I'm not wrong on methodology – but perhaps I did overstate this. The model was fine-tuned on papers describing exactly the thing they then asked questions about. In my experience models remember fine-tuning data precisely, and regurgitate it on demand.

→ View original post on X — @jeremyphoward