“AI”: what’s in a name?

In spite of my near-total isolation from news media, I have noticed that references to “AI” are ubiquitous. A perusal of such material suggests that “AI” means Machine Learning, which apparently means the newly discovered technique of multilevel neural networks. While this technique is spectacularly effective, it should be realized that it still is still basically a curve-fitting exercise (albeit with a novel family of curves). The result is a tool that lacks in the transparency that is taken for granted in older methods. In this essay I explore the question whether there is any connection between “AI”, as recently featured in the non-technical media, and the Artificial Intelligence envisaged by John McCarthy, who invented the term in 1955.

 

History

On August 31, 1955 a document [1] was submitted with the title “A Proposal For The Dartmouth Summer Research Project On Artificial Intelligence”. The authors were:

• J. McCarthy, Dartmouth College
• M. L. Minsky, Harvard University
• N. Rochester, IBM Corporation
• C.E. Shannon, Bell Telephone Laboratories

The abstract:

We propose that a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.

Although the term “artificial intelligence” (henceforth AI, without the quotes) was launched in this document, the idea had been in the air for a few years. Turing had discussed it during World War II. In 1958 a symposium was held with the title “Mechanisation of thought processes”. McCarthy contributed a paper “Programs with common sense” in which he presents his plan for a program called Advice Taker [2]. Though conceived as a computer program, it was a radical departure from any existing one: the processing was to be symbolic rather than numeric and, instead of being controlled by commands, it “took advice”.

The Advice Taker project needed a suitable programming language. McCarthy rejected Newell, Shaw, and Simon’s IPL, the only candidate. So a new language was called for. McCarthy liked Fortran’s “algebraic structure”, as he called it. He tinkered for a while with a Fortran compiler which was to accommodate IPL’s lists. Miraculously, what emerged from this unpromising start was LISP.

The Advice Taker concept was sharpened to the concept of a Question Answering program called QA3, and described by Cordell Green in his 1969 paper “Theorem-Proving by Resolution as a Basis for Question-Answering Systems” [3], which, among other things, was a precursor of logic programming.

What is “intelligence”, anyway?

In centuries past philosophers inquired into the nature of intelligence. This happened in response to external stimuli, sometimes in the form of new technology. One example is the concept of man as a machine, stimulated by the building of ever more ingenious and complex mechanisms in the 18th century.

Another such instance of new technology as mind stretcher started with the mechanisation of telegraphy in the 19th century, leading to the teleprinter. By 1917 this technology had matured sufficiently for Gilbert Vernam to invent a teleprinter that encrypted or decrypted automatically as it sent or received. A wireless version was used during the second world war. Much of Turing’s work was on decrypting messages encrypted by TUNNY, as the German version was called by the British [4], A digital electronic device named “Colossus” was built to facilitate decryption. Colossus was a precursor of the stored-program digital computer. The teleprinter was prime choice for communication with the early stored-program digital computers, the first of which became operational in 1949, in England.

It was an intriguing possibility that a person at a teletype would not know whether it is connected to another person at a teletype or to a computer. This possibility inspired Turing to conceive the Gedanken experiment that has become known as the Turing Test [5]. Apparently for Turing the test for intelligence in a stranger was the ability to strike up a congenial conversation such as one experiences occasionally with a stranger in a plane, on a train, or at a reception.

For others the test for intelligence is the ability to solve problems. At MIT a program was written to solve freshman calculus problems. Another program solved word problems. A program was written that solved planning problems in a blocks world from instructions typed in English.

Donald Michie, a collaborator of Turing during World War II and the founder of the Department of Machine Intelligence in the University of Edinburgh, pointed out that calculus problems can be solved algorithmically. Thus the mere fact that a program solves such problems is not an indication of its intelligence. Another example is chess. When Deep Blue gave Kasparov a hard time, it was an algorithm and an intelligent agent reaching the same high level of performance. The most blatant example of the contrasting approaches to solving problems appears in multiplying large numbers. There is a simple algorithm for this. When a human performs the same task, as some calculating prodigies can, it requires a sizeable knowledge base and the rapid formation of a plan; that is, it requires an intelligent agent.

