Tag: 互联网

因为 vitest 只能用 node 运行，所以 不得不吧 easy-office-2 的里所有直接调用 Bun Api 的代码全部改成兼容 Node Api 的形式。

质疑 node，理解 node，使用 node，但是我依然讨厌 node。

Anton Putra在 YouTube 上发了一条视频，对比了 FastAPI (Python) 和 Node.js 的服务端性能表现。第一轮是测试单纯的 Get 请求，第二轮是测试 PostgreSQL 数据库的写入。两轮测试下来，Python 的综合性能表现差不多是 Node.js 的 1/10。

一般来讲服务端的性能瓶颈都在数据库 IO，很少会遇到 CPU 瓶颈。Anton Putra 测试过不少语言和服务端框架，Python 是为数不多能在测试中段就能撞上 CPU 性能墙的。而且撞墙以后也没有恢复，后半段被锁死在 50% 的 CPU usage 上跑完了全程，也不知道为什么。

FastAPI 已经是 Python 生态内性能比较好的框架了，我真不敢想 Django 这种老乌龟得慢成什么样。没有 jit，再加上 gil 的限制，Python 的性能真的……配不上它今天的江湖地位。

最关键的是，Node.js 在 JavaScript 生态内的性能也不突出，甚至可以说是比较拉胯的……

又心痒痒了，想试试用 htmx + Python 写写小项目。

不过 Python 的速度真不行，感觉 Bun 会好很多。

但是 Python 的 Ai 生态更好，JavaScript 生态圈内优质的 Ai 开源项目太少了。

好纠结啊。

大语言模型「懂不懂你」取决于它能获取到多少你有价值的数据。所以能无感，或低门槛的让你把有价值的数据交给它至关重要。

从这个角度看，微软、谷歌是最有潜力打造出真正基于大语言模型的生产力工具。前者有 office 这个全世界最受欢迎的办公套件，后者也有一套用户基数很大的办公套件，以及持续收集用户数据的搜索引擎和浏览器。

国内在这方面最有潜力的还是字节跳动。飞书虽然市占率不高，但是飞书已经基于豆包推出了知识问答，可以全量获取用户飞书内的所有聊天记录、文档、知识库等数据。

字节在这方面唯一的短板是不像 OpenAI 能做开放平台打通第三方服务（让腾讯和阿里把自己的数据开放给字节？这不可能），也不像 Google 本身就是一个小商业生态系统，用户能在里面满足一部分需求（现在已经打通了 Google 酒旅和 YouTube）。

腾讯也掌握了微信这个数据富矿，可惜微信里大部分都是聊天这种垃圾数据。企业微信和腾讯文档本身产品力不行，远远不如飞书好用。我甚至怀疑他们底层的数据架构也不如飞书开放灵活，想要打通、整合，可能还需要付出一点代价。

阿里有钉钉，但是我不清楚钉钉文档有多少人在用。在我的印象里，钉钉文档应该远不如飞书那么好用。而且钉钉文档的底层数据架构应该也不如飞书。现在钉钉文档 AI 也不能全量获取用户数据，还需要用户手动导入知识库，支持的格式也有限。

DALL-E/Every illustration.

Sponsored By: Composer

_The internet is like water—we take its existence for granted, but its birth was by no means pre-ordained. A constellation of inventors, organizations, and efforts all contributed to its creation. In one of her signature deep dives, Contrary writer Anna-Sofia Lesiv excavates the history of digital communication infrastructure, from the invention of the telephone to the widespread installation of fiber-optic cable and big tech’s subsidization of undersea cables. Read this chronicle to understand how the internet’s decentralized origins led to its current state as fractured spaces governed by private entities—and its implications for its future accessibility. —__Kate Lee_

The internet is a universe of its own. For one, it codifies and processes the record of our society’s activities in a shared language, a language that can be transmitted across electric signals and electromagnetic waves at light speeds.

The infrastructure that makes this scale possible is similarly astounding—a massive, global web of physical hardware, consisting of more than 5 billion kilometers of fiber-optic cable, more than 574 active and planned submarine cables that span a over 1 million kilometers in length, and a constellation of more than 5,400 satellites offering connectivity from low earth orbit (LEO).

According to recent estimates, 328 million terabytes of data are created each day*. _There are billions of smartphone devices sold every year*, _and although it’s difficult to accurately count the total number of individually connected devices, some estimates put this number between 20 and 50 billion.

“The Internet is no longer tracking the population of humans and the level of human use. The growth of the Internet is no longer bounded by human population growth, nor the number of hours in the day when humans are awake,” writes Geoff Huston, chief scientist at the nonprofit Asia Pacific Network Information Center.

