My understanding is that LLM and Cyc are fundamentally different forms of AI. Even if you could turn OpenCyc into text rules, once ingested it would just dissolve into the ocean of training text data and would not significantly gain more apparent "common sense" than it already had. Maybe a more interesting combination could be to have both Cyc and LLM working side by side and comparing notes before agreeing on a result.
GPT-4 seems capable of creating CycL output from a text prompt. It might be an interesting guard against hallucination - much like a student being asked to show their work, you're forcing the LLM to go through the steps of framing the problem logically, in a way that's interpretable by the teacher independent from the student's thought processes.
That said, it certainly seems like there hasn't been recent work on hosting the OpenCyc knowledge graph in a reasonably modern way, much less the more recent closed-source work by Cycorp (https://cyc.com/). And it's likely GPT-4 doesn't know the full capabilities beyond whatever tutorials were on the web at the time of its training. If I were Cycorp I'd be seriously looking at developing this kind of hybrid model, with an agent model having access to recall their closed-source examples, as a paid cloud offering; there would likely be many who would desire this best-of-both-worlds.
