Posts by Zamfr@mstdn.social
(DIR) Post #AhPdLj5uqfc42pzeYS by Zamfr@mstdn.social
2024-04-30T11:31:41Z
0 likes, 0 repeats
@futurebird @dr2chase @baldur If the system runs on limited precision, then training and evaluation both run at that precision. And both go faster on optimised hardware.As for precision: these system have internal states with enormous numbers of variables, with every "feature" encoded in many variables (in a somewhat opaque way). It turns out that its better to reduce the precision of individual values, if that allows you to add more of them
(DIR) Post #Ai5PQYcn79xpUySzB2 by Zamfr@mstdn.social
2024-05-20T15:13:14Z
0 likes, 0 repeats
@interfluidity Yes, always has been