This was an attempt to use Zeta v098, a gpu BestFirst-search with thousands of
parallel threads, with anns, but one single thread is too slow for ann inference.