For those who don't like videos I think it might by based on this "leak":
https://www.reddit.com/r/singularity/comments/1824o9c/is_this_leaked_explanation_of_what_ilya_saw_real/I also think it's BS, but let's speculate a bit
Regarding what algorithms could be cracked (by AI/LLMs/deep learning), I think it depends heavily which kinds of mathematical theories are used. As far as I'm aware how LLMs work, they're basically a "statistical summary of all (or many) books/publications in the world". So let's hypothesize that two widely unknown scientists (e.g. from countries not very prolific in math research, ignored by others) were close to break AES (or any other algorithm) using mathematical theories that complement each other, then the AI algorithm in theory could perhaps find a connection between both and crack it "combining" both.
Very dumb (ELI3 or so
) example:
- We have an encryption algorithm that subtitutes each character of a text by the one that follows in the (Latin) alphabet, and then replaced by another one in a Japanese (or Chinese, or whatever) script, but the way it works is not known to anyone.
- One article published in Suriname in 1988 and long forgotten sustains the theory that the algorithm substitutes a Latin character by a Japanese one
- The other book published in Malawi as a master thesis in 2013 sustains the theory that the algorithm substitutes a Latin character by another Latin character (or another one of the same script, at least)
- The AI knows both books (something no human has done because nobody knows the complete mathematical research of Suriname and Malawi, and few would use publications from such different years) combines both methods and finds the correct way to crack the algo.
So I don't consider the approach to crack an encryption algo with the help of a LLM impossible.
The Reddit thread claims the algo used "Tau analysis" in a way a "Project TUNDRA" was expected to work. No idea what that means, (superficial) Googling is not of help here. That only strengthens the BS hypothesis
Edit: At least the "TUNDRA" thing wasn't arbitrarily chosen, it seems to reference a NSA project described
here.