RWKV is an RNN with transformer-level LLM performance. It can be directly trained like a GPT (parallelizable). So it's combining the best of RNN and transformer - great performance, fast inference, saves VRAM, fast training, "infinite" ctx_len, and free sentence embedding.
Curious + 1. But I guess suitable scaling up method is unknown. It can be layers, widths, gating, shortcuts... Probably need some empirical experiments and theoretical insights.
How about try to train a bigger model using RWKV? 1800B?