科技爱好者周刊（第 397 期）：财富正在向 AI 集中
TL;DR · AI Summary
本期科技爱好者周刊聚焦于AI技术的快速发展及其对社会财富分配的影响。文章指出，AI相关产业如内存、储存、CPU、服务器等的股价大幅上涨，表明财富正迅速向AI领域集中。此外，文章还探讨了AI在日常生活中的应用，如通过AI估算食物的碳水含量，但实验表明AI在这方面并不准确。微软宣布将淘汰短信验证码，转而采用更安全的验证方式，如Passkey。亚马逊推出供应链服务，开放其物流网络，可能对制造业产生影响。最后，文章介绍了一个机械打字机模型玩具，以及几篇技术文章，包括关于GitHub Pages域名盗用问题、JavaScript ShadowRealm API和Firefox配置指南。
Key Takeaways
- AI相关产业的股价大幅上涨，表明财富正迅速向AI领域集中。
- 实验表明，AI在估算食物碳水含量方面并不准确，不建议用于医疗估算。
- 微软将淘汰短信验证码，转而采用更安全的验证方式，如Passkey。
Outline
Jump quickly between sections.
Mindmap
See how the topics connect at a glance.
查看大纲文本（无障碍 / 无 JS 友好）
- 科技爱好者周刊第397期
Highlights
Key sentences worth saving and sharing.
内存、储存、CPU、服务器、液冷、光通信、变压器......股价全部在涨，更不要提，前期已经涨过的芯片、模型、算力了，甚至铜和铝也在涨。
社会财富正在重新分配，快速向 AI 集中。
四个模型给出的回答不一样。而且，同一张照片多次提交给同一个模型，回答也不一样。
微软公司本周确认，将放弃短信验证码，改用 Passkey、一次性时间码（TOTP）、以及验证过的邮件地址。
亚马逊宣布推出'亚马逊供应链服务'（ASCS），把自己的货运、配送、仓储和包裹运输向用户开放。
This records weekly tech content worth sharing, published on Fridays.
This magazine is open source, welcome to contribute. There's also a "Who's Hiring" service for posting programmer job openings. For cooperation, please [contact via email](mailto:yifeng.ruan@gmail.com) (yifeng.ruan@gmail.com).
Cover Image
Hainan Science and Technology Museum designed by Ma Yansong's firm, opened during this year's May Day holiday. (via)
Wealth is Concentrating Toward AI
Everything related to AI has been rising recently.
Just look at the stock market. Memory, storage, CPU, servers, liquid cooling, optical communication, transformers... all stock prices are rising, not to mention chips, models, and computing power that have already risen earlier, even copper and aluminum are rising.
The magnitude of these stock increases is breathtaking. Take memory as an example - among the world's three major memory manufacturers, two are in South Korea. These two companies alone pulled the Korean stock market from 2,600 points to 7,600 points within a year.
In October 2024, Samsung's chairman still apologized due to poor semiconductor performance, but this year the company will likely become the world's most profitable company.
Another major memory manufacturer SK Hynix is even more extreme. They have a labor agreement with their union to distribute 10% of profits to employees. Someone calculated that including guards, drivers, and receptionists, each employee can receive an average bonus of 6.1 million RMB this year.
As for AI model companies, everyone is a millionaire. OpenAI repurchased $6.6 billion worth of stock from 600 employees last year, averaging nearly $10 million per person.
These events indicate that social wealth is being redistributed, rapidly concentrating toward AI.
This affects everyone. Even if you don't use AI at all, rising prices and capital flowing from other industries to AI inevitably impact you.
Commonly used electronic devices (phones and computers), electronic components, basic production materials (like copper and aluminum) are all rising in price. If you're in an industry unrelated to AI, you likely won't benefit from AI's gains but instead be harmed by it, facing rising costs, insufficient demand, and reduced investment.
Ancient wisdom says "one successful general means countless bones," AI's rise is accompanied by the decline of many other industries.
Although every technological revolution inevitably involves wealth redistribution, this AI revolution is advancing exceptionally fast, with extraordinary intensity, causing particularly intense redistribution effects.
As an ordinary person, especially someone in internet and software industries, there seems no other choice but to follow wealth incentives and embrace AI trends.
Don't Use AI to Estimate Carbohydrate Content
Diabetic patients need to control sugar intake, not only eating less sugar but also less carbohydrates (rice and flour), since carbs eventually convert to sugar.
Therefore, they need to know how much carbohydrate is in food.
A natural idea is to take a photo of food before eating and let AI estimate carbohydrate content.
A British doctor conducted an experiment, submitting 13 food photos (like cheese sandwiches, Spanish paella, crème brûlée) to four major models—GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro, Gemini 3.1 Pro—to estimate carbohydrate content.
The results were shocking, with four models giving different answers. Moreover, submitting the same photo multiple times to the same model yielded different responses.
Take the above Spanish paella photo as an example, submitted multiple times to four models, the model estimates are shown below.
