Weekly Newsletter for Tech Enthusiasts (Issue 391): The Wealth Gap in AI

Here records the technology - related content worth sharing every week, and it is released on Fridays.

This magazine is open - sourced. Contributions are welcome here. There is also a “Who is Hiring” service for posting programmer recruitment information. For cooperation, please contact via email ([email protected]).

Cover Image

The wall decoration of a restaurant in Shanghai. (via monana3838@Threads)

The Wealth Gap in AI

I increasingly feel that AI is different from other technologies. It not only brings about technological changes but also social changes.

In short, AI will bring about the wealth gap.

Other technologies actually eliminate the wealth gap and achieve "consumer equality", that is, the poor and the rich consume the same things.

For example, everyone drinks the same Coca - Cola, uses the same iPhone, and drives the same Tesla. Even the Internet is like this. The world's richest man, Musk, uses the same websites and mobile apps as you.

However, AI models are not like this. The poor and the rich are not equal in the face of large - scale models.

In the future, ordinary people will definitely not be able to afford the top - tier large - scale models. In fact, this is already the case now. The most expensive AI programming package is the Max package of Claude Code, with a monthly fee of $200, and many people can no longer afford it.

OpenAI once envisioned a monthly package costing $20,000, providing the most top - tier and unlimited large - scale model services.

If it is really launched, only the wealthy can afford it.

This reflects a simple fact: The more expensive the cost, the better the model performance. Because the performance of the model is related to computing power. More computing power, a larger context, and more parameters all require money.

This is completely the opposite of industrial products. Industrial products have economies of scale. The higher the output, the lower the unit cost. Once mass - produced, the price will become cheaper and cheaper.

However, large - scale models do not have economies of scale. Mass production of models requires more servers, which will not reduce the unit cost. Instead, it may become more expensive due to the expansion of computer rooms and the transformation of circuits and waterways.

The future society will probably be like this: the rich and the poor use different models. The services of the most top - tier models - - planning, consulting, content generation, automation... - - require high usage fees, while ordinary people use free models, and of course, the performance is also ordinary.

However, I also saw that Musk recently said that there is another possibility in the future.

He means that computing power is essentially a form of energy conversion. Humanity will eventually achieve a large - scale supply of cheap energy (space solar energy?). So computing power will become cheap enough, and everyone will be able to use the best models.

Is it possible? I don't know. It seems that the former situation is more realistic.

A Way to Measure Model Capabilities

How to measure the capabilities of large - scale models?

The current method is to use a test set to calculate the score of the large - scale model. Its drawback is that it can only be used for horizontal comparison and is difficult to measure the progress speed.

Recently, a paper proposed a new measurement method.

Scientists first calculate how much time it takes for humans to complete a certain task. For example, it takes 2 seconds for humans to calculate 4 + 5 + 7, and it may take 1 minute to calculate 37 * 52 * 19.

Then, test whether the large - scale model can complete this task with a 50% success rate.

The study found that the time range of tasks that GPT - 2 can complete with a 50% success rate is 2 seconds; Claude 3.7 Sonnet is 50 minutes; O3 is close to 2 hours; and Opus 4.6 is about 12 hours.

In other words, for tasks that take 12 hours for humans to complete, Opus 4.6 has a 50% success rate.

The result is the above picture. It can be found that the evolution speed of large - scale models is a straight line on a logarithmic scale.

Every 7 months, the time range of tasks that large - scale models can complete with a 50% success rate doubles. According to this trend, large - scale models will be able to complete tasks that take a human expert a month to complete with a 50% success rate between 2027 and 2031.

If this paper is correct, it means that the models released at the end of the year will be twice as powerful as those released at the beginning of the year.

Technology News

1. Easter Egg in the User Agreement

The user agreements of software services are long and difficult to understand. Few users read them, but there is a lot of important content in them.

An American telecommunications operator, in order to show that it attaches great importance to user rights and interests, encourages everyone to read the "User Agreement" and secretly added an easter egg in it.

The highlighted sentence in the above picture reads: "If you read this sentence, please send an email to our mailbox to win a free trip to Switzerland."

Two weeks after it was launched, someone sent an email asking if this was true. Since only one person sent an email, she went to Switzerland for free.

