Weekly for Tech Enthusiasts (Issue 390): Without Corpus, Large - scale Models Are Intellectually Disabled

Here records the technological content worth sharing every week and 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 by email ([email protected]).

Cover Image

The colorful covered walkway in a residential community in Rizhao, Shandong. There is also a coffee shop in the forest at the entrance. (via)

Large Language Models are Stupid without Corpus

If we conduct a questionnaire survey now and ask people, "Do you think large language models are intelligent?"

I believe that most people will answer yes.

Even though we are only in the early stages of AI, large language models can already replace a lot of human intellectual labor, which is truly amazing.

However, we should not forget the real situation. Large language models are not magic, nor are they "silicon - based intelligent agents" with autonomous intelligence. They are language models based on statistical laws, and all their behaviors are based on mathematical calculations.

The best evidence is that if we ask them to solve problems that have not been trained on, that is, problems where there are no statistical laws, they simply cannot solve them.

This is an experiment I want to share today.

Two foreign researchers selected five mainstream large language models: GPT - 5.2, O4 - mini, Gemini 3 Pro, Qwen3 - 235B, and Kimi K2.

They asked the large language models to use five niche programming languages - Brainfuck, Befunge - 98, Whitespace, Unlambda, and Shakespeare - to program and solve various problems.

The common feature of these niche languages is that there is very little information about them on the Internet, so they cannot be used to train large language models. Guess what the result was?

The result of the experiment can be summed up in one sentence: the performance of the large language models was a mess.

The average accuracy rate of these five large language models in answering questions was only 3.8%, that is, they could answer 3.8 out of 100 questions correctly. In contrast, their accuracy rate in handling Python problems can reach 90%.

Even more embarrassing is that the few questions that were answered correctly were all at the introductory level. For more difficult levels (beginner, intermediate, advanced), the accuracy rate of all five large language models was 0.

This experiment fully shows that the performance (intelligence level) of large language models is first determined by the training materials: the more training corpus, the better the performance. For example, there is a large amount of Python corpus everywhere, so large language models are extremely good at solving Python problems; the less training corpus, the worse the performance of large language models, which is almost as useless as being stupid.

Then, a curious question arises: if there is no corpus for a certain niche language, but there is a very detailed "User Manual", and we let the large language model learn this manual, can it learn to program in this niche language?

MAI - Image - 2

This week, Microsoft released its own image - generation model MAI - Image - 2.

The quality of the images generated by this model is very high. Some reviews believe that it is currently second only to Google's nano - banana - 2.

Microsoft has opened the website MAI Playground (image below), and now free image generation is available.

After I tried it, the texture of the images is really good and very realistic. For example, a dog riding a bicycle in the sea.

However, it has many usage restrictions: (1) It will refuse to generate controversial or potentially offensive images; (2) The daily free quota is 15 images, and the interval between each generation is 30 seconds; (3) It can only generate images with an aspect ratio of 1:1, and other resolutions are not supported; (4) It does not provide image editing and processing, and can only be used for "text - to - image" generation.

If you need to generate high - quality images from text, you can give it a try.

Technological Trends

1. Playable Cover

Red Bull has launched a paper - based game magazine, 《GamePop》.

Its cover has a playable "Tetris", which is the world's first book with a playable game on the cover.

The secret is that a very thin flexible circuit board is embedded inside the cover.

This board is equipped with 180 RGB LED lights, 7 capacitive touch - buttons, and a 32 - bit ARM chip.

It also contains a rechargeable battery that can be charged via Type - C.

Unfortunately, this cover is a limited edition and not for public sale. It has been officially licensed by the Tetris Company, with only 150 sets released globally, and each set has an independent serial number.

2. Paid human customer service

Enterprises don't like to provide human telephone customer service because of the high cost and prefer to change it to machine - answered telephone customer service.

HP has come up with an idea to drive users towards machine customer service.

When users call HP's customer service hotline, they will hear a voice prompt asking them to visit the official website to find answers on their own. If they insist on human customer service, they have to wait online for 15 minutes.