As seen, each model's estimates for the same photo have large fluctuations. Among them, Gemini 2.5 Pro fluctuates the most, with carbohydrate estimates ranging from 55 grams to 484 grams, a difference of 429 grams! Claude Sonnet 4.6 has relatively concentrated estimates, but the fluctuation range is still significant.
Taking the cheese sandwich photo as another example, the packaging indicates 40 grams of carbohydrates, but GPT-5.4's average estimate was 74 grams, while other three models estimated 28 grams, all incorrect.
Additionally, large models cannot accurately identify food, sometimes thinking a cheese sandwich contains a piece of cooked meat.
Therefore, don't use large models to estimate food carbohydrate content, or let large models perform any precise medical calculations—they simply can't do it.
Microsoft Eliminates SMS Verification Codes
Many websites send you a verification code via SMS when logging in.
This practice has risks. First, attackers might deceive mobile carriers to bind your phone number to their SIM card, redirecting all SMS to their phone. Second, SMS is sent in plain text and easily leaked.
Microsoft confirmed this week that it will abandon SMS verification codes, switching to Passkey, one-time time-based codes (TOTP), and verified email addresses.
Among these, Passkey will be the primary authentication method after Windows 11.
This approach generates a key pair for each user, stored in Windows' password manager, accessible only through facial recognition, fingerprint scanner, or PIN code to activate private key verification, offering very high security.
Here's an introduction to Passkey (in English), focusing on concepts with concise writing for reference.
Amazon Supply Chain Services
This month, Amazon announced "Amazon Supply Chain Services" (ASCS), opening its freight, delivery, warehousing, and package transportation to users.
Any company can now build its merchandise business on Amazon's logistics infrastructure, with companies like P&G and 3M already using it.
This reminds me of 2006, when Amazon opened its network infrastructure as AWS (Amazon Web Services), directly leading to the cloud era.
Now it's opening warehousing and logistics again—will this change manufacturing?
In the future, creating physical products will become increasingly simple—just design the product, outsource production, and purchase standardized services for logistics and sales.
Typewriter Model
Recently, I saw a model toy that looked quite interesting.
This is a 19th-century mechanical typewriter model launched by a Shenzhen company. After purchasing, you need to assemble it yourself.
Most interestingly, after assembly, it can be used for typing, though with limited functionality—it cannot switch between uppercase and lowercase, supporting only uppercase letters.
It's not just a model toy but also helps understand typewriter principles, attracting attention in overseas markets as well.
Articles
- GitHub Pages has domain abuse issues (in English)
If your domain points to GitHub Pages, better read this article—under certain circumstances, your subdomain may be abused.
- Introduction to JavaScript ShadowRealm API (in English)
ShadowRealm is a new JS API allowing multiple domains within a single-threaded environment, each with its own global object, running independently without interference, and able to communicate with each other.
- Firefox Configuration Guide (in English)
The author introduces his approach to configuring Firefox into a good browser.
- Certificate Revocation Checking Mechanism (in English)
Website HTTPS certificates may be revoked by certificate authorities—how do users know? This article introduces three existing checking mechanisms. The problem is they all have flaws, and Chrome doesn't adopt any of them.
- Ephemeral Infrastructure (in English)
This article proposes that large systems should adopt stateless ephemeral infrastructure (like containers) for all parts except those absolutely requiring state maintenance (such as databases, message queues, etc.).
- Microscale Thermite Reaction (in English)
Thermite reaction refers to aluminum reacting violently with iron oxide (rust) at high temperatures, reaching 2500~3500°C. This article introduces a method to trigger microscale thermite reactions: two rusty iron balls, one wrapped in aluminum foil, colliding together.
Tools
A cross-platform desktop application that converts ePub/PDF/text files into audiobooks.
A web application that generates FFmpeg commands through visual graphics, open source code.
An open-source JavaScript SDK for fetching stock market data, pure front-end scraping without backend services, real-time quotes from public APIs like Tencent Finance and East Money. (Submitted by @chengzuopeng)
Open-source cross-platform desktop client for Mihomo/Clash. (Submitted by @Juwan-Hwang)
Open-source editing software that folds the timeline into multiple rows and displays inserted text above the timeline, all processing completed in the local browser. (Submitted by @hughfenghen)
A sticky note tool hidden inside the MacBook's notch, automatically expands when hovering with mouse. (Submitted by @oil-oil)
A rendering library for AI-generated streaming markdown, providing Vue implementation. (Submitted by @Simon-He95)
Cross-platform TXT file desktop reader that can colorize content, also supports chapter recognition, blank line compression, first-line indentation, text-to-speech, AI reading assistant and other features. (Submitted by @ssnangua)
A fork project of Gitea that can be used to build your own Git server, comes with a web frontend as a personal GitHub service.
Open-source VS Code extension for real-time editing and preview of HTML files.
AI Related
Command-line tool for removing AI image watermarks, also available as a Python package, capable of removing both visible and invisible watermarks (such as Google's SynthID).
