The tasks that GPT-3 can do are all language modelling tasks: given a sequence of tokens, what is the most likely sequence of tokens following from it? It's not surprising that this covers tasks like translation and question answering, which are language generation tasks. Other language models have been used in the same manner before, e.g. BERT. I don't remember metrics off the top of my head but I'm not sure GPT-3 is really top notch in many NLP tasks, if any.

In any case a language model is a language model. It has no other ability than calculating the probabilities of sequences of tokens. Leaving aside the question of how something like that can have "real understanding" just by being embodied in the world, how do you even "embody" a language model? I'm genuinely curious to hear how far you have thought about that and how clearly.

I mean in practical terms- you have a trained language model. You have a robotic body (not your own). How do you put them together to produce an embodied agent? What are the intermediarey steps that lead to a robot that can use its language model to... (what does an embodied language model do)?

It is impressive, but I'm not sure why doing stuff without training is such a good thing. Giving a choice, I'd rather have a model that works better with extensive domain-driven training than one that works worse without it. After all, when choosing an expert, you usually go for one that has the best experience in the field, not the one that can speak most eloquently on the widest variety of subjects (unless we're talking about politicians, there everybody does the opposite for some reason).