| Tension between machine learning and cryptography has been studied since at least the late 80's. Example classical result: Kearns and Valiant, "Cryptographic Limitations on Learning Boolean Formula and Finite Automata." Roughly, they show that efficient learning algorithms for Boolean formulae could break RSA --- so probably such learning algorithms do not exist. Recent breakthroughs are quite exciting: see page 4 of "On One-way Functions and Kolmogorov Complexity" which, for the first time, bases the security of private-key crypto on a natural and fundamentally computational problem instead of something from number theory. They get a solid connection to universal extrapolation via Impagliazzo, Levin '90. Currently, these results operate for overly-restrictive learning models. I hope we will slowly improve them to obtain win/win conditions: depending on the setup either strong cryptography or strong learning is possible, but not both. My point is: these are not fads. We have been working very hard on fundamental mathematical connections between learning and crypto for decades. Progress is slow, but meaningful. Those papers are theoretical, but modern crypto emerged from a purely theoretical and embryonic "computational number theory". I am also not a fan of the article. It would have been better if the author had chosen to define, for example, "one way functions" and the "PAC learning model" and explain why these theoretical constructions are not quite a match for large language models and blockchain-based protocols. However, I believe that bringing computational learning theory --- which is inextricably linked to cryptography --- closer and closer to LLMs is a promising direction for future research. Links to Cited Papers: https://www.cis.upenn.edu/~mkearns/papers/crypto.pdf https://eccc.weizmann.ac.il/report/2020/052/ |
My use of “fad” was primarily referring to the author’s reference to “crypto”, unfortunately these days more often short for “cryptocurrency” than “cryptography”. In context that is very clearly the case here:
> Blockchains are the right way for software to eat money
I think the link between cryptography and ML is fascinating, but cryptocurrency is absolutely a faddish cesspool of speculation and fraud.