[stopachka]

Update: I took a corpus of personal chat data (this way it wouldn't be seen in training), and tried asking it some paraphrased questions. It performed quite poorly.

[abraxas]

Which models did you try?

[stopachka]

Claude Sonnet 4.6