But without a designated steward, the internet faces challenges for its continued maintenance—and for the accessibility it provides. These are incredibly important questions. But in order to grasp them, it’s important to understand the internet in its entirety, from its development to where we are today.

Unlock the power of algorithmic trading with Composer. Our platform is designed for both novices and seasoned traders, offering tools to create automatic trading strategies without any coding knowledge.

With more than $1 billion in trading volume already, Composer is the platform of choice. Create your own trading strategies using AI or choose from over 1,000 community-shared strategies. Test their performance and invest with a single click.

The theory of information

In the analog era, every type of data had a designated medium. Text was transmitted via paper. Images were transmitted via canvas or photographs. Speech was communicated via sound waves.

A major breakthrough occurred when Alexander Graham Bell invented the telephone in 1876. Sound waves that were created on one end of the phone line were converted into electrical frequencies, which were then carried through a wire. At the other end, those same frequencies were reproduced as sound once again. Speech could now transcend physical proximity.

Unfortunately, while this system extended the range of conversations, it still suffered from the same drawbacks as conversations held in direct physical proximity. Just as background noise makes it harder to hear someone speak, electrical interference in the transfer line would introduce noise and scramble the message coming across the wire. Once noise was introduced, there was no real way to remove it and restore the original message. Even repeaters, which amplified signals, had the adverse effect of amplifying the noise from the interference. Over enough distance, the original message could become incomprehensible.

Still, the phone companies tried to make it work. The first transcontinental line was established in 1914, connecting customers between San Francisco and New York. It comprised 3,400 miles of wire hung from 130,000 poles.

In those days, the biggest telephone provider was the American Telephone and Telegraph Company (AT&T), which had absorbed the Bell Telephone Company in 1899. As long-distance communications exploded across the United States, Bell Labs, an internal research department of electrical engineers and mathematicians, started to think about expanding the network’s capacity. One of these engineers was Claude Shannon.

In 1941, Shannon arrived at Bell Labs from MIT, where the ideas behind the computer revolution were in their infancy. He studied under Norbert Wiener, the father of cybernetics, and worked on Vannevar Bush’s differential analyzer, a type of mechanical computer that could resolve differential equations by using arbitrarily designed circuits to produce specific calculations.

Source: __Computer History Museum_._

It was Shannon’s experience with the differential analyzer that inspired the idea for his master’s thesis. In 1937, he submitted “A Symbolic Analysis of Relay and Switching Circuits.” It was a breakthrough paper that pointed out that boolean algebra could be represented physically in electrical circuits. The beautiful thing about these boolean operators is that they require only two inputs—on and off.

It was an elegant way of standardizing the design of computer logic. And, if the computer’s operations could be standardized, perhaps the inputs the computer operated on could be standardized too.

When Shannon began working at Bell Labs during the Second World War, in part to study cryptographic communications as part of the American war effort, there was no clear definition of information. “Information” was a synonym for meaning or significance, its essence was largely ephemeral. As Shannon studied the structures of messages and language systems, he realized that there was a mathematical structure that underlied _information. _This meant that information could, in fact, be quantified. But to do so, information would need a unit of measurement.

Shannon coined the term “bit” to represent the smallest singular unit of information. This framework of quantification translated easily to the electronic signals in a digital computer, which could only be in one of two states—on or off. Shannon published these insights in his 1948 paper, “A Mathematical Theory of Communication,” just one year after the invention of the transistor by his colleagues at Bell Labs.

The paper didn’t simply discuss information encoding. It also created a mathematical framework to categorize the entire communication process in this way. For instance, Shannon noted that all information traveling from a sender to a recipient must pass through a channel, whether that channel be a wire or the atmosphere.

Shannon’s transformative insight was that every channel has a threshold—a maximum amount of information that can be delivered reliably to a sender. As long as the quantity of information carried through the channel fell below the threshold, it could be delivered to the sender intact, even if noise had scrambled some of the message during transmission. He used mathematics to prove that any message could be error-corrected into its original state if it traveled through a large-enough channel.

The enormity of this revolution is difficult to communicate today, mainly because we’re swimming in its consequences. Shannon’s theory implied that text, images, films, and even genetic material could be translated into his informational language of bits. It laid out the rules by which machines could talk to one another—about anything.

At the time that Shannon developed his theory, computers could not yet _communicate _with one another. If you wanted to transfer information from one computer to the other, you would have to physically walk over to the other computer and manually input the data yourself. However, talking machines were now an emerging possibility. And Shannon had just written the handbook for how to start building it.

