We have strong reason to suspect that GPT-3 has at least some model of what words mean, from experiments people have done where they asked it to use words in no

We have strong reason to suspect that GPT-3 has at least some model of what words mean, from experiments people have done where they asked it to use words in novel contexts and ascribe properties to them, which GPT-3 is surprisingly good at.