How it works

The team ran a Chinchilla-style IsoFLOP sweep on SpikeGPT models, varying width and compute budget, then compared loss against standard RWKV and Transformer baselines.