As an experiment I used a GPT-3-powered website [1] to see what GPT-3 has to say about bears, and the first answer was:
> "Weird that every day, there are so many cute/funny/entertaining bears to enjoy online but hardly any on the ground."
When asked about beards, the first answer has no relation with beards at all:
> "If a person doesn’t constantly outwit, outplay, outlast, others, the strong eat the weak."
And then there's that time when GPT-3 told someone to kill themselves [2].
While funny and (mostly) grammatically correct, these "thoughts" are nonsense and no amount of extra parameters is going to solve the disconnection between GPT-3 and reality. I imagine you could condition GPT-3 to generate text for a specific piece of data in such a way that guarantees the correctness of its output, but at that point you might as well throw GPT-3 away and write a rule-based system.
The site you tried is a tweet generator, not a question answering site. I prompted GPT-3 with "Bears and beards are different because" and got...
"Bears and beards are different because they are not the same thing.
Bears are animals. Beards are facial hair.
Bears are dangerous. Beards are not.
Bears live in the woods. Beards live on your face.
Bears eat people. Beards do not."
But my original point was mainly that this field is moving fast and the the old school NLG companies (I created one back in the day!) are toast.
Yes it is true that bears are animals.
No it is not true that, as GPT-3 said, "There aren't any on the ground"
It's younger sibling DALL-E is capable of language grounded in images, I expect the next version to be multi-modal as well. On another line of research there's effort to tame the horse (GPT) by attaching a secondary neural net. This can monitor language, topic, style and bias and ensure increased accuracy in tasks by auto-learning good prompts. It would make development of applications much easier because the base model which was super expensive to train can be reused many times while the secondary net is small and fast to train. Other efforts are related to including a search engine on an inner loop, to make the language model able to query large collections. Also, there's an open effort to create a huge text corpus, so far 800GB (The Pile). It improves on the GPT-3 training corpus on some categories that were lacking.
I think it's safe to say the article is way off the current research level.
But I agree there are some applications it is useful for, like education.
I don't know why that would imply that it knows nothing about the real world, unless the data corpus it is trained on likewise bears no relation to reality...
> Sometimes the results produced by GPT-2 and its ilk are quite startling in their apparent authenticity. More often than not they are just a bit off. And sometimes they are just gibberish. As is widely acknowledged, neural text generation as it stands today has a significant problem: driven as it is by information that is ultimately about language use, rather than directly about the real world, it roams untethered to the truth. While the output of such a process might be good enough for the presidential teleprompter, it would not cut it if you want the hard facts about how your pension fund is performing. So, at least for the time being, nobody who develops commercial applications of NLG technology is going to rely on this particular form of that technology.