He mentions that he posed to O1 the same challenge he posed to a previous GPT (which he also previously blogged about), so I am wondering how much O1 benefited from potentially "seeing" this discussion in its training set (which probably contains a very well recent snapshot of the world wide web).
I wonder if those responses could already be influenced by the fact that the cutoff for some of the models out there was indeed 2023 and people wrote about it all over the internet.
