Ep 45: Sakana AI and NVIDIA just showed how extreme sparsity in feed-forward layers can deliver real GPU speedups without retraining.

Models & Agents Sakana AI and NVIDIA just showed how extreme sparsity in feed-forward layers can deliver real GPU speedups without retraining. What You Need to Know: Sakana and NVIDIA released TwELL, a sparse inference approach that uses simple L1 regularization plus custom CUDA kernels to hit over 99% sparsity in LLM feed-forward layers. The result is measured 20.5% faster inference and 21.9% faster training with negligible quality loss. ... AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.