It feels a bit like a GOFAI way to teach machines to play chess— through hand-organized human knowledge versus more successful approaches that rely on computational scale. At the same time, it feels like it could give large language models a sense of “truth.”

The "computational scale" approach to semantic networks has been tried - that's pretty much what Freebase was. The Wikidata community refused a direct import of Freebase because the resulting dataset was full of errors; and even today it similarly rejects bot-generated data.

Right. Makes sense. But then GPT seems like it is on the verge of accomplishing something very similar but in a much more effective way. I wonder if it can use something like this for training or reinforcement learning, for instance.

GPT routinely dreams up totally fake data, it's just how the model works. It's a fun toy but absolutely not helpful for real-world scenarios.