more efficient, but unable to learn semantic representations. Recent efforts like those of the SPLADE models are changing that, yet they still can't quite manage the semantic performance of dense vectors

Realistically I think the near future best performing retrieval systems will use a mix of both in some hybrid retrieval system