zrm

> Meet in the middle is a specific term referring to an attack on block ciphers. If this was going to be characterized as something other than a pure collision...

> My point was more that if you're already mapping a public-key equivalent (HIP terminology "Identifier") to an IP address (HIP terminology: "Locator"), then you necessarily have some mapping DB...

> Except that mapping HITs to (HI, IP) pairs _is not the same thing_ as a naming system. Unless you're relying on the cryptographic properties of the HIT, anything that...

> Nobody wants to spend a lot of effort building apps for a platform nobody has installed, and nobody wants to install a system level platform that has no or...

Snow is not an anonymity service. It uses virtual addressing but the daemon itself still requires an underlying transport. Receiving packets anonymously requires some device (a privacy VPN, a Tor...

Snow uses a DHT like a STUN server, so yes. Can you say more about the specific network configuration where you find STUN/ICE to be slow? Do you know the...

Not sure if this one is alright (9d058c2096b9f1f300e1ee16f5740a6a0a342917). On machines with high thread synchronization overhead, distributing GGML_TASK_FINALIZE to the thread pool adds about 10% to eval time, so only do...

> I recommend to re-do the analysis with always using --no-mmap and see if there is benefit from NUMA in that case `master`: ./main -n 512 -m models/7B/ggml-model-q4_0.bin -p "I...

Let's see if I can better explain how this is working. `ggml_compute_forward_mul_mat_f16_f32()`: https://github.com/ggerganov/llama.cpp/blob/ac7876ac20124a15a44fd6317721ff1aa2538806/ggml.c#L9802-L9807 Thread *ith* always gets the same rows of the model, and they're several memory pages in size....

> It doesn't seem to work on the windows platform. I didn't write a Windows implementation because the APIs are different, I'm not especially familiar with them and I haven't...