AI language models like GPT-3 can achieve 97% accuracy on some Winograd schematics, but understanding the language is not the same as understanding the world.– It’s simple enough that the AI seems to understand the data, but designing a real test of machine knowledge has proven difficult.
