New Space Race A few weeks ago I posted about the five major Chinese projects to develop reusable rockets, as they never seem to get much attention over here. Since that post, China has now successfully demonstrated 2 reusable rocket projects: • Their own cable catch barge • Folding landing legs (falcon 9 style) There are 3 other major Chinese rocket programs, and there are a lot of smaller minor programs behind those. China’s efforts are serious and progressing fast. What should be very obvious to anyone following SpaceX is that space (or rather Earth orbit) is really a numbers game. A space data center is actually not a “center” at all, it is a satellite constellation, it is not a single large object assembled in space it is a constellation of satellites much like Starlink already is. The way to think about it is that each satellite is basically 1 server rack. A single satellite is optimally around 150-250kW, which is EXACTLY where current AI racks are in the 2020s. So the fit seems fairly attractive even if you have to swallow the launch cost, and the lack of synchronicity that comes from being spread out in a constellation. Space data is solar powered and China already makes 95% of the world’s solar, so going after Elon is a very natural strategy for China, they already totally dominate one of the main industrial substrates, the solar cells. But space is only 1 branch of this story. Nvidia’s next gen rack technology Kyber moves to 800VDC, and to liquid cooling, this pushes rack power density to 600kW and possibly 1-1.5MW, well beyond the solar satellite optimal package size. At those power densities you start to go beyond what is optimal for a satellite and so the space data center window is opening in terms of access, just as Nvidia is moving chips toward something much denser. Chip tech is about to fork by heat sink. • Radiative cooling into deep space • Conductive/convective cooling into water This is why SpaceX needs their own silicon, this is Terafab is being built. Space chips need to run at higher temperatures to dissipate heat radiatively and be more energy efficient, whereas conductive and convective cooled chips just want to cram as much power and compute as possible into the synchronous operation horizon (the distance that light travels during 1 clock tick). For example at 3GHz clock speed light can travel 10cm between each tick of the clock, this is why fast chips must be small, and why a satellite constellation cannot run synchronous operations. Radiative cooling in space means distributed asynchronous compute. Liquid cooling at home means densified synchronous compute. So this new Space Race sits entirely on one side of the fork. Either way, China seems to be doing their fast follower thing here, and is following Elon into building a big space constellation of hot asynchronous chips in space. China has pre-built the solar fabs to support scaling and has 5 launch programs pushing to give them access, 2 of which are now demonstrated. Space silicon with high chip temperatures and limited energy budget, is not Nvidia’s direction, and the two tech forks lead to very different application strengths and weaknesses and different applications in that layer will fall under one fork or the other.

其实，和其他物体的关系和作用，正代表了这个物体所有可能的性质。而一个物体的所有可能性质，也就完全定义了这个物体本身。归根到底，也许我们根本就没有物体，只有一大堆关系。而物体这一抽象的概念，以及物体所有可能的性质，是由这一堆关系来定义的。这就是范畴学的精神。 文小刚

The Information还在穷追猛打Anthropic创始人达里奥·阿莫代伊的幕后信息，太刺激了。 这次深扒的是达里奥的妻子卡米拉，看完之后，我只能说，果然成功者都能忍常人所不能忍啊⋯⋯ 达里奥这个人的心理问题大家应该都已经知道了，既有浓厚的弥赛亚情结，也属于资深的怀疑论者，这导致他猜忌心很重，当初创办Anthropic时，就带着自己的妹妹过来担任总裁，在他还想把自己的妻子卡米拉也安排进创始团队时，遭到了其他成员的不满，于是只得作罢。 但因为枕边这层关系始终存在，卡米拉还是成为了能在背后左右公司的关键人物，哪怕她没有任何职权，甚至Anthropic的第一笔融资，就是卡米拉帮忙谈成的，或者说，是卡米拉劝自己的前男友出钱支持自己的现男友。 😅😅😅 卡米拉早年曾和谷歌前CEO施密特谈过恋爱，分手之后，卡米拉找了当时还在斯坦福读博的达里奥当男友，一直陪着他从高校到企业，从百度到OpenAI，最后出走自己创业。 那会儿OpenAI已经肝出了GPT-3，处于盛名前夜，达里奥在行业里并不为人所知，所以Anthropic也很难找到投资者，是卡米拉把施密特介绍给了达里奥，让他俩在自己的公寓里彻夜长谈，最终敲定了投资协议。 搁一般人可能很难接受这样的命运，在最缺钱的时候，女朋友出面给多金的前男友打电话，求他帮忙给点支持，而达里奥也忍了下来，面不改色的接受了施密特的帮忙。 当然此一时彼一时，昔日的穷小子，如今变成了地球上最值钱的AI公司的老板，施密特这种老登反而排不上号了，很有爽文剧本的味道。 卡米拉出生于1979年，家庭条件普通，大学读的建筑专业，毕业后去了知名工学椅生产商赫曼米勒做销售，后来创业想做女性向的成人电影，认为Pornhub的内容太垃圾了。 她一度找过爱泼斯坦投资，不过当时并不知道后者的丑闻，爱泼斯坦也没理她，于是成人电影制片梦想破碎，她转而结识了施密特，开始新的创业项目，一家护理女性健康的连锁诊所。 当时卡米拉30岁出头，而施密特已经快60岁了，标准的崩老头情节，施密特这个人对自己的女友们都挺大方的——这个可以另外再扒——二话不说就给卡米拉投了一笔启动资金。 而卡米拉在筹备公司时，从斯坦福大学请了一个顾问，她在BP里把这个顾问的名字填了进去：达里奥·阿莫代伊博士。 基本上是无缝衔接了…… 反过来看，卡米拉的选人眼光确实有一手，曾经拿下亿万富翁施密特就不说了，当她和一文不名的达里奥在一起时，也坚定看好这个男人的潜力，很早就把他的论文整理成册，发给自己认识的大佬们，希望他们能注意到达里奥。 现在她取得了最大的成功，被圈子里称作Anthropic的「第一夫人」，是达里奥最重要的参谋，同时也不再需要抛头露面，她把LinkedIn、Instagram的账号都注销了，专心致志的辅佐丈夫。

@royxy 而它们背后的共同语言就是： 线性算子 + 特征模式 + 特征值​ 如果你继续往下学，会发现傅里叶变换、矩阵特征值、振动模态、量子力学、PDE、信号处理最后都会汇聚到一个非常漂亮的概念：“把复杂问题变换到一个由特征模式组成的坐标系里，让原本耦合的问题变成许多个简单的问题。”