If they hang up the phone halfway and call again, they need to wait another 15 minutes. The system will also remind them at the 5th, 10th, and 13th minutes that they can visit the website or contact by email.

Although this practice is abominable, it may become the norm in the future: only AI or robot customer service is free, and human customer service requires an additional fee.

3. How to play frisbee

How can we throw a frisbee fast and far?

An American physicist conducted an experiment with dozens of students, throwing frisbees with different gestures and angles. He measured the flight speed and torque and wrote a paper on the results.

He found that placing the thumb about 3 cm from the outer edge of the frisbee can achieve the best results in terms of average rotation speed and initial speed.

He also found that there is a linear correlation between the rotation speed and the initial speed. The higher the rotation speed, the higher the initial speed.

So, the next time you play frisbee, place your thumb in the right position, then use all your strength and throw it backhand for the best results.

Articles

1. The slow collapse of MkDocs (English)

MkDocs is a well - known document website generation tool, but there are intense conflicts and confrontations among the main contributors, leading to the fragmentation of this project. This article sorts out this matter.

2. Large - model prediction of coffee heat dissipation (English)

The author asked various large models to give formulas for the heat - dissipation time of coffee, then measured the actual heat - dissipation time and obtained a ranking list.

3. The next app is likely to be a headless app (English)

If we all use mobile phones through AI assistants in the future, then various apps won't need a display module (headless) and only need to provide data interfaces to AI assistants.

4. A method of front - end data compression on the web (English)

This article introduces how to compress data into an image through the canvas on the front - end.

5. Ruby is the best language for building AI applications (English)

The author used Python, JavaScript, and Ruby to write an AI Agent. After comparison, he believes that Ruby is the most convenient for writing AI applications.

6. Ancient Roman concrete architecture (English)

The ancient Romans discovered concrete and learned to use it to pour buildings. As a result, ancient Roman buildings had the largest interior spaces in ancient times and were extremely sturdy, and have been preserved to this day.

Tools

1、proxychains-rs

A Rust implementation of proxychains4 that specifies which process goes through the proxy chain. (Submitted by @tianrking)

2、Flare Stack Blog

A blog system based on Cloudflare Worker, integrating services such as D1, R2, KV, and Workflow. (Submitted by @du2333)

3、Tunelo

Expose local services to the public network with a single - line command, only requiring a single 4MB binary file and using the QUIC protocol. (Submitted by @jiweiyuan)

4、ReadAny

An e - book reading tool for desktops and Android, with built - in AI functions, voice reading, and multi - device synchronization. (Submitted by @codedogQBY)

5、RaTeX

A pure Rust implementation of a KaTeX - compatible math rendering engine that natively parses and typesets LaTeX math formulas and supports various environments. (Submitted by @erweixin)

6、Work Review

An open - source Win/Mac desktop application that continuously records the applications used and websites visited on the same day in the background, making it easy to organize into a personal work trajectory. (Submitted by @wm94i)

7、Valdi

A UI framework released by SnapChat, which allows you to write components using syntax similar to React and then compile them into native iOS, Android, and macOS applications.

8、Npflared

A tool for setting up a private NPM mirror, suitable for enterprises to provide internal JS packages.

9、Chokidar

A Node.js module for listening to various file - system events (addition, deletion, editing, etc.), which is more powerful than the native fs.watch / fs.watchFile functions.

AI - related

1、WeChat's Lobster API

WeChat officially released the Lobster API this week, allowing AI robots to send messages to WeChat.

Many projects use this API for secondary development to facilitate the access of various Bots and Agent gateways.

• wechat-rs-sdk (Submitted by @tianrking)
• WeChatBot (Submitted by @jiweiyuan)
• WeClaw (Submitted by @idoubi)
• WeChat-Channel (Submitted by contribution)

2. AI CLI Complete Notify

A cross - platform desktop application that sends task - completion reminders when AI command - line tasks (Claude code/Codex/Gemini) are completed. It supports various channels (Feishu/DingTalk/Qiwei Webhook, Telegram, email, desktop/sound prompts). (@ZekerTop contribution)

3. Claude Config Manager

A desktop management tool for Claude resources (Skills, MCP, Agent) on macOS, providing a graphical central console. (@Daydayoneup contribution)

4. TrustClaw

A modified version of Lobster OpenClaw that tries to eliminate the risk points in the code.

