[ironrabbit]

Is there any evidence that GPT-3 responses are edited/filtered before being returned to users? My understanding is that some GPT-3 responses are annotated post-hoc, and this data is used to fine-tune later versions of GPT-3 (InstructGPT). This article seems extremely misleading.

[ImPostingOnHN]

it seems there is evidence that GPT-3 is being overtrained in response to well publicized bad inputs, without regard to the generalizability of the PR-driven spot-edits, which is what the article describes