Tech Enthusiast Weekly (Issue 391): The Wealth Gap in the Age of AI
Tech Enthusiast Weekly (Issue 391): AI Wealth Inequality
This column records noteworthy tech content worth sharing each week, published on Fridays.
This magazine is open source, and we welcome contributions. We also offer a "Who's Hiring" service that publishes programmer recruitment information. For partnerships, please contact via email ([email protected]).
Cover Image
Wall decoration from a Shanghai restaurant. (via monana3838@Threads)
AI Wealth Inequality
I increasingly feel that AI is different from other technologies. It brings not just technological change, but also social transformation.
Simply put, AI will create wealth inequality.
Other technologies actually eliminate wealth inequality and achieve "consumer equality"—where rich and poor consume the same products.
For example, everyone drinks the same Coca-Cola, uses the same Apple iPhone, drives the same Tesla. Even the internet is like this; the world's richest person Elon Musk uses the same websites and mobile apps as you do.
But large language models are different. In front of large models, the poor and the rich are unequal.
In the future, ordinary people definitely cannot afford top-tier large models. In fact, this is already the case. The most expensive AI coding package is Claude's Max plan at $200 per month, which many people already cannot afford.
OpenAI once envisioned a $20,000 per month package offering top-tier, unlimited access to large language models.
If they actually launched it, only the wealthy could afford it.
This reflects a simple fact: the more expensive the service, the better the model performance. Because model performance is related to computing power, more computing resources, larger context windows, and more parameters all require money.
This is completely opposite to industrial products. Industrial goods have economies of scale—higher production volumes lead to lower unit costs. Once mass production is achieved, prices keep dropping.
But large models lack economies of scale. Mass-scale operation of models requires more servers, which doesn't reduce unit costs; it might even increase costs due to data center expansion, circuit redesign, and water system upgrades.
Future society will probably look like this: the rich and poor use different models. Premium model services—planning, consulting, content generation, automation—will require high fees, while ordinary people use free models with ordinary results.
However, I also noticed that Musk recently said there might be another possibility.
His point is that computing power is essentially a form of energy conversion. Humanity will eventually achieve abundant cheap energy supply (space-based solar power?), making computing power cheap enough that everyone can use the best models.
Is it possible? I don't know. The first scenario seems more realistic.
A Method for Measuring Model Capability
How do we measure large language model capabilities?
The current method uses a test set to calculate model scores. Its drawback is that it only allows horizontal comparison and makes it hard to measure progress speed.
Recently, a paper proposed a new measurement method.
Scientists first calculated how much time humans need to complete certain tasks. For example, calculating 4 + 5 + 7 takes humans 2 seconds, while calculating 37 * 52 * 19 might take 1 minute.
Then they tested whether large models could complete these tasks with 50% success rates.
The research found that GPT-2 can complete tasks with 50% success rate in a 2-second timeframe; Claude 3.7 Sonnet is 50 minutes; O3 is close to 2 hours; Opus 4.6 is about 12 hours.
In other words, Opus 4.6 has a 50% success rate on tasks that humans need 12 hours to complete.
The result is shown in the graph above, which reveals that large model evolution follows a straight line on a logarithmic scale.
Every 7 months, large models can complete tasks with 50% success rate in a timeframe double the previous one. By this trend, large models will be able to complete tasks requiring human experts one month of work with 50% success rates between 2027 and 2031.
If this paper is correct, it means models released at year-end will be twice as powerful as those from the beginning of the year.
Tech Updates
1、Easter Egg in User Agreements
Software service user agreements are long and hard to understand, and few users read them, yet they contain many important provisions.
An American telecommunications operator, wanting to show it values user rights and encourage everyone to read the Terms of Service, secretly added an Easter egg inside.
The highlighted sentence in the image reads: "If you read this sentence, email us and win a free trip to Switzerland."
Two weeks after launch, only one person emailed asking if it was real. Since only one person responded, she got the free trip to Switzerland.
