>>106561614
>>106561622
my experience with kimi-k2 is it's absurdly overdramatic, like some kind of marvel movie script that a 12yo would find epic
for example, characters don't argue with you, they backflip through the air doing somersaults and land with a knife one cm from your eyeball. i was testing an argument prompt on every model and the kimi-k2 one was fucking absurdly hilariously over the top.

also it outputs 4x more than other models. whichever chinks they RLHF'd it on definitely think that more text == automatically better. not surprising from the race in this video who have some kind of pathological peasant concern with having "the most" of everything

the only interesting chinkmodel is GLM 4.5, it is quite creative and "integrated" in some ways. it almost makes me think there is some sort of unexplored advantage to a model being small/dense - perhaps it can access more neurons more easily