We train large language models on large corpuses of clean text. That might be good enough for many things, but I think this way of training is very limited. This idea came to my mind when I was looking for a quote from John Adams from a letter to his wife, Abigail. The original manuscript found on the Massachusetts Historical Society is

which reads

I must study Politicks and War that my sons may have liberty to study Painting and Poetry Mathematicks and Philosophy. My sons ought to study Mathematicks and Philosophy, Geography, natural History, Naval Architecture, navigation, Commerce and Agriculture, in order to give their Children a right to study Painting, Poetry, Musick, Architecture, Statuary, Tapestry and Porcelaine.

This is far more revealing than the clean, polished version we usually see.

I must study Politicks and War that my sons may have liberty to study Mathematicks and Philosophy. My sons ought to study Mathematicks and Philosophy, Geography, natural History, Naval Architecture, navigation, Commerce and Agriculture, in order to give their Children a right to study Painting, Poetry, Musick, Architecture, Statuary, Tapestry and Porcelaine.

The edit—crossing out Painting and Poetry for Mathematicks and Philosophy is his reasoning made visible. The clean text gives us his conclusion; the messy text shows us how he got there.

This highlights a fundamental limitation in how we train AI. Our models learn from the final draft, the published book, and the finished article. They are experts on the results of human thought, but they are completely ignorant of the messy, iterative process of revision, debate, and discovery that created it.