文章正文

侠客行

“Kimi时刻”反而利好芯片?华尔街:“杰文斯悖论”将再度应验!_我的网站

名侦探柯南

一 |     

Qingyangshen, or Root of Auricledleaf Swallowwort, a traditional ethnic medicinal plant long used in Southwest China's Yunnan Province. Photo: Scrrenshot from website
    Qingyangshen, or Root of Auricledleaf Swallowwort, a traditional ethnic medicinal plant long used in Southwest China's Yunnan Province. Photo: Scrrenshot from website
A Chinese research team has identified a new potential target to develop antidepressant drug from Qingyangshen, or Root of Auricledleaf Swallowwort, a traditional ethnic medicinal plant long used in Southwest China's Yunnan Province, according to Kunming University of Science and Technology, Science and Technology Daily reported Sunday.
The team, led by professors Xu Tianrui and Xiang Cheng from the university's School of Life Science and Technology, found that Qingyangshen glycosides, the plant's core active compound, can alleviate depression-related behaviors by simultaneously suppressing inflammation, reducing oxidative damage and stabilizing neurotransmitters in the brain.
The findings were published in the international journal Phytomedicine.  
Depression has become a widespread mental health disorder globally, while existing clinical drugs fail to provide effective treatment for more than 30 percent of patients, according to the researchers. Previous studies have established a strong link between chronic inflammation and depression, but existing anti-inflammatory drugs have limited targeting capabilities, underscoring the need for new therapeutic targets and lead compounds.
Qingyangshen, a plant of the Cynanchum genus in the Apocynaceae family, has long been used as a traditional medicine by ethnic groups including the Naxi, Yi and Bai peoples in Yunnan, mainly for treating neurological conditions such as epilepsy and neurasthenia. 
The researchers systematically examined the pharmacological effects of Qingyangshen extract and its core active compound, Qingyangshen glycosides. Tests using mouse, zebrafish and fruit fly models repeatedly showed that the compound significantly improved depression-related behaviors induced by inflammation, according to the report.
To identify the compound's target, the team developed a dual-discrimination model for four subtypes of prostaglandin E2 receptors and combined it with AI-powered machine learning. The researchers screened hundreds of G protein-coupled receptors and identified the prostaglandin E2 subtype 4 receptor as the direct target of Qingyangshen glycosides.
Multiple molecular experiments confirmed that the compound binds directly to the receptor and acts as an inhibitory ligand. Further experiments involving animals with the receptor knocked out in the hippocampus showed that the antidepressant effects of Qingyangshen glycosides disappeared completely when the receptor was absent, providing the first in-vivo evidence that the receptor is essential for the compound's therapeutic effects.
An illustration of the research. Photo: Screenshot from the website
    An illustration of the research. Photo: Screenshot from the website
The study establishes the prostaglandin E2 subtype 4 receptor as a potential new target for antidepressant drug development and points to a new approach to treating depression through inflammatory pathways. It also identifies Qingyangshen glycosides as a potential lead compound for developing highly targeted anti-inflammatory antidepressants, creating a research chain from ethnic medicinal plants and active compounds to drug targets and new drug development.
Global Times
。    财联社7月21日讯(编辑 潇湘)上周,一款来自大洋彼岸中国的极具竞争力的AI模型,重新引发了华尔街对于美国AI行业在存储组件和算力上巨额支出是否合理的争论——费城半导体指数也在上周五跌入了技术性熊市区域。         然而,在经历了一个周末的热议后,一些华尔街分析师和美国科技行业人士认为,低成本中国AI的出现对于美国芯片巨头来说并不一定是一件坏事,尽管这些芯片巨头的股价此前曾因外界对AI硬件的强劲需求而飙升。

二 |          他们的理据是——即使AI成本下降,这些芯片公司最终也可能会从中受益。         上周四,中国AI开发商月之暗面发布了最新的Kimi K3模型,这是一款开放权重模型,在某些基准测试中其排名可与领先的美国模型媲美。

三 |          在Arena综合文本排名中,Kimi K3目前以1679分位居前端代码排行榜的榜首,领先于Anthropic的Claude Fable 5(1631分)和OpenAI的GPT-5.6 Sol(1618分)——这两者均属于闭源模型。该排名是基于对98个模型进行的约484000次两两对决盲测投票得出的。鉴于K3的前代产品排名第18位——因此,月之暗面在极短的时间内实现了性能的重大飞跃。         一家中国实验室在资金资源较少且面临美国对英伟达最新、最强大芯片出口管制的情况下,能够开发出一个具备近前沿能力的模型,这给华尔街和硅谷带来了震动。Kimi K3的发布加剧了上周五美国芯片板块的抛售势头。