Switching to packets

The telephone system was the only interconnected network by the mid-20th century. AT&T was the largest telephone network at the time. It had a monstrous continental web with hanging copper wires criss-crossing across the continent.

The telephone network worked primarily through circuit switching. Every pair of callers would get a dedicated “line” for the duration of their conversation. When it ended, an operator would reassign that line to connect other pairs of callers, and so on.

At the time, it was possible to get computers “on the network” by converting their digital signals into analog signals, and sending the analog signals through the telephone lines. But reserving an entire line for a single computer-to-computer interaction was seen as hugely wasteful.

Leonard Kleinrock, a student of Shannon’s at MIT, began to explore the design for a digital communications network—one that could transmit digital bits instead of analog sound waves.

His solution, which he wrote up as his graduate dissertation, was a packet-switching system that involved breaking up digital messages into a series of smaller pieces known as packets. Packet switching shared resources among connected computers. Rather than having a single computer’s long communiqué take up an entire line, that line could instead be shared among several users’ packets. This design allowed more messages to get to their destinations more efficiently.

For this scheme to work, there would need to be a network mechanism responsible for granting access to different packets very quickly. To prevent bottlenecks, this mechanism would need to know how to calculate the most efficient, opportunistic path to take a packet to its destination. And this mechanism couldn’t be a central point in the system that could get stuck with traffic—it would need to be a distributed mechanism that worked at each node in the network.

Kleinrock approached AT&T and asked if the company would be interested in implementing such a system. AT&T rejected his proposal—most demand was still in analog communications. Instead, they told him to use the regular phone lines to send his digital communications—but that made no economic sense.

“It takes you 35 seconds to dial up a call. You charge me for a minimum of three minutes, and I want to send a hundredth-of-a-second of data,” Kleinrock said.

It would take the U.S. government to resolve this impasse and command such a network into existence. In the late 1960s, shaken by the Soviet Union’s success in launching Sputnik into orbit, the U.S. Department of Defense began investing heavily in new research and development. It created ARPA, the Advanced Research Projects Agency, which funded various research labs across the country.

Robert Taylor, who was tasked with monitoring the programs’ progress from the Pentagon, had set up three separate Teletype terminals for each of the ARPA-funded programs. At a time when computers cost anywhere from $500,000 to several million dollars, three computers sitting side-by-side seemed like a tremendous waste of money.

“Once you saw that there were these three different terminals to these three distinct places the obvious question that would come to anyone's mind [was]: why don't we just have a network such that we have one terminal and we can go anywhere we want?” Taylor asked.

This was the perfect application for packet switching. Taylor, familiar with Kleinrock’s work, commissioned an electronics company to build the types of packet switchers Kleinrock had envisioned. These packet switchers were known as interface message processors (IMPs). The first two IMPs were connected to mainframes at UCLA and Stanford Research Institute (SRI), using the telephone service between them as the communications backbone. On October 29, 1969, the first message between UCLA and SRI was sent. ARPANET was born.

ARPANET grew rapidly. By 1973, there were 40 computers connected to IMPs across the country. As the network grew faster, it became clear that a more robust packet-switching protocol would need to be developed. ARPANET’s protocol had a few properties that prevented it from scaling easily. It struggled to deal with packets arriving out of order, didn’t have a great way to prioritize them, and lacked an optimized system to deal with computer addresses.

Source: __Computer History Museum_._

By 1974, researchers Vinton Cerf and Robert Khan came out with “A Protocol for Packet Network Intercommunication.” They outlined the ideas that would eventually become Transmission Control Protocol (TCP) and Internet Protocol (IP)—the two fundamental standards of the internet today. The core idea that enabled both was a “datagram,” which wrapped the packets in a little envelope. That envelope would act as a little header at the front of each packet that would include the address it was going to, along with other helpful bits of info.

In Cerf and Khan’s conception, the TCP would run on the end-nodes of the network—meaning that it wouldn’t run on the routers and obstruct traffic, but instead on users’ computers. The TCP would do everything from breaking messages into packets, placing the packets into datagrams, ordering the packets correctly at the receiver’s end, and performing error correction.

Packets would then be routed via IP through the network, which ran on all the packet-directing routers. IP only looked at the destination of the packet, while remaining entirely blind to the contents it was transmitting, enabling both speed and privacy.

