[Tenoke]

Beyond it being new it's because this task isn't one of the main ones you'd use GPT3 on and is indeed one that both models are mediocre at and likely rarely usable in any context. The title is just a tad misleading.*

Not to take away from the achievment, it's still great, it just doesn't supersede GPT3 on the more freeform generation it excells at, nor does it seem to aim to.

* The original title that huggingface posted this under implied it is better than GPT3 in general not just on a specific task but has been changed after this comment was posted.