AI in the 80s? How a Simple Animal Guessing Game Pioneered Machine Learning

True. For example, the Apple ][+ came with a demo disk full of programs for the then-new Applesoft BASIC language, and this was one of them. The questions were saved by POKEing them into arrays hard-coded into the program, allowing you to SAVE the modified program after running it.

It seemed like a neat trick at the time. There was also a crude CRUD database that worked the same way, retaining up to 50 names and phone numbers.

Actually that disk had a lot of really cool programs, now that I think about it. A biorhythm plotter, Woz's "Little Brick Out" Breakout clone, and a few other demos by luminaries like Bruce Tognazzini and Bob Bishop. And of course, ELIZA, the mother of all LLMs... only it was called FREUD for some reason.

I believe that the intention was to say "No ability to save progress _between sessions_" in the original program, whereas the C++ implementation saves to text files.

Another portion of the article says more explicitly:

  Limitations:
    Maximum number of nodes: 200.
    The structure is stored only in memory (no disk saving).

Further suggested by the imagery used being AI generated

I read about this idea and actually built a variant in college for the Washington University ACM (our CS student club). We ran a demonstration at one of the engineering school events and ended up refilling the matchboxes multiple times that day - it was a hit!

If I recall, there is a way to significantly reduce the number of matchboxes needed by taking advantage of mirrored board conditions. Somewhere I have a Perl script that generated HTML for all the matchbox covers and possible next states.

That's immensely more interesting than a program that plays 20 Questions. I'm amazed I've never heard of it!

True. The program in the article is an expert system, not a decision tree learner. The confusion probably stems from the fact that decision tree learners are (still?) a very popular machine learning algorithm; and because most people who are into AI and ML today don't know anything about its past.

And what happens to those who don't remember the past? Ask George Santayana.

Well, dear Peggy, the program in question is, in fact, known to have existed in 1974 on Multics (https://multicians.org/terminals.html), so you're not half wrong.

The other major issue is that it isn't machine learning—it's a classic example of an expert system, even if there is a bit of a gray area around whether a binary decision tree qualifies as ML.

The worst part, of course, is that it takes less time to slap "GUESS THE ANIMAL" on a stock image of a glass terminal than it does to prompt a diffusion model to generate the same thing with passable fidelity... and it still wouldn't be an accurate depiction of what the program actually did.

I cried "LLM garbage" too. Catchy title, article full of bizarre errors and repetition.

It would be a prank--LLM writing about AI, badly.

But the Medium account behind it has only articles like this.

It can be difficult to find hero images for articles like this, so people generate them. Why is that bad to you?

There is a very important reason why "decision tree" is both a deductive inference procedure and the "model" (really, a propositional logic theory) learned by decision tree learners.

The reason is that the way decision trees, well, make decisions, is the way that expert systems of the '80s made decisions. The first expert system was called DENDRAL (capitals mandatory), derived from the Greek word for "tree", "dendron" (also see: "dendrite"; little tree). DENDRAL and similar systems implement deductive inference procedures that traverse a tree (either bottom-up or top-down, with backward or forward chaining). In traditional expert systems the edges in the tree are production rules of the form IF-THEN-ELSE and the nodes are intermediary goals of what is essentially a proof of a propositional logic formula (the decision tree; it's a boolean formula).

Expert systems' Achilles heel was the so-called "knowledge acquisition bottleneck" which refers, in short, to the difficulty of constructing and maintaining a huge rule-base as needed for an expert system that can do something more interesting than play 20 questions like the animal identification system above. Because of this difficulty, expert systems never quite fulfilled their promises and they suffered from the brittleness that is one of the first thing anyone hears about expert systems today.

And what did people come up with in order to overcome the knowledge acquisition bottleneck? Why, machine learning, of course. The early days of machine learning as a field --remember, we're back in the '80s, a bit before the AI winter hits-- were exactly systems that tried to learn the IF-THEN-ELSE rules for expert systems' rule bases.

Decision tree learners, like ID3 and C4.5 by J. Ross Quinlan, and similar algorithms, that were wildly popular even well after the ImageNet moment of deep neural nets come from that era. Instead of learning only the rules though, decision tree learners, learn an entire decision tree; in short, an entire expert system, from top to bottom.