From this, we can see that even with an easter egg, no one reads the "User Agreement". My current practice is to ask large - scale models for help and ask "What are the aspects in this agreement that are beneficial to users?" places that are unfavorable to households" and quickly got the answer.

2. Touch - screen Nail Polish

One problem with widely - used capacitive touchscreens is that they don't work when wearing gloves.

The reason is that it requires the touching object (such as a finger) to be conductive so that the screen can generate an electric - field disturbance to determine the touch position.

The solution is simple: apply a layer of nail polish to the fingertips of the gloves. The metal shavings in the nail polish can conduct electricity.

An undergraduate in the chemistry department in the United States, while studying cosmetic chemistry, invented an improved transparent nail polish specifically for using touchscreens while wearing gloves.

This nail polish is transparent. It's invisible when applied to gloves and can also be applied to bare nails as a nail polish.

3. Copilot Ads

Copilot is an AI assistant launched by GitHub. Last week, users found that it would automatically insert ads.

The above picture is a Pull Request automatically submitted by Copilot. It added an ad at the end of the commit message (in the red - framed area) to introduce the application Raycast.

Searching on GitHub shows that more than 11,400 PRs already contain the same advertising words.

After user protests, GitHub temporarily stopped this feature. But this is a dangerous sign, indicating that GitHub wants to use users to increase revenue.

Articles

1. Review of Xiaomi MiMo v2 Pro (English)

Xiaomi released the MiMo V2 series of large - language models. This is a review by foreign media, which gives a high evaluation.

2. I Generated a JavaScript Engine with AI (English)

The author spent six weeks generating a JavaScript engine that 100% passes the test262 test suite, covering all 98,426 scenarios. This article introduces this matter.

3. Anatomy of the .claude/ Directory (English)

Claude Code will generate a .claude/ sub - directory, and all the underlying data processed by AI is stored in it. This article studies what's actually in this directory.

4. Introduction to Consistent Hashing (English)

Consistent hashing is a cache - location algorithm. In the case of adding or reducing cache servers, it can keep the original cache locations unchanged.

5. How to Use a Laptop as an HDMI Monitor for a Single - Board Computer (English)

The author used an HDMI - to - USB capture card to use a laptop as a monitor for a Raspberry Pi.

Tools

1. EmDash

An AI - generated replica of WordPress, based on the TypeScript language, supports plugins. It is said to have basically the same functions. See the introductory article.

2. SubsTracker

A subscription management system based on Cloudflare Workers that can send various subscription expiration notifications through notification channels such as Telegram and Webhook. (Submitted by @wangwangit)

3. OpeniLink Hub

An open - source WeChat robot message management platform with a built - in app market. By clicking to install apps, functions can be added to the WeChat Bot. (Submitted by @xixihhhh)

There is also a similar project wxWebHook, which sends messages to WeChat users through WebHook. (Submitted by @aristorechina)

4. Lixian.Online

A tool to obtain offline installation packages of VSCode plugins, Chrome extensions, and Docker images, with open - source code. (Submitted by @LiaoGuoYin)

5. Rename.Tools

A batch file - renaming tool on the browser side that supports various rule settings, with open - source code. (Submitted by @chenz24)

6. FontInAss

An open - source subtitle font subsetting tool that embeds the required font glyphs into subtitle files. (Submitted by @Yuri-NagaSaki)

7. pretext.video

A small application based on Pretext (a text - layout calculation library) that displays the human body outline captured by the camera in real - time through text layout. (Submitted by @fifteen42)

8. OxideTerm

A cross - platform SSH terminal based on the Rust language with many functions, using the Tauri desktop framework. (Submitted by @AnalyseDeCircuit)

9. wtree

A graphical management interface for git worktree. (Submitted by @FatDoge)

AI - related

1. Open Agent SDK

An open - source alternative to claude - agent - sdk implemented based on the Claude Code source code, used for the development of AI Agents, fully compatible with the original interface and not dependent on local cli processes. (Submitted by @idoubi contribution)

2. Antigravity Gateway

A web console for unified management of all local AI agents, supporting multi - workspace isolation, Feishu remote collaboration, Skills ecosystem, etc. (@Mr - ZhangBo contribution)

3. ArcReel

An open - source AI video generation workbench. Input a novel and it can automatically complete the script, character design, storyboard, and short - video generation. (@Pollo3470 contribution)

4. TermCanvas

An open - source desktop application where all terminals are laid out on an infinite canvas for easy management of AI programming tools. (@blueberrycongee contribution)

There is also a similar project OpenCove. (@DeadWaveWave contribution)

Resources

1. Claude Code Hands - on Tutorial

An interactive tutorial for Claude Code. Master this AI programming tool through 11 small exercises.

