Encryption backdoors: The trade-off between security and privacy
Explore the impact of encryption backdoors on privacy and security, and its conflict with efficiency
GLM-5.3-Flash: The AI Model Revolution Driven by Hardware Acceleration in China
Discuss the acceleration effect of China hardware on GLM-5.3-Flash performance and the technical logic behind it
Legal consequences of the Trump administration's ban on Anthropic: The battle between technological freedom and political interference
Analyze how technology companies respond to political interference through legal channels, and explore the boundary between technological freedom and political power
The unit price of Mixed Hy4 is 6 times more expensive, but the measured bill is only 1.8 times more expensive.
Tencent Hunyuan today released and open-source a new generation of Hy4 preview. The parameters of the press conference can be seen elsewhere. What I care about is how expensive and stronger it is than the previous generation. The input unit price of Hy4 is 6.3 times that of Hy3, and the context gives 1.04 million tokens. The same question is run on both sides: Hy4's three questions total 37 seconds and US$0.0044, and Hy3's 50 seconds and US$0.0024. The unit price difference is 6.3 times, and the actual bill is only 1.86 times, because Hy4 thinks much less. The previous generation burned all the 2000 budget on the issue of writing code. 4556 words were written in the thinking area, and not a word of the text was output.
Efficiency revolution in the era of small models
Explore how small models can drive AI development through efficiency improvements
Free big model, who told you how much was left, who made you hit the wall?
Sooner or later, I will encounter the same problem when using a free model: how many times can I adjust it? I asked this again, first to see if each family responded to the question whether the amount was given, and then to send 8 requests in succession to see who would stop me first. Only Groq and Mistral will tell you how much is left in the 10 houses, and the other 8 houses will not have a word. The one with the hardest hit in Card is precisely the one that says nothing: Kimi's fourth request is 429, only 3 times per minute. On the other hand, Mistral nominally fired 4 shots per minute, but actually fired 8 shots in a row but failed to stop them. In addition, NVIDIA's Nemotron-49B officially reached EOL yesterday, and I was still recommending it two days ago.
Qwen3.8-Flash-Next: The future of open source models
Explore the balance between commercialization and community development of open source models
Project Nitter: Exploring the Boundary between Open Source and Copyright
The balance between open source and commercial copyright based on the Nitter project ban
The fastest free model from 8 days ago starts charging today
I tested 21 free large model entrances eight days ago. The fastest one at that time was gpt-oss-120b on Cerebras, 593 milliseconds, and I recommended it as my first choice. Run the same script again today, and it returned "Payment required to access this resource". Another entrance at Cerebras is even more straightforward, and models have been archived. The top one is Groq, 545 milliseconds, becoming the new fastest, and it runs exactly the same model. One more thing: my script in the first round declared these two clear deaths "empty returns" for reasons worth taking a look at by anyone doing usability monitoring.
Childbirth costs: Family choices under efficiency
Explore how childbirth costs reflect the impact of efficiency in family decision-making
Are AI cloning offices really more efficient?
AI cloning colleagues did not improve overall output, but instead led to reduced efficiency due to coordination and supervision costs
Misrecording of military base calls: The efficiency paradox of privacy leaks
Explore the conflict between tool efficiency and personal responsibility in privacy breaches
He used AI to defraud 16,000 yuan, and the county could not buy durian
Residents of Hengshan County, Hunan Province were unable to buy durian and cherries on e-commerce platforms for a while. It was not that they were out of stock, but that multiple platforms listed the entire county as a high-risk area. The cause was that a local unemployed man born in the 1990s placed more than 100 fruit orders in four months. After receiving the goods, he used AI to repair the fresh fruits to a mouldy look. He applied for "refund only and no return". The money was refunded, and none of the fruits were returned. Refund, and sell them at a low price. Police found 75 mobile phones for the crime in his home. But none of his fake pictures were exposed, and the platform discovered that he relied on another thing.
jav.hk VIP: The Price of Business Promise
Analyze the relationship between the fulfillment of business commitments and Internet business integrity
The lawyer asked the AI to find the case, but the judge checked it all made it up
A lawyer submitted two cases to the court, one of which was (2022) Hu 01 Min Zhong Zhong No. 12345. The judge searched this case number and found that the real case was private lending and had nothing to do with the equity holding in this case. He admitted that these two cases were obtained by repeatedly asking AI after refining keywords and did not conduct any verification. This case was included in the People's Court case database this year. Among the 5538 articles in the entire database, only one "artificial intelligence false case" hit it. The case database's description of these two fake cases is the most fatal: it is "highly consistent with the factual details, legal disputes, and adjudication logic of the case, and seems to have strong reference value."
