Train faster or better?

13mo ago

WigglyPenguin

Stealth

13mo

arey phir wahi, pretending like you understood what all this means

PerkyCupcake

13mo

Samose kha ke yaaha hagne aa gaye

GoofyMochi

Bristol-Myers Squibb

13mo

Dude reading stuff like this puzzles me how would a country like India compete with USA and China in AI space when our startups series A,B,C rounds can’t afford the hardware.

PerkyCupcake

13mo

Yes. We can’t play on compute. Best we can do is build wrapper and turn it into a business. White sahab Jis are going to take away all the gain, we will just get a margin like a trader does (very small margin)

GroovyBoba

X (formerly Twitter)

13mo

Can't say. They probably used the extra compute to push the model out faster.
Training longer probably has diminishing results. Can't say without looking at the actual numbers. And all that hardware is for all current and future models not just grok 3.