2. Claude Code Unpacked

Based on the leaked source code of Claude Code, it graphically demonstrates step - by - step how the software processes input prompts.

3. Machine Learning Primer

A machine - learning tutorial for engineers, explaining basic concepts.

Images

1. European Tree of the Year

There is a "European Tree of the Year" selection in Europe. At first, it may seem strange, but upon closer thought, you will find that this activity has many benefits: increasing city popularity, promoting ecological protection, and boosting tourism...

Here is this year's "European Tree of the Year" in China.

An oak tree in Rukai Village, Lithuania, with a tree age of 400 years.

Here are other trees that made it to the final.

A wild apple tree in Slovakia

An elm tree in Poland

A linden tree in Latvia

A cypress tree in Portugal

Digest

1. The More I Use AI, the Less I Worry

The more time I spend on AI programming, the less worried I am about my career, even as AI's programming ability gets stronger and stronger.

Because I find that AI programming is just one part of the process. My job is not just writing code.

My real job is to identify problems that can be solved with code, then solve them, and verify whether the solutions are effective.

AI may eventually be able to fully undertake the middle - stage coding

the code part, and help with the first and last parts. However, someone still needs to identify problems, define them, and confirm that they have been solved.

This is 80% of my job.

2、The Unsustainability of Moore's Law

Moore's Law states that the number of transistors on a chip roughly doubles every two years.

However, it also has a concomitant effect that is rarely mentioned. That is, roughly every five years, the construction cost of a chip factory doubles, while the number of chip companies that can afford such costs is halved.

Twenty - five years ago, there were about 40 companies that could build chip factories, with each factory costing about $2 - 4 billion to build. Today, only two or three chip companies (the number depends on your optimism about Intel) can build the most advanced chip factories, and the construction cost has soared to tens of billions of dollars.

If this trend continues for another 10 years and the construction cost of chip factories continues to soar and double, perhaps only one company or no company at all will be able to afford such costs.

Currently, the manufacturing process of chips has approached 1 nanometer. If it develops further, both the technological and financial barriers will approach their limits simultaneously.

I expect that Moore's Law will soon fail, and future growth will mainly lie in computing power rather than the computing power of a single chip.

Future chips will be like second - hand cars, with similar driving speeds, only differing in newness and oldness. I even think that there will be almost no substantial difference between a chip produced in 2035 and one produced in 2065.

Opinions

1、

Accidentally publishing the map file of the source code to npm may seem like an impossible mistake, but it becomes easier to understand when you realize that a large part of the codebase is likely written by the AI you are publishing.

-- Netizen's comment on the Claude Code source - code leak incident

2、

The vigorous development of artificial intelligence may mean that the demand for certain office jobs is not that great, while it will create a large number of jobs for electricians, welders, and plumbers.

Previously, we told all young people to go to college and work in the banking, media, or legal industries. Now, we need to balance this. Some people may be more suitable for manual labor, and they can also have a successful career in fields such as plumbing and electrical work.

-- Larry Fink, the boss of BlackRock, a major US financial group

3、

The purpose of writing is not to finish writing, but to enhance your own understanding and, in turn, the understanding of those around you.

Letting AI write for you is like paying someone to exercise for you.

-- Don't Let AI Write for You

4、

A programmer's job is not programming, but managing software complexity through abstraction. If you do this, then programming will be easy.

-- Your Job Isn't Programming

Retrospect of Previous Years

Manufacturing is Becoming "Gig - like" (#344)

Thoughts on the Battle of Yamen (#294)

Big Data is Dead (#244)

Pessimists are Right, Optimists Succeed (#194)

(End)

    <h3>Document Information</h3>

• Copyright notice: Free reprinting - non - commercial - non - derivative - keep attribution (Creative Commons 3.0 license)
• Publication date: April 3, 2026

This article is auto-translated by AI.