Working Up

by Dwayne Phillips

The old ways of doing AI are still better than some of the new ways in some cases. The answer, of course, is to combine the best of all to do something better.

Feed the following into a chatbot: “The cow jumped” The chatbot will finish with “over the moon.”

That is from the 16th century English nursery rhyme “Hey Diddle Diddle.”

In Large Language Models (LLMs), the machines “learn” from vast amounts of text acquired from the Internet. In one respect, the “memory bank” of the computer is filled with all that stuff and the LLM mimics what is there. This machine learning works pretty well.

Now to reality: cows don’t jump and they certainly don’t jump over the moon. The expression “over the moon” could be replaced by “for joy” as in “The cow jumped for joy.” Still, cows don’t jump (very high).

Now back to the old days of AI which are now called Symbolic AI. There was no scraping the world and mimicry. People wrote “rules” about things. Some of the cow rules were:

• have four legs
• females provide milk for people who can digest lactose
• males have horns
• are mammals
• have hide can be tanned and used for clothing humans
• walk slowly
• etc.

Note that “can jump” is not one of the rules. Hence, symbolic AI techniques would never conclude that the cow jumped over the moon. Hence, symbolic AI was much better at avoiding that mistake.

Symbolic AI, given its better performance in regards to cows and jumping, also had problems not experienced by LLMs and machine learning. The attributes of cows above is pretty short. It can be longer, much longer. How long does it need to be? We never discovered that. We never really wrote a good rule set describing cows.

Alas, symbolic AI works to prevent us stating that cows can jump over the moon. LLM works to do many wonderful things other than cows and jumping over moon. The solution, of course, is to use the good in both and avoid the bad in both. Simple idea, difficult to implement.

Let’s try to do better.