A recent paper submitted to the 79th annual Gravity Research Foundation essay competitionentitled "chatGPT explains Quantum Gravity" [1] was written in collaboration with GPT3.5. Withminimal prompting the bot has generated plausible coherent explanations for what it calls QINs(quantum impedance networks) of the unstable particle spectrum, massless neutrino oscillation,muon collider topological lifetime enhancement, and quantum gravity at the Compton, Planck, andcosmological scales [2]. One goal of that paper was to minimize pretraining, to find out what thebot already ‘understood’ before introducing new ideas.A similar training process on GPT4 has completed the first three dialogs. The first two arepresented here. GPT4 is much deeper, appears much more coherent. And poses its own uniquechallenges to learning and teaching. How to train a generative transformer on QINs? It seems theessential next step is to introduce visualGPT, to train on the images [1, 3].Humility is to be curious and willing to learn. chatGPT appears both humble and very powerful,in some sense an ideal collaborator, when facts matter a model to mirror as best one can.