π“Ÿ host1.bitchan.one

𓀐 @aESDJeRA

⏱ 5d seenβ›“ 0πŸ‘ 0β§– 1

filter [-]

languages:onother

it means the relationship between each concept is described with less subtlety the less bits you give it, so concepts and relationships that were comparable but one was slightly more related than another will be squashed to the same weight. This may or may not make the output worse for a given case, but overall always feels like an IQ reduction in model ability. How much worse it gets in real-world performance is on a kind of exponential curve where going from eg 32-bits to 16 or even 8 doesn't hurt much, but going to 2 or 1 basically scoops out the prefrontal cortex.⏱ 5d↩