Keep on to blur preview images; turn off to show them clearly
RT @HuggingPapers: Facebook just released the OMC25 dataset on Hugging Face. This massive collection of over 27M molecular crystal structu…
Co-founder & CEO @HuggingFace 🤗, the open and collaborative platform for AI builders
RT @aiDotEngineer: 🆕 Leadership talks: is all this AI code being spammed out any good? how do we prove any ROI? @yegordb is back with an e…
achieve ambition with intentionality, intensity, & integrity i made: - @dxtipshq - @cognition - @sveltesociety - @aidotengineer - @latentspacepod + @smol_ai
CUDA-L2 uses reinforcement learning to outperform cuBLAS on matrix multiplication. Tested across 1000 HGEMM configs, it beats torch.matmul, cuBLAS, and cuBLASLt AutoTuning on A100. +22% in offline mode. +28.7% in server mode. LLMs are now tuning kernels.
📄 Paper: https://t.co/8u9Vmul5KX 🔗 GitHub: https://t.co/xijapZLAgY
Ted Underwood et. al. also trained a GPT-1914 which might well still the largest historical model to date (nearly 800M). https://t.co/wSA1WCmz2D
Artisanal baker of reasoning models @pleiasfr
i had thought we were past peak webpack but apparently not, what happened? (vite not shown due to data error but yes has overtaken webpack)
achieve ambition with intentionality, intensity, & integrity i made: - @dxtipshq - @cognition - @sveltesociety - @aidotengineer - @latentspacepod + @smol_ai
Example of historically grounded generation (model has zero idea who Trump is and revert to early 20th century). I see model is actually still popular on HF. https://t.co/1ye5VEGerd
Ted Underwood et. al. also trained a GPT-1914 which might well still the largest historical model to date (nearly 800M). https://t.co/wSA1WCmz2D