Note that that kind of research was only possible on the (smaller) open models. When the new models are proprietary ones then outsiders can't do that kind of an

Note that that kind of research was only possible on the (smaller) open models. When the new models are proprietary ones then outsiders can't do that kind of analysis. If they had regenerated the tokenization dictionary to match the training data instead of reusing an older set then this wouldn't transfer from GPT2 to ChatGPT.