This folder contains research on RSA/DH/ECC cryptography.
This is hand-"written" notebooks, NOT AI-generated (except for one). Chats with AI (GPT-OSS, "critical" "thinking" model) prove it was generating fiction (eg, talking to me about lattice theory which is useless for practical work with DLP). As well as one or two notebooks, I asked to generate (ends with "-ai").
The bottom line: fastexp overhead is non-monotonic. Asymmetric encryption has "cat in a bag" security - defender has to spend as much energy/resources as attacker to ensure security of a given pubkey.
No free lunch, but illusion that there is.
Statistical estimates of "overheads of fast exp for DLP", which NIST and others (eg Blackberry) are using as an excuse, are not applicable to non-monotonic problems; made-up useless soviet 'proofs' there, coming from worst european ideas.
Additionally, ECC is even worse than DH, speaking of crypto - polynomials introduce more holes on top of illusions. Concretely - in one of the html notebooks here (I wrote by hand) - you can notice that naive billinear map shows that sign of encrypted number can be guessed sometimes, since a bit is flipping faster when you progress with decimals. There are trivial statistical dependencies, even on the strongest bitcoin curve.
ECC creates jobs for cryptoanalysts simply.
DLP Note on key restauration with deterministic slowexp (the only inductively proven way): here I show that local monotonicities can be exploited. perfect power (logN speedified with memoization) is used to skip through local monotonic interval.
P.S. I accidentally re-invented recently invented algorithm for fast perfect power (logN, I used divide and conquer), aka polynomial DLP-solution for non-cyclic integers. When I asked AI about it - it refered to David Harvey (2019).
About AI
AI was a bit useful in my "abstract machine research", not presented here (spent a month or so), but here, hand-"written". It, at least, gave nice formulations and was able to list challenges (with hallucinations though). But it gave me very bad code! Anything novel - u have to babysit it which is a trick to collect data simply, with no permission.
I think chat based on pure-search (rather than NN) would work better than GPT non-sense. Can add transform-grammars for reasoning (and context awareness), long-range masks for long-range dependencies/dropouts, and PoW to emphasize relevant truth (certify with energy). I have concept presented somewhere (called YaQui Search). Don't even ask AI about it - I asked it if ngrams would replace it - it told me yes (GPT-OSS, GPT-5), while in reality transform grammars (more complex framework) is required.
Energy conservation (or why not consumer/producer symmetry) also applies to TRNGs, but proof requires to replicate whole TRNG pipeline.
And without proof - cannot improve crypto
I partially replicated TRNG pipeline in a tool. PRs are welcome.
Bitcoin TRNG audit tool is here: https://github.com/dk14/crypto/tree/main/chats/btc-audit
Donations BTC: bc1qekvmkczge3hxrvwdf2lj3yyvgjnparn3fdf9lg