[ACS_Solver]

Kryptos K4 seems to me like a potential candidate for AI systems to solve if they're capable of actual innovation. So far I find LLMs to be useful tools if carefully guided, but more like an IDE's refactoring feature on steroids than an actual thinking system.

LLMs know (as in have training data) everything about Kryptos. The first three messages, how they were solved including failed attempts, years of Usenet / forum messages and papers about K4, the official clues, it knows about the World Clock in Berlin, including things published in German, it can certainly write Python scripts that would replicate any viable pen-and-paper technique in milliseconds, and so on.

Yet as far as I know (though I don't actively follow K4 work), LLMs haven't produced any ideas or code useful to solving K4, let alone a solution.

[thephyber]

Yeah, you would suspect that the individual elements of solving K4 exist in some LLM, but so far the LLM slop answers are just very confident and very wrong.

My biggest complaint is that the users aren’t skeptical. They don’t even ask the LLM to verify if the answer it just generated matches the known hints from the puzzle artist. Beyond that, they don’t ask it to verify whether the decryption method actually yields the plaintext it confidently spit out.

I’m super impressed with Claude Code, though. For my use case, planning and building iOS app prototypes, it is amazing.