But embodiment being a bottleneck could indicate that it's a data/training issue, rather than an architectural issue. Multimodal training data improves GPT-4 already, but that's still very little data compared to growing up to a full human adult. There are still many things to try.