I love “the product of writing is meaning.” The same kind of statement can be made about LLM-generated software. The product of software development is solutions, not code. Same category error mistaking the finger for the moon it is pointing at.

I wonder also if we might see cyclic attractors and maybe expect that models whose training trajectories fall into them might generalize better.

“So much precision with so little detail” is a really nice concept. Learning from art!

This is one advantage of having the historical memory of authoritarianism, as I have from Brazil: you know that, when the regime is gone, the people who collaborated and acquiesced look terrible in retrospect. It may look reasonable and justifiable now, but believe me, it will age like milk.