Ability to solve problems is apparently not a satisfactory criterion for intelligence. A better criterion may be the ability to quickly learn a wide variety of things. But what is “learning”? This could be defined as “an agent has learned something when it has knowledge that it didn’t have before”. This is problematic because of “knowledge”.

We must not be deterred because we can’t define “intelligence” or “knowledge”. We can make progress by trying best-effort answers, even when they send us around in circles. Otto Neurath, a philosopher and economist who was active in the first half of the 20th century, described the predicament as follows:

We are like sailors who on the open sea must reconstruct their ship but are never able to start afresh from the bottom. Where a beam is taken away a new one must at once be put there, and for this the rest of the ship is used as support. In this way, by using the old beams and driftwood the ship can be shaped entirely anew, but only by gradual reconstruction [6].

Expert systems

How does an agent acquire knowledge? McCarthy was not interested in reward schedules for rats running mazes. He wanted a computer program that learned by being told, which is one of the capabilities of his projected Advice Taker.

Advice Taker was a general-purpose program. One can also imagine a program that is specialized to acquire the knowledge of an expert in a specialized area. Such a system is called “expert system”. An attractive area was that of a specialist physician. Hence INTERNIST-I, a system was designed to capture the expertise of just one man, Dr Jack D. Myers. By the late 1970s fifteen person-years of work had gone into its development. It had some educational use, but was not usable in practice. Such a tool is direly needed. The fact that it is not widely used implies that the approach is a dead end.

Dr Myers is not the first medical specialist held in awe by his colleagues for miraculous diagnostic powers: in New York Dr Emanuel Libman had a similar reputation. David Ogilvy met Libman in the late 1930s. In his memoir [7] Ogilvy reports:

Dr. Alexis Carrel, who was then head of the Rockefeller Institute of Medical Research, told me that the most important thing in medicine was to persuade Libman to write down his methods of diagnosis before he died. But this tiny, whitefaced old man got childish pleasure out of mystifying his fellow doctors and never did so. He published more than a hundred papers, but held back his magic tricks.

The childish pleasure is Ogilvy’s guess. I have a better theory: Libman could not write down his methods. In his book The Tacit Dimension [8] the philosopher Michael Polanyi writes on page 4: “we can know more than we can tell”. He devotes the rest of the chapter “Tacit Knowing” to substantiating this claim. Dr Carrel’s belief was erroneous, but plausible. It was shared by the instigators of the INTERNIST-I project. One wonders whether machine learning applied to the data collected might be a way to gain access the tacit knowing of Dr. Myers.

Alien Intelligence

I.J. Good introduced the concept of an Ultra Intelligent Machine, which he defined as follows.

Let an ultra-intelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultra-intelligent machine could design even better machines; there would then unquestionably be an “intelligence explosion,” and the intelligence of man would be left far behind. Thus the first ultra-intelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control. It is curious that this point is made so seldom outside of science fiction. It is sometimes worthwhile to take science fiction seriously. [9]

This was in 1963. After having lain dormant for a long time, the concept has given rise to a flurry of apocalyptic speculation (Bostrom, Tegmark, Harari, Lovelock). Opinion on the possibility of an Ultra Intelligent Machine is divided. There are those whose urgent concern is to prevent its realization. There are those who are resigned to, or even welcome, the advent of the UIM. Finally there are those who believe the possibility is near enough to nil not to worry.

I read Good as saying that an intelligent machine is unlikely have a level of intelligence that is in the same order of magnitude as that of humans. The lower limit is nil; such machines are familiar and ubiquitous. As we have seen here, there are machines whose level of intelligence, while debatable, is not such that visions of UIM arise. Good points out that there is no upper limit. Apparently Good views intelligence as totally ordered, that is, that any two intelligences can be compared with the result that they are equal, that one is greater than the other or the other way around.

