Machine Logic

Memories: artificial intelligence at Stanford in the 70s

11 Jan 2023

These days, artificial intelligence (AI) is synonymous with neural networks and machine learning (ML), but old-timers can remember when symbolic approachs were king: we now have the wonderful acronym GOFAI. I was at Stanford University from 1977 until 1982 and can vividly recall the unique atmosphere of the Stanford Artificial Intelligence Laboratory.

AI in the 1960s

My own interest in AI (as a high school student in the USA) arose from references to some of MIT’s luminaries in Martin Gardner’s “Mathematical Games” column in Scientific American and to a hagiographic article probably in National Geographic, also focused on MIT.¹ (Can’t find it, sorry.) I also saw a filmed demo of Terry Winograd’s SHRDLU. The dialogue that Winograd carried out with his robot about blocks in a virtual world gave a strong impression of sentience. Winograd himself later admitted that it was largely bogus:

Pressure was for something you could demo. … I think AI suffered from that a lot, because it led to “Potemkin villages”, things which - for the things they actually did in the demo looked good, but when you looked behind that there wasn’t enough structure to make it really work more generally.

The film 2001: A Space Odyssey had come out in 1968, envisaging that within 32 years, computers would be capable of carrying out intelligent dialogue (and committing murder). One of its scientific advisers was MIT’s Marvin Minsky, the de facto leader of the global AI community.

I had the opportunity attend MIT, having been offered an undergraduate place in 1973. Bizarrely, given my enthusiasm for AI at that time, I instead took up an offer from Caltech, where little AI was done. The year 1973 also saw the highly critical Lighthill Report in Britain, which led to severe funding cuts in both the USA and the UK. MIT philosopher Hubert Dreyfus published his influential critique of AI (at least, the prevailing practice of AI).

Stanford AI Lab, 1977

In the 1970s, the Stanford AI Lab (SAIL) was located some distance from the main campus, in the DC Power Building, a wooden semicircular structure that was slowly decaying. I can’t recall precisely why I used to go there, since I wasn’t doing AI. But it was well worth the half hour bike ride. It had a unique vibe.

It had one of the world’s first laser printers,² called the XGP (Xerox Graphics Printer). It was the size of a refrigerator and printed on thermal paper that came in rolls. It cut the paper into pages using an automatic guillotine, and the last page was typically a little strip (waste not want not). This printer supported Stanford’s modified version of the ASCII character set, extended with logical and mathematical symbols. An ancient line printer also supported this character set—its outstanding quality can be seen in the text of this report—and somehow the weird green terminals did too.

A PDP-10 ran a bespoke operating system based on an old version of DEC’s TOPS-10. Email pioneer Mark Crispin looked after it. MIT also used a heavily modified version, the famous Incompatible Timesharing System. DEC’s OS presumably had been made available in source form, unthinkable now. The advantages of rolling your own OS had to be weighed against the security updates they didn’t get, which already in the 1970s left them vulnerable to hacker attacks from across the ARPAnet. That’s another story.

John McCarthy, the legendary inventor of Lisp, was the scientific director, with Lester Earnest as lab manager. Strangely enough, McCarthy was only 50 when I arrived, but seemed to have left research behind him,³ and I was left with the impression that his main interest was the promotion of nuclear power. But exciting things were happening. Donald Knuth released the first version of TeX, written in SAIL’s eponymous programming language (a heavily extended Algol). Early robots were put through their paces. Key verification researchers, including Mike Gordon and Robin Milner, also spent time there. Robin created Stanford LCF, which as described elsewhere later became Edinburgh LCF, the first modern proof assistant. Robin returned to give a seminar on ML, which I attended, only to ask an extremely stupid question.

Expert systems

AI as a field seems particularly prone to schisms. Expert systems work at Stanford was, as far as I could tell, entirely separate from McCarthy’s world, and not even conducted at the AI Lab. Expert systems had somehow escaped the gloom that had fallen over AI as a whole. MYCIN, one of the earliest such, diagnosed bacterial infections and recommended antibiotics guided by approximately 600 rules that had been obtained by interviewing doctors. Entirely different in architecture from modern ML-based systems, it was similar in that it used a body of knowledge to deal with new situations. Crucially, it could explain its answers in terms of those rules, which in turn could be traced back to the doctors themselves. With machine learning we do not need the laborious manual curation and can work with vastly larger knowledge sources, but lose this accountability.

Is theorem proving AI?

Logic Theorist, regarded by many as the very first AI program, was designed to prove theorems from Whitehead and Russell’s Principia Mathematica. LT proved 38 theorems from the first two chapters. Two years later, Hao Wang did rather better:

the whole list of over 200 theorems of the first five chapters of Principia Mathematica were proved … the actual proving time for over 200 theorems was less than 3 minutes.

It’s odd that LT is celebrated while the far superior work of Wang is generally overlooked. The reason perhaps is that LT was seen as a cognitive simulation, while Wang merely used an algorithm. As it happens, algorithms often win.

The AI world at that time struggled over issues such as whether knowledge was best captured procedurally (in the form of executable code), or declaratively, and if the latter, in what sort of knowledge representation language. John McCarthy was a fan of first-order logic. For a time, the whole enterprise of automating logic, such as the CADE conference series, could be regarded as AI.

But human thought is not logical. People are spectacularly bad at logic. Actually, Hubert Dreyfus said essentially the same thing. But we have to say that theorem proving is AI, with so many people applying machine learning to it.

Postscript

I was amused to stumble upon Roger Needham’s response to the Lighthill Report.

Artificial Intelligence is a rather pernicious label to attach to a very mixed bunch of activities, and one could argue that the sooner we forget it the better. It would be disastrous to conclude that AI was a Bad Thing and should not be supported, and it would be disastrous to conclude that it was a Good Thing and should have privileged access to the money tap. The former would tend to penalise well-based efforts to make computers do complicated things which had not been programmed before, and the latter would be a great waste of resources. AI does not refer to anything definite enough to have a coherent policy about in this way.

Roger could always see to the heart of any matter. “AI” still refers to a very mixed bunch of things. This thoughtful paper (also here) considers the relationship between the exciting progress accomplished using neural networks and more classical AI techniques. And there’s an interesting survey of early AI. See also this 1969-1974 Stanford AI Lab photo album.