well, with all respect to Thinky, there's a damn good reasonTrinky getting up to speed on training a V3+ tier model is a great win for…
beyond mHC: scaling the residual stream by effectively 16xXiangdong Zhang: Hyper-Connections never fade. They evolve.To push model intelligence further, we need to scale residual streams wider…
I love Ryan Reading this now is deliciousThough J.S. Helberg's Claudeslop is still sweeterSorry guys, you haven't manifested a sublime irreplaceable power yetkeep trying though! that's…
> Kimi K3 pretrain is around halfway between Opus 4 and Opus 4.5. So ~10 months behind Anthropic.That is very likely true for irreducible compute costAnthropic's…
…does Moonshot even have a blogpost about Kimi in Chinese? It's a bit strange that we have English-language distribution channels and I don't see any easy…
RT GDPThere is one X account that is writing about existence and rapid rise of Chinese data labelling industry. He has just 15K follower. Please go…
Holy shit Meta has cookedthey have a Fable/K3/Sol level model!…at least functionallyAlexandr Wang: Muse Spark 1.1 is now available on OpenRouter!This was highly requested by developers!…
we don't know enough about infra side, but I think Kimi's margins are a bit fatter than DeepSeek's, they are VERY supply-constrained, and objectively they could…
just like thatCharacteristics of a King momentBill Bishop: So how can he still be friends w Xi after what he says the prc just did in…
this is just trueFurthermore, with full supply chain access and investable high tech companies, I think China could have been a $32T economy (nominal; maybe $55T…