I propose that intelligences are partially ordered, with comparability as an exception rather than the rule. Intelligences other than the human kind are likely to be incomparable, so alien rather than “ultra”. The case for powerful alien intelligences is suggested to me by Chomsky’s book What kind of creatures are we? [10].

The Materialist Thesis (MT) holds that all mental phenomena have a physical basis. This is universally accepted by scientists, but only in the weak sense that the alternative is inconceivable. Chomsky introduces a stronger version: our physical brain structure determines both the scope and limitations of what we can know. I shall refer to this as SMT, the Strong Materialist Thesis.

The SMT makes it interesting and even urgent to develop intelligences with a different physical substrate than the human brain and therefore with different scopes and different limitations. Such intelligences are the ones I call alien. The possibility of these is not suggested by Chomsky but seems to me a consequence of the SMT. Chomsky is concerned with the SMT.

It remains to give an idea how Chomsky justifies the SMT. As for the MT, it is hard to disagree, but it is inconsequential. The SMT is consequential, but is hard to believe. Chomsky justifies the SMT in two chapters of [10]. An outline follows.

What is language?

No language, no thought. The ability to acquire language is innate. It is supported by special-purpose brain structures. These structures arrive in two stages: stage I is complete at birth; stage II develops with the still rapidly growing brain in the first few years of life. During this stage exposure to language users is essential.

Language is composed of two components: I-language and E-language. I-language (I for Internal) is the Language of Thought. It has a grammar according to which non-sequential structures are manipulated. E-language (E for External) is sequential. It is used in speech or writing.

What can we understand?

These features of language imply that the scope and the limitations of what we can know is determined by features of brain structure. This is the SMT. The limitations imply that only some unknowns are knowable. Chomsky calls these “problems”. The other unknowns he calls “mysteries”.

Alien Intelligence

Chomsky’s theory suggests that thought can occur in non-human brain structures. I would call this Alien Intelligence. Kowalski argues [11] that logic is the human Language of Thought. This thesis can be made less vulnerable by splitting it in two parts, namely his answers to the questions: (1) What is the human Language of Thought? (2) Is logic adequate as Language of Thought for some kind of thinking agent? An affirmative answer would open the way to an interesting intelligence, though possibly alien. LPS, Logic Production Systems, represents progress by Kowalski et al. in this direction since [11].

References

[1] “A Proposal for the Dartmouth summer research project on Artificial Intelligence” by J. McCarthy, M. L. Minsky, N. Rochester, and C.E. Shannon. 1955.
[2] “Programs with common sense” by J. McCarthy. Symposium on Mechanization of Thought Processes. National Physical Laboratory, Teddington, England (1958).
[3] “Theorem proving by resolution as a basis for question-answering systems.” by Cordell Green. Machine intelligence 4 (1969): 183-205.
[4] Breaking Teleprinter Ciphers at Bletchley Park: General Report on Tunny with Emphasis on Statistical Methods (1945) Editor(s): James A. Reeds, Whitfield Diffie, and J. V. Field. Published 2015 by Wiley. Consists of the 1945 report written by I.J. Good and D. Michie and supporting articles.
[5] “Computing Machinery and Intelligence” by Alan Turing. Mind, LIX (236): 433–460.
[6] Problems in War Economics by Otto Neurath. According Wikipedia article “Neurathian Bootstrap”, October 4, 2019.
[7] Blood, Brains, and Beer by David Ogilvy. Atheneum, 1978.
[8] The Tacit Dimension by Michael Polanyi. Peter Smith, Gloucester, Mass., 1983. University of Chicago Press, 2009.
[9] “Speculations concerning the first ultra-intelligent machine” by I.J. Good. Advances in computers. Vol. 6. Elsevier, 1966. 31-88.
[10] What kind of creatures are we? by Noam Chomsky. Columbia University Press, 2015.
[11] Computational Logic and Human Thinking by R.A. Kowalski. Cambridge University Press, 2011.