These protocols were trialed on a number of nodes within the ARPANET, and the standards for TCP and IP were officially published in 1981. What was exceedingly clever about this suite of protocols was its generality. TCP and IP did not care which carrier technology transmitted its packets, whether it be copper wire, fiber-optic cable, or radio. And they imposed no constraints on what the bits could be formatted into—video text, simple messages, or even web pages formatted in a browser.

_Source: __David D. Clark,_ Designing an Internet.

This gave the system a lot of freedom and potential. Every use case could be built and distributed to any machine with an IP address in the network. Even then, it was difficult to foresee just how massive the internet would one day become.

David Clark, one of the architects of the original internet, wrote in 1978 that “we should … prepare for the day when there are more than 256 networks in the Internet.” He now looks upon that comment with some humor. Many assumptions about the nature of computer networking have changed since then, primarily the explosion in the number of personal computers. Today, billions of individual devices are connected across hundreds of thousands of smaller networks. Remarkably, they all still do so using IP.

Although ARPANET was decommissioned in 1986, the rest of the connected computers kept going. Residences with personal computers used dial-up to get email access. After 1989, a new virtual knowledge base was invented with the World Wide Web.

With the advent of the web, new infrastructure, consisting of web servers, emerged to ensure the web was always available to users, and programs like web browsers allowed end nodes to view the information and web pages stored in the servers.

Source: __Our World in Data_._

As the number of connected people increased in hockey-stick fashion, carriers finally began realizing that dial-up—converting digital to analog signals—was not going to cut it anymore. They would need to rebuild the physical connectivity layer by making it digital-first.

The single biggest development that would enable this and alter the internet forever was the mass installment of fiber-optic cable throughout the 1990s. Fiber optics use photons traveling through thin glass to increase the speed of information flow. The fastest connection possible with copper wire was about 45 million bits per second (mbps). Fiber optics made that connection more than 2,000 times faster. Today, residences can hook into a fiber-optic connection that can deliver them 100 billion bits per second (gbps).

Fiber was initially laid down by telecom companies offering high-quality cable television service to homes. The same lines would be used to provide internet access to these households. However, these service speeds were so fast that a whole new category of behavior became possible online. Information moved fast enough to make applications like video calling or video streaming a reality.

The connection was so good that video would no longer have to go through the cable company’s digital link to your television. It could be transmitted through those same IP packets and viewed with the same experience on your computer.

YouTube debuted in 2004 and Netflix began streaming in 2007. The data consumption of American households skyrocketed. Streaming a film or a movie requires about 1 to 3 gigabytes of data per hour. In 2013, the median household consumed 20-60 gigabytes of data per month. Today, that number falls somewhere about 587 gigabytes.

And while it may have been the government and small research groups that kickstarted the birth of the internet, its evolution henceforth was dictated by market forces, including service providers that offered cheaper-than-ever communication channels and users that primarily wanted to use those channels for entertainment.

A new kind of internet emerges

If the internet imagined by Cerf and Kahn was a distributed network of routers and endpoints that shared data in a peer-to-peer fashion, the internet of our day is a wildly different beast.

The biggest reason for this is that the internet today is not primarily used for back-and-forth networking and communications—the vast majority of users treat it as a high-speed channel for content delivery.

In 2022, video streaming comprised nearly 58 percent of all Internet traffic. Netflix and YouTube alone accounted for 15 and 11 percent, respectively.

This even shows up in internet service provision statistics. Far more capacity is granted for downlink to end nodes than for uplink—meaning there is more capacity to provide information to end-user nodes than to send data through networks. Typical cable speeds for downlink might reach over 1,000 mbps, but only about 35 mbps are granted for uplink. It’s not really a two-way street anymore.

Even though the downlink speeds enabled by fiber were blazingly fast, the laws of physics still imposed some harsh realities for global internet companies with servers headquartered in the United States. The image below shows the “round-trip time” for various global users to connect to Facebook in 2011.

Source: __Geoff Huston_._

At the time, Facebook users in Asia or Africa had a completely different experience to their counterparts in the U.S. Their connection to a Facebook server had to travel halfway around the world, while users in the U.S. or Canada could enjoy nearly instantaneous service. To combat this, larger companies like Google, Facebook, Netflix, and others began storing their content physically closer to users through CDNs, or “content delivery networks.”

These hubs would store caches of the websites’ data so that global users wouldn’t need to ping Facebook’s main servers—they could merely interact with the CDNs. The largest companies realized that they could go even further. If their client base was global, they had an economic incentive to build a global service infrastructure. Instead of simply owning the CDNs that host your data, why not own the literal fiber cable that connects servers from the United States to the rest of the world?

