I'm developing an implementation of what I call Hydra – Multi-Head Prediction Embeddings [1], which I believe represents the next evolution in transformer architectures.

[1] https://www.researchgate.net/publication/381009719_Hydra_Enh...