Episode 33: Tri Dao, Stanford: On FlashAttention and sparsity, quantization, and efficient inference
Generally Intelligent
Exploring Alternatives to Attention in Language Models
Approaches like recurrent neural networks and state space methods have been used in the past, but they may not work well for language models./nModifications were made to improve the efficiency of models with more multiplicative interaction, resulting in models with fewer attention layers./nAttention may become unwieldy or a bottleneck when dealing with long sequence lengths./nSome people are skeptical of alternative architectures and prefer to stick with proven recipes using big transformers and lots of data./nAs a researcher, taking contrarian bets and exploring alternative architectures can lead to promising results, even if the success rate is low./nBeing open to different approaches and adjusting quickly can help researchers stay ahead of the competition.
Remember Everything You Learn from Podcasts
