John Huang

CIO of LiZen | | HK/China
Opinions are solely my own and do not reflect the views of my firm, also not financial advice, dyor

[John Huang @zzjohnh]

@MichaelAArouet I envy retards who behaves like a 2b param local model, hallucinate data, make up city names, generate an incorrect map, and then post it on X sounding completely confident 🤣

[John Huang @zzjohnh]

@staysaasy It just goes to show, when there’s a will, there’s a way, look at this mofo right here, Chinese lab open source AI for all of humanity, somehow he still manage to find an angle for “fucking china man”

[John Huang @zzjohnh]

@Xianzhong_1953 🤣🤣🤣井底之蛙

[John Huang @zzjohnh]

@notjazii GLM 5.2 is slow and noway that website only use 2% of your weekly limit, so yeah, share the prompt and output, otherwise you're just full of shit.

[John Huang @zzjohnh]

It's a very large country and very fragmented, your experience at different corp size/city will be very different.
Probably I'd say in Hangzhou/Shanghai/Shenzhen/Beijing you are getting close to western capitalistic rule with a strong communist/government authority ceiling, basically below that ceiling you're living in the dream world where everything is perfect, but any attempt to hit/break the ceiling will get you absolutely run-over/crushed.
Lower tier cities are really anything goes, the lower you go probably the more similar to India

[John Huang @zzjohnh]

@aoxzh 买国际版的呗 https://t.co/zf1hhvbUzp ，但是他们算力真的不够用

[John Huang @zzjohnh]

@hongsv11 出乎意料的面对现实呢

[John Huang @zzjohnh]

@JoeAnima 英诺赛科在港股上市的呀？而且全球唯一一个有产能的SiC，其他都是fabless

[John Huang @zzjohnh]

@TeksEdge @OpenRouter the rate limit is insane on glm though, i'm getting like maybe 1/5 call through.
we need alibaba/google or someone to actually host these and make it reliable

[John Huang @zzjohnh]

glm 5.2 我一个付费用户，10个api call，9个给我
"HTTP 429: The service may be temporarily overloaded, please try again later"
我真服了

[John Huang @zzjohnh]

in actual overall cost, cache hit rate is a huge factor as well, Chinese model are consistently hitting 80-90% cache hit rate, while openai/anthropic model usually in low 50%.
Since you can't post train the cache (KV pair compression space is done pre-training), to me this just implies US models are on a less compute efficient architecture vs. chinese ones.

[John Huang @zzjohnh]

平心而论，你就说这内需刺不刺激？

[John Huang @zzjohnh]

国内经济全方位灭杀

[John Huang @zzjohnh]

@jukan05 @grok who's the largest producers of GaN semi? show listed companies only, look at US/China/Hong Kong/Taiwan/Korea only.
Show me along whether they are fab/fabless and current market share.

[John Huang @zzjohnh]

@_7t 对公司股东来说这是好事啊？AI肯定还是比人便宜，如果可以砍掉50%的人那利润率不直接飞升了？

[Satya Nadella @satyanadella]

https://t.co/vLmiBKTtX3

[John Huang @zzjohnh]

@TheStalwart I also found out very recently this is not from some sort of shared data by government network. The nav app built this by deriving it from user data.

[John Huang @zzjohnh]

@DissentFu sounds like somebody who woke up a loser