- AVC (Agent View Controller)
Converts confirmation text from AI Agent terminals into interactive web popups, can be used as a skill for Agents. (Submitted by @study8677)
Connect locally running AI Agents to WeChat/Telegram, allowing you to receive updates, approve permissions, reply to questions, and continue conversations even when away from your computer. (Submitted by @tuchg)
Generating 10 tokens per second, is that fast or slow? What about 20? This website lets you experience token generation speed online.
Resources
Gaussian Splatting is an algorithm that can synthesize 3D images from multiple photos.
The website SuperSplat provides online Gaussian Splatting processing, where you can upload photos for processing and browse many pre-made 3D models.
Check out the work Strawberry, the effect is excellent.
Free IP geolocation database available for free download and script-based queries.
Images
Chile is the world's largest copper producer, with nearly half of its exports being copper, accounting for up to 60% of global annual supply, still maintaining 25% today.
Chile's largest copper production center is in the Atacama Desert, where there's a sculpture called "Hands Supporting Chile" at the entrance town of the desert.
This sculpture is a pair of giant copper hands holding up a map of Chile, symbolizing the significant contribution of copper mining in this area to Chile.
Below are photos of Chilean desert copper mines, showing huge mining pits and sedimentation pools next to the ore.
There's an empirical rule: all growth curves eventually become S-shaped curves.
For example, below is an exponential growth curve.
Eventually, growth slows down and becomes an S-shaped curve.
This is a universal phenomenon; all growth curves will eventually look like this, flattening after reaching a certain stage.
Digest
Since becoming a freelancer taking on independent projects, I've constantly felt the tension between generalist programmers and specialist programmers.
I found a statement that captures it accurately: "In theory, all companies prefer adaptable programmers who can handle multiple roles. In practice, most job postings favor specialists."
This means that even if you're a generalist programmer, you must present yourself as a specialist when job hunting.
My problem is that I don't feel my skills belong to any specific category. I see myself as a problem-solving hacker driven by curiosity, solving problems that interest me.
I learn to use whatever tools are suitable for solving a particular problem.
However, after going independent, calling myself a generalist significantly affected my ability to attract clients. Clients want to hear that you're a specialist, not vague statements like "I'm a computer generalist." They ask questions like: Are you a front-end or back-end engineer? Do you use .NET or Python? Are you specialized in AWS or Azure?......
I don't blame them... clients face concrete problems and it's genuinely difficult to assess the value of hiring a generalist programmer.
So I decided to change my communication strategy and say what people want to hear. My positioning became: a Rust expert programmer focused on systems programming and open-source software. I packaged myself accordingly.
It worked quickly, keeping me busy with contracts all year!
After a long bath or swim, skin develops wrinkles and may even change color. Why?
You might not believe it, but this remains an unsolved mystery to this day, with no widely accepted explanation.
Previous thinking was that it's due to skin swelling after absorbing water, causing swelling and bending. But researchers in the 1930s observed that people with damaged finger nerves don't develop wrinkles when immersed in water. This suggests that wrinkles aren't related to skin but to the nervous system.
Now, the mainstream scientific explanation is that when hands and feet contact water for more than a few minutes, sweat ducts in the skin open, allowing water to flow into skin tissues.
Increased water in the skin reduces salt proportion in the skin. The nervous system sends signals to the brain about reduced salt levels, then responds by constricting blood vessels to reduce water infiltration.
When blood vessels constrict but the outermost layer of skin doesn't shrink, the outer skin loses support and caves in like a deflated balloon, forming wrinkles. This is similar to how dried grapes wrinkle---they lose more internal volume than surface area.
This vasoconstriction also makes the skin pale because there's less blood in the subcutaneous vessels. This is exactly opposite to what happens when you take a hot bath, when blood vessels dilate and skin turns red.
So, skin wrinkling is triggered by the nervous system. If nerves are damaged, blood vessels won't receive brain signals, won't contract, and skin won't wrinkle.
Quotes
1.
Over the past one or two years, I've completely relied on AI for programming, not writing a single line of code myself.
One day, I wrote an article myself without using AI, but after finishing, I couldn't help copying and pasting it into Claude to see what AI thought, because I worried the article had confused logic, childish writing, or missed something. This is the consequence of heavy AI usage—it breeds self-doubt, and I now trust AI more than myself.
-- "Damn AI Is Making Me Dumb"
2.
No matter how far tools advance, no matter how powerful they become, they remain just tools. They cannot replace our reasoning and values; it's still up to you to choose what's worth creating.
3.
I suggest slowing down and giving yourself some time to think about what exactly you're building and why. That way you have the opportunity to tell yourself "no, we don't need this."
4.
There's a saying in the oil industry: A healthy oil company is run by a geologist, a mature oil company is run by an engineer, a declining oil company is run by an accountant, and a dying oil company is run by a lawyer.
Previous Year in Review
Java's 30th Anniversary (#350)
Thirty Years, Solving Life's Three Major Problems (#300)
The Greatest Risk of New Technologies (#250)
Low Expectations, More Experimentation (#200)
(End)