Relying on self-reported names is a pretty bad idea. Name training is always shallow. Sonnet 3.6 used to call itself GPT-3 (not even 3.5) with some irrelevant d

Relying on self-reported names is a pretty bad idea. Name training is always shallow. Sonnet 3.6 used to call itself GPT-3 (not even 3.5) with some irrelevant distractors in the context, and they clearly didn't train it on GPT-3 outputs. It's just dataset contamination.