4 ms·

And if you believe nothing ChatGPT says then you simply suffer from the opposite bias. The dilemma is that it is mostly right, most of the time, isn't it? Almos

• tripletao 1y ago
  "Mostly right" is the most insidious option, because the frequent true statements or implications will lull you into believing the infrequent false ones. For Fisher-Yates, if the RNG is good then the output permutation is independent of the input permutation. That's not true for your proposed algorithm, and neither you nor the the LLM has analyzed to determine how quickly the sensitivity to the input permutation decays. That sensitivity will have complicated structure, so it's a nontrivial problem in cryptanalysis. The problem where the low bits of old, not cryptographically secure PRNGs were particularly non-random is real. That would be a problem for any algorithm using it though--imagine if your riffle took rand() % 2. The resolution (use the high bits, not the low bits) is the same for any algorithm. It's possible that your proposed algorithm is less sensitive because the much greater number of RNG calls makes it somehow average out, but neither you nor the LLM has analyzed that, because the residual structure is again a cryptanalysis problem. Any potentially adversarial application should use a CSPRNG, moving the cryptographic security into that existing, well-studied block. So the problem you're trying to solve doesn't exist in practice. The LLM can probably be prompted to tell you that; but since you stated that the new shuffling algorithm was your idea, it worked to find reasons it was a good one. I don't think this is a good use of LLMs, but if you're going to then you have to avoid any suggestion that the new idea is yours, to avoid triggering the sycophancy.
• kerkeslager 1y ago
  > And if you believe nothing ChatGPT says then you simply suffer from the opposite bias. No, you're making a logical error. It's not that I believe the opposite of what ChatGPT says, it's that I don't believe what ChatGPT says based on ChatGPT saying it alone. If ChatGPT says your algorithm is right, it's not that I automatically believe your algorithm is wrong, it's that I believe I still don't know it's right because ChatGPT is shit at being right. > The dilemma is that it is mostly right, most of the time, isn't it? Almost like... people? I mean, yeah. Which is why I don't believe things just because some random person says them, either. I tend to assign confidence intervals to beliefs in my head, and even and expert saying something doesn't make me super confident in it. I tend to believe things strongly if I can understand the hard evidence for them, not because any person, even an expert, says it. With security, we have tens of thousands of examples of unproven security assertions being proven wrong later with disastrous effect, so the standard of proof is very high, and ChatGPT saying it is absolutely nowhere near close enough to proof. The only acceptable proof with a security algorithm is mathematical proof. > If ChatGPT says things that are empirically provable, then it's not wrong, is it? ChatGPT also says things that are empirically disprovable, so basically you always have to go to empirical proof to tell whether what ChatGPT has said is true, making ChatGPT next-to-useless if accuracy is at all a concern. Just go to empirical proof and skip the pointless ChatGPT step. And accuracy really, really is a very big concern in secure random shuffling algorithms, so you really really need to empirically prove that they work which you really really have not done because you stopped thinking when you trusted ChatGPT. Your shuffling algorithm is not proven, and very likely is not secure.