Resources

1. Project N.O.M.A.D.

A Linux application that integrates various human knowledge (Wikipedia, world maps, online courses, local AI assistants), etc., for offline access. (@15x3 contribution)

2. AI Coding Agent for Data Analysis (English)

The teaching materials of the training course by the famous developer Simon Willison, which uses AI tools for data analysis with detailed steps.

3. The Concise TypeScript Book

An open - source TypeScript tutorial with a Chinese version.

Images

1. Apple Wallpaper Easter Egg

Apple recently released a new laptop, the MacBook Neo. As before, it comes with a special wallpaper.

The product name is embedded in the wallpaper, and previous wallpapers also had this easter egg.

iMac

MacBook Pro

iPad Air

MacBook Air

iPad Mini

iPad Pro

1. Child Mortality

It's hard for modern people to imagine that for most of human history, the child mortality rate (death before adulthood) has been close to 50%.

In the above - figure, the red line represents the infant mortality rate, which has remained stable at around 50% until the late 19th century, when it began to decline rapidly.

In 2020, the global average infant mortality rate was 4.3%, and the lowest - rate countries had reached 0.3%.

Abstracts

1. Don't Become a Machine

I recently saw a sentence: "Only slaves quantify their own existence value through productivity."

Yes, the higher the productivity, the more valuable the slave.

This reminds me that today's social media is full of a lot of hustle culture, and many people show how hard they are working to improve their personal productivity.

In my opinion, this is comparing oneself to a machine. People believe that if they can receive instructions and efficiently achieve a certain goal like a machine, the more valuable they are and the more likely they are to achieve success in life.

On social media, this "hustle culture" has many manifestations: (1) You're not working hard enough. (2) You have to get up at 5 am. (3) You have to be the first to arrive and the last to leave.

Behind this culture, is the requirement for people to become machines.

Machines are indeed very efficient, but there is a problem: they are rigid, operating in a fixed pattern and at a linear speed, unable to automatically adapt to environmental changes or learn the rules of the game.

You are not a machine; you are a human being. Your characteristic should be flexibility and the ability to adapt quickly. Instead of pursuing extreme hard - work, you should find the most valuable solutions. You should focus on the truly important factors: speed, efficiency, or quality, and don't be obsessed with boring work.

Quotes

1、

We have created a civilization in which the most important elements are deeply dependent on science and technology, but we have also made science and technology so difficult to understand. This will lead to disaster. We may escape by luck for the time being, but sooner or later, this combustible mixture of ignorance and power will explode.

-- Carl Sagan

2、

There used to be traffic jams in Paris all the time. The mayor came up with a solution by greatly reducing the number of parking spaces, and indeed, fewer people drove later.

-- CNN

3、

A study found that under remote teaching, the assignment scores of students with outstanding looks were lower than those under face - to - face teaching.

-- Economics Letters

4、

The thing that has influenced me the most in recent years is that I have become a "day - type" person.

I used to stay up late often, sometimes until dawn. In the past five years, I have forced myself to develop the habit of getting up early. Now, my life is all during the day. Seeing the dawn and dusk with my own eyes makes me feel at peace, and my life is in harmony with the natural cycle.

-- Becoming a Day - Person

5、

AI is very good at turning clear ideas into runnable code. What really takes time is figuring out exactly what I want to develop.

-- lustin.fr

Retrospective of Previous Years

How to Stop AI Crawlers (#343)

A Week Is 2% of a Year (#293)

Chatting with Confucius AI (#243)

Which Is Harder, Front - end or Back - end? (#193)

(End)

    <h3>Document Information</h3>

• Copyright notice: Free reprinting - non - commercial - non - derivative - keep attribution ( Creative Commons 3.0 license )
• Publication date: March 27, 2026

This article is auto-translated by AI.