this says it drops parameters by 25x using
   retrieval: [1]https://www.deepmind.com/publications/improving-language-
   models-by-retrieving-from-trillions-of-tokens
   if we're stuck copying work we bump into, then it makes sense to figure
   out what's cited that paper (or bump into something else)

References

   1. https://www.deepmind.com/publications/improving-language-models-by-retrieving-from-trillions-of-tokens