四 |          然而,一些业内人士表示,对Kimi K3暴涨的需求表明,中国的AI模型在算力饥渴程度上可能与西方同行不相上下。

五 | 月之暗面上周末表示,对Kimi K3爆发式的需求已将其容量推至极限,导致该公司暂停了C端新用户订阅,以优先保障现有会员的算力调配。

六 |          杰文斯悖论将再度应验?          BluBird Capital管理合伙人Anni Sen明确表示,这一动态“对内存芯片交易而言是绝对的利好顺风”。         开源社区Hugging Face的数据显示,在Kimi K3高达2.8万亿的总参数中,其激活参数仅约500亿个。

七 | 参数是AI模型在训练阶段用于学习和识别特征规律的变量,相当于模型的“记忆单元”。         Anni Sen解释称,虽然这种设计极大地提升了模型的运行效率,但要在主存中实时驻留2.8万亿个参数,意味着普通企业几乎不可能在自建服务器上独立运行该模型。

八 |          Kimi K3目前对每百万个输出token收费15美元——相比之下,OpenAI的GPT-5.6 Sol收费30美元,Anthropic的Claude Fable 5收费50美元。         华尔街人士认为,作为“杰文斯悖论”的一个体现——即提高一种资源的使用效率可能会增加其总需求,Kimi K3更低的token成本可能会导致AI工作负载的增加。         Sen认为,“随着开发者利用其较低的成本结构,Kimi将创造更多的应用场景”,进而也会增加对推理的需求。Sen表示,企业将越来越多地将“中端”AI 任务迁移至更便宜的模型,同时把极其昂贵的前沿模型留给更为复杂的多步骤智能体流程。         AI研究员兼Interconnects AI博客创始人Nathan Lambert在周一的一篇文章中写道:“开放权重模型通过降低特定性能水平下的智能门槛,对整个经济中的AI普及起到了加速作用。”          他补充道,开放模型提供了额外的定制化机会,这可以使它们对特定企业更有用。         芯片需求将进一步增长          Wedbush分析师Matt Bryson表示,随着AI模型变得越来越大,支持更多参数对内存的需求也将增长。Bryson在周一的一份报告中表示,这意味着单个AI芯片配备的内存容量必须大幅提升,要么意味着必须通过扩充芯片集群规模,来分摊承载庞大的权重数据。         因此,Bryson表示,如果中国AI模型继续获得青睐,考虑到对高性能内存的需求将会增加,这对内存供应商来说“可以说是一件好事”。此外,由于模型参数增加而导致芯片集群变大,这也可能会给网络设备提供商带来推动力。         美光科技、SK海力士和三星电子是目前全球高带宽内存(HBM)的三大主要供应商,内存组件的短缺赋予了这些公司极强的定价权。         D.A. Davidson常务董事Gil Luria表示,随着运行AI模型的成本下降,对算力的需求可能会增加,这将使芯片制造业受益——尤其是在正经历供应紧张的内存芯片方面。

九 |          PurePlay ETFs首席执行官Joseph DeYonker也表达了类似的观点,即随着AI模型变得更加高效和易于获取,用户采用率的激增“会立即考验物理基础设施的极限”——就如Kimi K3的情况一样。在他看来,这证实了芯片制造商的长期增长轨迹依然完好。         DeYonker表示,“运行具有庞大上下文窗口的高级前沿模型,正在给内存和芯片制造供应链施加前所未有的巨大压力。

十 | ” 因此,他预计专注于HBM和先进芯片封装的公司将保持强大的定价权。         同时,DeYonker还预计GPU和定制AI芯片的设计者——如英伟达和博通,将继续看到强劲的订单积压,因为企业和超大规模算力厂商正争先恐后地获取必要的硬件,以避免发生这些具体的运营瓶颈。         在他看来,对Kimi K3订阅的限制是“对半导体行业结构性的确认,而不是周期性的警示信号”,投资者应该将持续的内存和算力瓶颈,视为芯片板块具备长期可持续盈利增长潜力的迹象。         (财联社 潇湘)。

十一 |

Current article:http://www.bianhajieleigoukusunmen.pics/list_5op/yq3vk.html

Published on:07:21:00


|