>Because the honest incremental results are still useful: the negative result on the dictionary matcher, and the failure mode of the autoencoder on out-of-distribution images, are real findings that took real testing to get. Better to document what's true than claim more than what's built.

Reads like LLM cope.

It's a neat idea, could be potentially impactful given the prevalance of such images, but definitely a long way to go.