If a frozen model invents novel math, I expect it will use terminology which is comprehensible to itself in not too many inference steps, which also likely correlates to being at least somewhat comprehensible to humans who have the necessary background.
Which, to be clear, will at some point be zero humans. But I think you need to have models learning from other models to get that effect—models won’t stray too far from the ontology they know only by in-context learning.
If a frozen model invents novel math, I expect it will use terminology which is comprehensible to itself in not too many inference steps, which also likely correlates to being at least somewhat comprehensible to humans who have the necessary background.
Which, to be clear, will at some point be zero humans. But I think you need to have models learning from other models to get that effect—models won’t stray too far from the ontology they know only by in-context learning.