This shows that even with Easter eggs, no one reads Terms of Service. My current approach is to use large language models to help by asking "what unfavorable terms does this agreement have for users," and I get answers quickly.
The widely used capacitive touchscreen has a problem: it stops working when you wear gloves.
This is because it requires the touch object (like your finger) to be conductive, allowing the screen to sense electrical field disruption and determine touch location.
The solution is simple: apply nail polish with metallic particles on glove fingertips. These metallic particles conduct electricity.
An American chemistry undergraduate, while studying cosmetic chemistry, invented an improved transparent nail polish specifically for using touchscreens while wearing gloves.
This nail polish is transparent, invisible on gloves, and can also be applied to bare nails as a polish.
Copilot is GitHub's AI assistant. Last week, users discovered it automatically inserts ads.
The image shows a Pull Request automatically submitted by Copilot, where it added an ad in the submission description (red box area) promoting the Raycast application.
Searching on GitHub, you'll find over 11,400 PRs already contain the same ad text.
After user protests, GitHub temporarily stopped this feature. But this is a dangerous signal showing GitHub wants to use users to increase revenue.
Articles
1、Xiaomi MiMo v2 Pro Review (English)
Xiaomi released the MiMo V2 series of large models. This is a review from international media giving very high praise.
2、I Generated a JavaScript Engine with AI (English)
The author spent six weeks generating a JavaScript engine that passes 100% of the test262 test suite, covering all 98,426 scenarios. This article introduces this achievement.
3、Anatomy of the .claude/ Directory (English)
Claude Code generates a .claude/ subdirectory where all AI-processed underlying data is stored. This article investigates what's inside this directory.
4、Introduction to Consistent Hashing (English)
Consistent hashing is a cache location algorithm that doesn't change the original cache location when adding or removing cache servers.
5、Using Your Laptop as an HDMI Display for Single-Board Computers (English)
The author used an HDMI to USB capture card to use a laptop as a display for Raspberry Pi.
Tools
1、EmDash
An AI-generated WordPress clone built with TypeScript, supporting plugins with basically identical functionality. See the introduction article.
A subscription management system based on Cloudflare Workers that sends expiration notifications for various subscriptions through Telegram, Webhook, and other channels. (@wangwangit contribution)
An open-source WeChat bot message management platform with a built-in app store where you can click to install applications to add functionality to your WeChat Bot. (@xixihhhh contribution)
There's also a similar project wxWebHook that sends messages to WeChat users through webhooks. (@aristorechina contribution)
A tool for obtaining offline installation packages for VSCode plugins, Chrome extensions, and Docker images. Code is open source. (@LiaoGuoYin contribution)
A browser-based batch file renaming tool supporting various rule settings. Code is open source. (@chenz24 contribution)
An open-source subtitle font subsetting tool that embeds required font glyphs into subtitle files. (@Yuri-NagaSaki contribution)
A small application based on Pretext (a text layout calculation library) that captures human silhouettes from cameras and displays them in real-time through text layout. (
📝 本文来自抖文 www.douwen.me ，转载请保留出处。 原文链接：https://douwen.me/archives/637/
💬 评论 (1)
This issue hits hard. We're watching wealth concentration happen in real-time with AI, and most people don't even realize it. The tech giants hoarding these models while regular folks can't compete—it's the new gold rush and we're all just spectators.|Nickname2|Great timing on this piece. Quick question though: do you think open-source AI projects can actually level the playing field, or are we fooling ourselves?|Nickname3|Finally someone talking about this! I've been saying for months that AI will make inequality worse before it gets better. The access gap is real.|Nickname4|Solid analysis but I'd add that education is key here. We need to train the next generation NOW on AI tools, otherwise the wealth gap becomes a skills gap too. There's some promising initiatives in this space that deserve coverage.|Nickname5|Honestly feeling pretty pessimistic after reading this. If the wealthy control the AI infrastructure, what leverage do workers even have? This feels inevitable.|Nickname6|Appreciate the perspective. One thing missing though—what about the countries being left behind? This isn't just about rich vs poor individuals, it's about which nations get to participate in the AI economy at all.