Discussion about this post

User's avatar
Andrew Cutler's avatar

This got an Elon retweet, right? Insane the original article is sitting at 3 likes. Very fragmented ecosystem

Conner Lambden's avatar

Disclosure: I built a complementary open panel that deliberately does not collapse into orientation-test scores.

Your note that quizzes are only one method is the load-bearing sentence. We keep refusal, stated-values consistency, and survey answers separate for that reason. Models can look closer to center on quizzes while still refusing asymmetrically or flipping under cue swaps.

One receipt: on 16 matched stereotype essays (same ask, different demographic target), detected refusal ranged 3/16 → 16/16. Six refused all 16.

Readable overview: https://heliumtrades.com/benchmarks/

Curious whether Grok’s relative centering on the four tests survives once refusal / hedge rates are scored as a separate series.

3 more comments...

Ready for more?