NVIDIA's Gated DeltaNet-2 splits erase and write gates in linear attention, aiming for cleaner memory updates and stronger benchmarks without attention costs.