✧ NEW RELEASE ✧ my song "Fluorite" for Gakuen iDOLM@STER is out now everywhere ♡ stream here → nex-tone.link/EwejxAE7K lyrics by やぎぬまかな やぎぬまかな vocals by 七瀬つむぎ 七瀬つむぎ

Visualizing LLM basics: How do LLMs store facts in its weights? Came across a video by Grant Sanderson on how LLMs store facts in its feed-forward layers (ffwd), something that a lot fewer videos / blogs touch on compared to the famous attention layers. However, ffwd is the key

Download pre-compiled, optimized kernels from the Kernel Hub! Battle tested in transformers and TGI, let us know if you use it in other PyTorch projects 🚀 huggingface.co/blog/hello-hf-…

I wish more of the best AI researchers work on advancing medical AI

I really didn't expect Avey to get much notice at all, thank you everyone c:

クオリティ高すぎるのいた

Expectation: "When I get a job as ML researcher, I'm going to push infinite-ctx-length diffusion model that faithfully diffuses via second order expansion of fokker-planck equation, with adaptive guidance on control signal and efficient gradient estimation" Your dataset:

Done some midnight coding and added Avey to BlaGPT it is crazy slow and uses way more VRAM at training. I think it needs some iteration to be practical however it landed well amongst other transformer alternatives github.com/erogol/BlaGPT

this basically replaces tokenization by pooling bytes through u-net into chunks and predict the next chunk, which would also predict multiple bytes/words at once arxiv.org/abs/2506.14761