Tootfinder

Excited about the new xLSTM model release. There are many well-though designs compared to transformers: recurrence (which should allows composability), gating (like Mamba & LSTM which is based on, which allows time complexity independent of the input size), state tracking (unlike Mamba & transformers). For now, these advantage aren’t apparent on benchmarks, but most training techniques are secrets, and the recent advances of LLMs evidenced that they matter a lot.