5h agoSakana AI launches DiffusionBlocks, training neural networks block-by-block to cut memory requirements to a single block's footprintIt matches standard training performance across ViTs and LLMs.SentimentSentimentPos65.2%Neg34.8%Positive users praise Sakana AI's DiffusionBlocks as a clever novel way to train networks block by block for memory efficiency while negative users dismiss the approach as unoriginal gradient hacks or marketing fluff.52 comments with sentiment. View comments.