Cyc started in the mid-80s when rule-based AI systems were the norm. Everything in AI looked like a theorem prover back then. However, in the 90s, statistical methods began to produce better results as techniques were refined and computing power increased. What's better: a system that produces the correct (or good) results 85% of the time, or one that produces no result until the database has a necessary rule added? Industry prefers the former (although there are definitely uses for the latter).

Its connection to AI still remains, but the field has largely moved away from the rule-based paradigm Cyc is based on. I mean, you can throw a whole wikipedia segment into Doc2vec (a lightly supervised deep-learning technique) and it can discover non-trivial links between semantic objects with a few days work. If you tried that with Cyc, it would take longer.

It was a bad approach to the problem.

Perhaps it could be made into a good approach if it were reworked in terms of topic modeling or clustering categories and them feeding them through a machine learning KNN or bayesian model enabling prediction.