3 ms·

Nice!

• cubefox 8mo ago
  Unfortunately the paper seems to have been mostly overlooked. It has only a few citations. I think one practical issue is that that existing training hardware is optimized for floating point operations.