1/
There's a new kind of AI researcher showing up.
Not someone racing to build the next giant model.
Someone who takes a model that already works and asks: how do I make this small, fast, and cheap enough to actually use?
6/ Still experimental and early-stage, but definitely worth watching for people working on:
1. AI infrastructure
2. CUDA kernels
3. Compilers
4. HPC
5. ML systems
would you write GPU kernels in Rust instead of CUDA C++?
#GPUProgramming#SystemsProgramming
1/ NVIDIA Labs(@nvidia) just released cuda-oxide - an experimental Rust-to-CUDA compiler.
You can now write CUDA GPU kernels directly in idiomatic Rust.
No C++.
No DSLs.
No foreign language bindings.
https://t.co/LzF9GWLpUg
#CUDAProgramming#SystesProgramming
5/ The bigger trend is interesting:
AI infrastructure is increasingly becoming:
1. Compiler-driven
2. Kernel-driven
3. Memory-bandwidth constrained
Languages and tooling around GPU kernels are becoming strategically important again.
#MachineLearning#Cuda
1/ 🚀 I’ve started publishing on Substack!
If you're into AI: algorithms, research papers, benchmarks, datasets, agents, and more, you’ll love it.
https://t.co/x4P0zcUS3z
1/ Not all progress in AI is generative.
VL-JEPA serves as a strong reminder of that.
I recently explored the paper “VL-JEPA: Vision–Language Joint Embedding Predictive Architecture,” which takes a fascinating non-generative approach to multimodal learning. 🧵
4/ The broader takeaway for me:
We don’t need to generate everything to understand everything.
Predictive representation learning could enhance efficiency, stability, and generalization as multimodal systems scale.