hearing lots of misconceptions about grass’s role in AI training. the protocol is built to retrieve realtime knowledge from the web at scale, mainly for the purpose of informing a model in real-time of new developments. new information appears on the internet, grass “learns” this information, and supplies it to the model as it’s doing some task. it doesn’t matter if this new knowledge was created by a human or a bot - it’s the semantic information that matters. we’ve been talking for a long time now about how scaling limits have been saturated for text-based models (hence the RL rabbit hole / importance synthetic data ie distillation), but multimodal models are still far far far from reaching similar saturation. today you cannot find someone who can generate 10 million hours of “synthetic” video data in an economic and useable way. tldr - realtime is where the network will accrue most value long term, but right now multimodal capabilities are mostly unmatched & it would be very silly not to use this.