Let’s be frank, the internet was taken over by LLMs and AI slop. As time goes on, it is increasingly hard for humans to distill truthful information from the oceans of LLM-generated garbage. Unfortunately, these texts often look legitimate due to elaborate grammar and a convincing way of prose, while they often contain idiotic mistakes and produce nonsense information (for juicy details, check out a great write-up by Kyle Kingsbury titled The Future of Everything is Lies, I Guess: Information Ecology, especially the Information Ecology section).

Yet, due to the way LLMs work, some simple tasks, such as reading analog clocks or solving simple riddles in one word, are out of their reach. Once, I was curious whether LLMs can decode a weird kind of text–the one encoded with ASCII-art fonts. Trying this, I got truly amused by the results, and a little bit surprised. After all, the way it handles ASCII art is very revealing about the LLM’s inner workings.

While implementing an ECDSA request signature algorithm I was writing unit tests for it. The tests should work in a reproducible way: we generate the same private key each time, sign constant data with the same random (which is needed for ECDSA) - we get the same signature. It seems quite straightforward.

Using reproducible random stream

Looking at the documentation, we see that we should pass elliptic curve parameters and the random stream we want to use.