kind of assertion fascinating. I am not saying it is wrong; just fascinating. Actually I do not know if it is right or wrong.

It is very simple to see why I do not know (and why I believe Mr. Rey also does not know). For example I can with some assurance say that I “understand” the basics of linear algebra. Even so, if you ask me to find the eigenvalues of a small matrix I might occasionally make a mistake, and if you ask me to prove one of the classic theorems in the field I might occasionally get stuck. If you ask an LLM the same questions it will get the answers right much of the time but may also occasionally mess up (“hallucinate”). So what enables me to say that I — or for that matter Mr. Rey, who I see from his LinkedIn CV went to a school that I also attended, so that I know its entrance requirements — are more intelligent than Claude or Gemini? I actually suspect that the LLM will get answers right more often than I do, but that hunch is irrelevant: both the LLM and I will get many answers right and some wrong.

I can hear many possible retorts, but they are all of the same form: the AI tool only appears to understand, I really understand. It does not matter that I make mistakes once in a while, they are only superficial mistakes, whereas the tool's hallucinations shows that it has no clue whatsoever. All beautiful arguments — but really worthless. They are worthless because they do not satisfy the basic criterion of scientific arguments: they are not falsifiable. Falsifiability would mean that we can construct a reliable experiment to determine that a tool is not intelligent. All such experiments, at least all those devised so far, can only measure outcomes. Think of the Turing Test, or Searle's Chinese Room arguments. Both today's best tools and competent humans will pass them. Throw in enough complexity and tools will actually fare better than humans. Does it mean they are more intelligent? Do they understand less? Might they possibly (horribile dictu!) understand more?

We are back to the unanswered question: what does it mean that a human or a tool “understands” a question or concept?

I believe that much of the debate is due to a fundamental misunderstanding or, to be more precise, clashing understandings of what intelligence is about.

The clash rings very personal since it reminds me of the shock I experienced when as a student at Stanford I first came to the legendary AI lab, SAIL, then at its zenith with luminaries including not just the SAIL director, John McCarthy, but people such as Art Samuel, Zohar Manna and Terry Winograd. “Intelligence” was on everyone's lips and I vividly remember discovering that the working definition was, from the American Heritage Dictionary, where the entry started: “Intelligent usually implies the ability to cope with new problems and to use the power of reasoning and inference effectively.” That view was scandalous to me, coming from a European intellectual perspective. These Americans, I thought, are so utilitarian, prosaic, earth-bound, pedestrian, mercenary! Surely there has to be something deeper to intelligence than knowing how to react to circumstances: you have to understand the situation. I had studied Latin and new that etymology was on my side: intelligo means “I understand”.

As I soon found out, the issue was not just with me but reflected a difference between continental European and Anglo-Saxon views. The Larousse definition, for example, starts with “the set of mental functions whose object is conceptual and rational knowledge".

I now think for a large part that I was wrong and the Americans had it right. Of course there is a long and fascinating European tradition of explaining things and as a result sounding very smart. The French in particular have made a specialty of visiting a country for a few months or just a few weeks, bothering or not to learn the language, and then explaining it to the gobsmacked natives. Tocqueville is the most famous example, but there is also Roland Barthes on Japan. Not French and not harmless, we have Marx and Freud who respectively “understood” all about human history and human psychology and explained it to us. It is really petty to point out that these theories had zero success in predicting future outcomes. Or that in the first case their main result was to destroy countries and civilizations and lead to the death of millions of people. Who is to quibble about such minor outliers when these theories “explain” things so elegantly! 

Serious scientific theories do explain, too, and make us understand complex things.  The difference is that they predict correctly, and are falsifiable. The way relativity  “explained” the basics of time and space was not just to present convincing ideas but to predict that, under the circumstances of a certain eclipse, at a certain place, light would bend not by 0.87 arcseconds, as Newton would have had it, but by twice as much. Had the measurement been different, Eddington would have disproved the theory.

The difference between the two concepts of intelligence — ability to understand, versus ability to act successfully — is also the difference between deductive approaches, which start from a theory and attempt to verify it through facts, and inductive approaches, which start from facts and build up a theory. It is a deep difference and and goes back far in the history of thought. It Among philosophers — simplifying things, as detailed views can be more nuanced — we find, on the conceptual/deductive side, Descartes and Kant and on the empirical/inductive side such English and American thinkers as Hume, John Stuart Mill and behaviorists typified by Skinner.

The first view says “I am intelligent because I understand”. It is appealing, elegant, and propitious to powerful speeches. But how do we validate or falsify it? On just about any topic, the conspiracy theorists (including the Marxists and the Freudians) also make beautiful speeches. If you and I both have explanations for something, but they are incompatible, how do I convince you that mine is right and yours is wrong? The second view says “I am intelligent because I can make predictions that turn out right most of the time”. But how do we know that this is really intelligence and not just careful record-keeping?

Old-AI, with its expert systems and logical deduction tools, was of the first kind, deductive. The consensus is that it failed. Modern-AI is almost entirely (at least in the current, obviously intermediate state of evolution) of the second kind, inductive. Modern-AI is machine-learning: it builds answers to new queries by relying on a large body of validated answers to previous queries. In its flagship areas of application, the new answers should be right most of the time. Is it intelligence? Is the human-level translation of today's translation tools intelligent? Is a vibe-coding tool more intelligent than the programmer who uses it, or less? Is a medical-image analysis tool which produces fewer false negatives and false positives than a Stanford Hospital radiologist more or less intelligent than that doctor? For that matter, are non-AI programs such as EiffelStudio or even gcc intelligent (after all, no human would be able to compile a 100,000-line program in any reasonable time and with any serious likelihood of correctness)? I do not know.

I do not even know if these are interesting questions. But for many people, they are. And I do know that to get the discussion even started it is necessary to settle on a definition of intelligence, that there are two of them, and that they are radically different.

Cover photo: decoration of a Loge (Primo Ordine / Dress Circle / Premier Balcon) at the Teatro La Fenice, Venice. (By the way, all the photos illustrating this newsletter are by me. No connection is claimed with the newsletter's topics.)