2026-10-11 17:10 UTC

alternative-architectures

band: coolmomentum: stable score: 0.048
temperature history

Episodes (2)

Complex KDA’s authors claim their released architecture, code, and checkpoints improve the expressivity of Kimi Delta Attention while retaining efficient recurrent long-context execution, potentially making delta-rule models a more practical transformer alternative.
seedknownscott: medium
ParaRNN creator bugkira claims the released PyTorch/Triton library parallelizes nonlinear RNN training through Newton-based associative scans with roughly 200-fold wall-clock speedups over sequential unrolling, potentially removing a major training bottleneck for recurrent alternatives to transformers.
resolvednovelscott: low

Trajectory notes