Literally just yesterday I asked why this wasn't being done:
https://news.ycombinator.com/item?id=47195212
> "Models have unlocked advanced collaborative reasoning, yet they remain shackled by the inefficiency of discrete text communication, which imposes significant runtime overhead and information quantization loss."
