This is kind of a stupid argument. How about make a slightly stronger claim like "models won't break symmetric crypto" ? I mean, language models aren't even trained to break symmetric crypto. There is not good reason to think they will. It seems possible to train a large model to do it though.
To train a large model to do what? Break AES? How would that work?
Train on plaintext, ciphertext -> key.
I suggest you do this experiment yourself. You can try model architectures as big as your computer can fit. It won't work because these algorithms are designed to have no patterns at all. People have already tried.
More classically, you can try feeding the problem into a SAT solver. People have tried that too. Doesn't work - it just grinds until you run out of memory or patience, finding no useful results.
You can also try doing it by hand and see if you get anywhere (you won't). People have tried.
This is an adversarial problem. The problem is literally designed to be resistant to all kinds of analysis. That's the point. Even real attacks, like SHAttered (different kinds of attack on a different kind of algorithm) manage to find conditions where the probability of finding a solution is raised to 2^-70 or so, and then they let it grind on their biggest compute clusters until they find one. And that problem (finding a collision in a cryptographic hash function) is one that's especially amenable to grinding. If you're mounting a known plaintext attack it's unlikely your adversary will answer 2^70 encryption requests for you.
>> It won't work because these algorithms are designed to have no patterns at all
Every encryption algorithm proposal has this property of being designed to have no patterns....
And some of them still have no known patterns! Even the "broken" ones have only really subtle patterns. SHAttered was only 2^17 times faster than brute force. It took 2^63 attempts instead of 2^80 that bruteforce would take (which was arguably already too low).
Yup, sure. But my claim is that I believe it's better than 20-1 against that a model can break it. The fact that many encryption schemes have later been shown to be fragile and that very very large models seem to be able to things we can't explain well, and that you can create enormous amounts of training data for this problem makes my claim not so far fetched.
Try it and see then. If you can reverse double-SHA256 with some constraints, you can print yourself infinite bitcoin. If you can reverse whatever they use for public keys (RIPEMD and some elliptic curve) you can steal yourself infinite bitcoin.
Yes, indeed. And if I really had a good idea for an architecture, and millions to blow trying different ideas... I would be doing that and not on hacker news.
LLMs aren't literally science fiction.
Well, again, i'm not talking about a language model....
And, just because what I'm saying isn't especially likely to work, it's not obvious that it cannot. Very large models are doing all manner of things that very smart people thought were not possible just 6 or 7 years ago.
It's unclear what you are talking about then. Because the idea of training "ciphertext -> plaintext" for language models is absolutely bonkers, so what are you suggesting?
I can't tell if you are serious at this point. I literally say, twice, that I'm not talking about a language model. And I also say the model would predict the key...not the plaintext.
Have you ever built a NN model? Have you ever broken a crypto system, even a small one?
And I'm asking you to describe the model.
Ouput: 128 logits. Input: maybe 10 samples of plaintext,ciphertext (using the same key), so maybe a 2560 length tensor. Loss function: binary cross entropy on the true key bits. Architecture: anyone's guess.
If you were in a place to debate this, you would have known the above (or something similar) is what I was suggesting when i said train on plaintext, cipertext -> key, and you'd have some deep mathematical insight as to why no architecture known is likely to work. And you would also know I wouldn't be here talking to you about it if I really had a solid idea of an architecture that is likely to work.
I'm not debating you at all. I'm asking what the model looks like since you've stated (and I've agreed) that a language model wouldn't work.
I think it would make sense to explain how a theoretical model could do better than SAT. Otherwise, is the idea here just "magic is possible"?
Yes, "magic is possible" if you defined "magic" as "very large models approximating functions in a way that people didn't think would work".
Current SOTA language and vision models, or models used to predict protein shapes are magic by the standards of 2016. As for why could it be better than a SAT? Why couldn't it be? Models are better than deterministic, logically written software for lots of situations. You can create infinite training data for this problem. The number of humans that work on encryption is tiny. The idea that because humans haven't figured out how to break some encryption schemes it can't be done is kind of absurd.
Agreed that many of the articles claims are a bit weak. One point is reasonably strong though: symmetric crypto may not be breakable (battle tested).
It probably isn't. But if you laid 20-1 I'd bet a large model will break an industry used standard within 10 years. That's a loose framing of a bet, but I think you get my point, even if you think my numbers suggest too much optimism.
[dead]