This pic goes so hard with the military and intelligence community in the crowd at the Citadel Military College This event is sick so much happening here I asked for permission to tweet this pic

I don't even see the nn.Modules anymore. All I see is memory-bound, memory-bound, compute-bound, memory-bound, host-overhead bound, register spill fiasco, memory-bound, shared memory bank conflict galore, compute-bound

Fiy thats what a cunty zipper looks like Dylan Patel

Met great people in Seoul 📸 Simon Guo 🦝

A major mistake I made in my undergrad is that I focused way too much on mathematical lens of computing - computability, decidability, asymptotic complexity etc. And too little on physical lens - energy/heat of state change, data locality, parallelism, computer architecture. The

To predict means to compress

Releasing INTELLECT-2: We’re open-sourcing the first 32B parameter model trained via globally distributed reinforcement learning: • Detailed Technical Report • INTELLECT-2 model checkpoint primeintellect.ai/blog/intellect…

Cohere has a Jax team

If you are based in Zurich (or anywhere rly) and write code for ML accelerators (including cuda/rocm) HMU

FP4 numbers are like quantum systems. You know that they exist and have one of those 16 values, but the moment you need to observe them, they immediately attach to something (8 bits) and lose their original meaning.

deep learning compilers when you give them a model with dynamic shapes and control flow

What is the crime? Smuggling a pytorch_model.bin ? A succulent safetensor?

ML hopecore is how the homies Lucas Beyer (bl16) Alexander Xiaohua Zhai ride or die together

Shaowei from our infra team actually wrote about the decisions we made on the Kimi K2 architecture. zhihu.com/question/19271… I suggest reading it with Kimi K2 as your translator. :)