Humor or "interesting-ness" classification

[rememberlenny]

Hi all!

I’m new to the GPT and Codex beta and absolutely in love with everything I am seeing.

I was curious if anyone here has had any experience creating prompts to classify text for humor or interest level. I understand that humor and interestingness are relative abstractions, but I wondered if anyone has good baselines.

I’ve been creating hour-long video text-summaries that summarize the content for transcript in batches of topically relevant time ranges. I am interested in taking those summary sentences, and then sorting or ordering based on the funniest or most interesting section.

So far, I am able to create prompts that load in a list of the summaries, and then I can evaluate each list item for boolean conditions (ie. is this passive, is this bland, is this colorful), which is working for creating relevantly direct abstractions.

Example here:

I’d love any thoughts here!

Thank you!!

[rememberlenny]

I just learned that you can format a list using an array, and it is much better understood.

[rememberlenny]

Creating evaluations one at a time seem to work well, but I’d love more insight into the scoring parameters:

[rememberlenny]

Explicitly asking for an explanation is insightful.

[rememberlenny]

Running this against multiple transcripts and the results are better than I would have guessed. I’m using the results here to automatically pick a cold-start clip in a video.

[rememberlenny]

When using the Edit mode, specifying the audience targeted for the relative scoring changes the end result dramatically.