I don't think we've bottomed out on what we can do with embedding models. They're these tiny models that absolutely rip on modern cpus with 8 bit int optimizations. Like in my app we can say pretty definitive things about hundreds of millions of places in the world on retrieval tasks on regular hardware.