In the 2020s, the largest internet companies have done just that. Most of the world’s submarine cable capacity is now either partially or entirely owned by a FAANG company—meaning Facebook (Meta), Amazon, Apple, Netflix, or Google (Alphabet). Below is a map of some of the sub-sea cables that Facebook has played a part in financing.

Source: __Telegeography_._

These cable systems are increasingly impressive. Google, which owns a number of sub-sea cables across the Atlantic and Pacific, can deliver hundreds of terabits per second through its infrastructure.

In other words, these applications have become so popular that they have had to leave traditional internet infrastructure and operate their services within their own private networks. These networks not only handle the physical layer, but also create new transfer protocols —totally disconnected from IP or TCP. Data is transferred on their own private protocols, essentially creating digital fiefdoms.

This verticalization around an enclosed network has offered a number of benefits for such companies. If IP poses security risks that are inconvenient for these companies to deal with, they can just stop using IP. If the nature by which TCP delivers data to the end-nodes is not efficient enough for the company’s purposes, they can create their own protocols to do it better.

On the other hand, the fracturing of the internet from a common digital space to a tapestry of private networks raises important questions about its future as a public good.

For instance, as provision becomes more privatized, it is difficult to answer whose shoulders the responsibility of providing access to the internet as a “human right,” as the U.N. describes, will fall on.

And even though the internet has become the de facto record of recent society’s activities, there is no one with the dedicated role of helping maintain and preserve these records. Already, the problem known as link rot is beginning to affect everyone from the Harvard Law Review, where, according to Jonathan Zittrain, three quarters of all links cited no longer function. This occurs even at The New York Times, where roughly half of all articles contain at least one rotted link.

The consolation is that the story of the internet is nowhere near over. It is a dynamic and constantly evolving structure. Just as high-speed fiber optics reshaped how we use the internet, forthcoming technologies may have a similarly transformative effect on the structure of our networks.

SpaceX’s Starlink is already unlocking a completely new way of providing service to millions. Its data packets, which travel to users via radio waves from low earth orbit, may soon be one of the fastest and most economical ways of delivering internet access to a majority of users on Earth. After all, the distance from LEO to the surface of the Earth is just a fraction of the length of subsea cables across the Atlantic and Pacific oceans. Astranis, another satellite internet service provider that parks its small sats in geostationary orbit, may deliver a similarly game-changing service for many. Internet from space may one day become a kind of common global provider. We will need to wait and see what kind of opportunities a sea change like this may unlock.

Still, it is undeniable that what was once a unified network has, over time, fractured into smaller spaces, governed independently of the whole. If the initial problems of networking involved the feasibility of digital communications, present and future considerations will center on the social aspects of a network that is provided by private entities, used by private entities, but relied on by the public.

Anna-Sofia Lesiv is a writer at venture capital firm __Contrary_, where she originally published __this piece_. She graduated from Stanford with a degree in economics and has worked at Bridgewater, Founders Fund, and 8VC.

Astro 是个好框架，但是从实用性的角度上讲，彻底放弃 SSR 好像不太现实 🤔

最近两天在学 Astro。

Astro 的一大优势就是能静态生成网站。通过编译的方式，将 JavaScript 写的网站编译成只包含 HTML 和 CSS 的静态文件。不管是上到服务器用 Nginx 转发，还是放到静态网站托管平台，还是直接放 CDN，都能运行，不挑环境。

编译的好处就是能一处编译，处处运行。不依赖运行时和解释器。

现在 JavaScript 的运行时这么多，为啥没有人用 golang 写一个？然后继承 Golang 可以直接编译为二进制的优势，让 JavaScript 也能编译为二进制，做到一处编译，处处运行。

在国内最适合做 Ai Workspace 产品的，还得是微信，毕竟我们老中人最爱用微信办公了。微信沉淀的工作群聊记录和各种 word、ppt 文件才是最真实有价值的检索数据。

Ai 要真正提升生产力，做出真正不一样的产品，还是需要和已有的数据整合起来。

从这一点来看谷歌的优势太大了。目前谷歌的 Ai Workspace 是整合最好的 Ai office 产品。微软的 copilot 365 输在了 office 三件套用户体验不行。notion ai 能整合的资源有限，价格没有竞争力。wps 的用户基本不上云，金山没有数据可用。国内最有希望的还是字节，飞书体验没有问题，就看字节能把豆包优化成什么样了。

Open Ai、anthropic 起了个大早，赶了个晚集。如果 chrome 真的能被剥离，这俩公司谁能买下来，谁就有一线生机。

分享也是一种促使我去思考的动力。这就是为什么我要让这个网站可以被公开访问。