[halleyscomet]

i made this for a few reasons: a) to help understand how to make llms produce less slop. b) ironically, this also helps me understand how to improve detection. comparing multiple versions of the same text can be very helpful. the "ai-tells" become more pronounced and easy to spot.

as a bonus, it allows me to assess current sota landscape in ai-detection, in a world that is going to become more and more reliant on verifying the authorship of any published content.