[athrowaway3z]

So this is a tangent on a thought i had after reading the title, but it might be a cool idea that I'll not have the time to do anything with so feel free to use it:

Human checkable fingerprints for pubkeys/hashes dont really work. None of the schemes i've seen hold up under somebody willing to spend compute to get a near-enough collision to fool most people most of the time.

But we can take those random bits and transform them and feed them into a seeded image generation LLM, and then have a person remember/compare the deterministic output image.

You might even make the case its the perfect machine to create memorable-2-human image artifact from random data.

[wiml]

There are a bunch of methods for transforming a hash into something that's easier to compare. You've probably already seen the RandomArt thing that openssh uses for comparing host keys on first use. Some apps produce a sequence of emoji for fingerprint comparison. It's a small but fertile little research niche.

I can't offhand think of anything that an LLM image generator would do to improve the process; it'd be an interesting research task. You'd need a way to transform the 256-bit hash into LLM input in a way that would maximize the perceptual difference in generated images. The problem is that it's absolutely critical that two different implementations work the same, which means the spec would need to specify the exact set of model weights to use.