The efficiency paradox of AI painting tools: Qwen-Image-Edit-2511-LoRAS-Fast
Explore the business strategies and efficiency paradoxes behind the efficiency improvement of AI painting tools
The personal matters you told the AI were used for training
You have asked AI about what you feel uncomfortable, quarreled with your partner, whether you still want to do the job, how to manage your children. You have asked AI all these things that you are embarrassed to ask. Where are those words now? I read the original agreement of the seven mainstream AI assistants one by one: It is the norm to use your input for training by default, but Tencent Yuanbao is an exception. Its agreement states in black and white,"Unless you start an experience optimization plan, we will not Use the foregoing content for model optimization." I can turn off the other few stores. I have listed all the switching paths. Only the text part of Qianwen needs to contact customer service. There are two other places that you can't control even if you turn off the switch. You press one almost every day.
Health coverage: The dual benefits of saving lives and the economy
Explore how global health coverage can achieve dual improvements in economic benefits and social well-being by reducing medical costs and improving productivity
He made 85,000 yuan using AI and was fined 485,000 yuan
On the evening of November 30, 2023, a Baijia account posted 13 articles in 42 minutes, saying that a certain soda ash company had damaged its factory by the earthquake, all of which were based on online rumors to get AI to expand. He was also making soda ash futures and earned 85,028.85 yuan in two days. The decision in July this year confiscated all the money and fined another 400,000 yuan. Those 17 articles added up were only viewed 8794 times, which is less than twice the number of one of my hits. None of the five defenses did not stand still. The term used to characterize them was not "knowing it was false" but "unverified." Two and a half years ago, there was an almost identical case. The person changed a word manually, but he was still losing money, so he still fined 200,000 yuan.
I have measured a line of what paid software can free AI replace?
I asked the free AI to recite Article 97 of the Labor Contract Law. It was recited in quotation marks and the format was neat. However, the real Article 97 talked about something else. The rules it compiled were even in the opposite direction. But with the same model, translation, polishing, summary, sorting out messy feedback into tables, and multi-step calculations with system documents, all six items are passed. After twelve tests, the dividing line is not whether the task is difficult or not, but where the content comes from.
Three APIs that claim to require overseas networks can be used directly in China
When it comes to Groq, Cerebras, NVIDIA, and OpenRouter, most strategies add the sentence "requires an overseas network environment." I have classified it this way in previous articles myself, but I have never verified it. This time, I took a clean machine in a domestic computer room and directly tested the 13 entrances equipped in my gateway, and sent a real reasoning request with the key: Cloudflare, OpenRouter, and NVIDIA, which were labeled, did not actually need overseas networks, while Cerebras returned a 1009 error code rejected by region by Cloudflare. Four other networks are completely open, and accounts are stuck.
GitHub Incident: Crisis of Trust in the Open Source Ecosystem
Explore the impact of the GitHub incident on trust in the open source ecosystem and its governance mechanism
Of the 21 free large model entrances, only 9 are actually available
The article six days ago said that 4 out of 13 entrances were hung, and a few more have been added in the past few days. I have adjusted all 21 entrances one by one: 9 are usable, 5 are current-limited, and 7 are dead. But what is more useful is not these numbers, but that the same model is now often hung on several platforms, with completely different usability. The three entrances of gpt-oss-120b have two connections and one limit. The most counter-intuitive one is GLM-4.7. Smart Spectrum has its own current limit, but the one managed by Cerebras can be used instead.
AI smart downgrade: The business strategy and efficiency paradox behind it
Analyze the business strategies behind the intelligent downgrade of AI models and their impact on efficiency