Totally agree. Wasn’t trying to sound like I was bagging on nostr, I think it’s a perfect fit for machine-to-machine, which is a different design space than decentralized social network.

That's fair, "unusable" is probably the wrong word. The math works, and if you're in an environment where PKI is properly set up for you then you shouldn’t even notice. The friction is more around trust and portability. And from an adoption perspective, unfortunately most normal people don't care or think about trust authorities or attestation, let alone managing identity through cryptography. The machines though, just tell them the mechanics.

The things that make PKI unusable for humans disappear when the users are machines.

Shout out to all the "it's just next token prediction" folks this morning.

You know a new platform is on the verge of breaking out when all of the "this is nothing special, I could have built this" comments start to increase.

Asked Kimi to compare repos…. Yours: You're using Jean-Luc Pons' secp256k1 implementation from VanitySearch—battle-tested, heavily optimized PTX assembly for carry chains (add.cc.u64, addc.cc.u64, mad.hi.cc.u64). That stuff is tight. His: Clean handwritten Jacobian math with 32-bit words and standard CUDA. It's readable, but it's doing ~2x the work per operation. That is indeed what I started with FWIW https://github.com/rossbates/rummage/blob/main/src/CPU/SECP256K1.cpp

Most of the optimization work focused on batching. Run nvidia-smi during your job and see it it’s actually using 100% of the GPU

Built something similar, getting 42 million/sec on an 3070 and 170 M on an H200 https://github.com/rossbates/rummage/blob/main/docs/PERFORMANCE.md

Totally, I’m running openclaw with a variety of self hosted apps and it’s incredible. The quality is so dependent on Opus though.

What model are you based on?