Current Information#
Time Period: 25/02/12-18
Keywords: Liang Jianzhang, demography, university, education return rate, elite education, interpersonal communication, privilege, family relationships, marriage, deepseek, thinking chain, AI, Jiang Nengjie, vulnerable groups, independent producers, documentary, independent film, cloud storage director, film industry
Information Stream#
【Why did Ctrip's Liang Jianzhang publish an article in TOP, asking why you must go to university today when the education return is extremely low?】#
(Liang Jianzhang, demography, university, education return rate) https://mp.weixin.qq.com/s/upcy9SCpFftjmhPz-lEPPw
Simulation is more intuitive: If we simply expand the enrollment of university students (supply shock), the wages of all educational groups will fall, which is inconsistent with reality. Only when the demand for skills simultaneously increases (demand shock) can we reproduce the real trend of "newcomers falling, old-timers rising." This indicates that the essence of university enrollment expansion is an active adaptation to industrial upgrading and cannot simply be labeled as "educational surplus." Therefore, human capital investment must consider long-term value and cannot deny educational input due to short-term market fluctuations. Even if developing countries accelerate enrollment expansion, it is difficult to replicate the decades of accumulated experience of skilled talent reserves in developed countries in the short term. Just like brewing requires time, the "aging process" of experience accumulation cannot be rushed, reminding our policymakers to view talent cultivation with more patience.
【Why is Chinese education just as competitive as elite education in Europe and America, yet few criticize elite education in the West?】#
(elite education, interpersonal communication, privilege) https://www.zhihu.com/question/3291267744/answer/91515820151
Humans, as naked apes, are extremely social animals, unconsciously absorbing values, thoughts, skills, and desires from those around them. Those who ultimately achieve extraordinary things are often surrounded by a group of equally extraordinary adults during their critical growth periods. … These parents strive to create a rich knowledge environment.
【What is the most brutal truth behind the "cat-eating forced marriage" incident in Guangxi?】#
(family relationships, marriage) https://mp.weixin.qq.com/s/lLDpb4DvVv9YtNEi0X4J_A
This is a test of obedience. It is simply to make the child obedient. Killing what you care about most, if you do not speak up or resist, then fine, from now on you will be firmly controlled by them. If you resist, you are angry, you struggle, then you are unfilial, disregarding family ties. In many details, elders must demonstrate absolute authority in front of juniors, so they constantly test. In the end, they gain an obedient child, and their goal is achieved.
【A ten-thousand-word analysis of the beauty of creation by DeepSeek: How was DeepSeek R1 developed?】#
(deepseek, thinking chain, AI) https://mp.weixin.qq.com/s/Pbha8I1-j174ra9uWXG32Q
Peak once shared a penetrating insight with me. As we know, the essence of language models is to activate a vast neural network matrix. When a token is input, it can activate certain parts of the matrix, but this activation is limited. When more tokens are input, more parts can be activated, and the amount of information increases accordingly. Therefore, more tokens mean the model can obtain sufficient information to make more precise decisions.
The model needs more tokens to "think," which led us to propose the concept of the Reasoning Model.
What is a reasoning model? For example, we can illustrate with a question: "How many stops does it take to get from Wangjing West to Xizhimen by subway?" A "direct answer type" model might directly respond on the left side of the image: "Nine stops."
In contrast, a reasoning model would provide the answer on the right. It would first consider various transfer routes, then compare the number of transfer stations for each route, and finally derive the best solution. The reasoning model not only provides an answer but also shows its thinking process.
……
First, the article mentions that on September 12, o1 was released, shocking the world, and then the team noticed that long CoT was extremely effective. They realized they had to invest in long CoT, or they would be left behind. Thus, they began to think about how to draw inspiration from OpenAI's work and discovered two key videos during their research.
These two videos released by OpenAI were not from September but earlier talks by Noam Brown and Hyung Wong Chung. These videos were only made public when o1 was released, prompting them to wonder: why release these videos at this time? It must be related to the training of o1.
When I saw this analysis, I thought, "This perspective is amazing." So, they delved into these two videos and first discovered a key slide in Noam's video that mentioned AlphaGo and its subsequent version AlphaGo Zero. Everyone knows that AlphaGo Zero is a version entirely based on reinforcement learning (RL), and this slide emphasized Test-Time Search.
Many believe Noam emphasized this part to explain AlphaGo's MCTS, i.e., Monte Carlo Tree Search—exploring multiple paths, evaluating scores, and ultimately finding the optimal solution. However, the Kimi team had a non-consensus judgment: they believed Noam was actually emphasizing the S in MCTS, which is Search itself, rather than the specific MCTS. This understanding led to their first key idea: let the model search on its own! Allow the model to learn to explore different paths rather than artificially limiting its thinking.
This reminded them of Richard Sutton's famous talk "The Bitter Lesson."
The content of the second video was equally crucial. They summarized a core point: "Don’t teach, incentivize." In other words, do not "teach" the model, but rather "incentivize" it to explore autonomously.
In many experiments, the less structural constraints a model has, the higher the ultimate performance ceiling when computational resources increase. Conversely, if too many structural constraints are imposed on the model early on, its final performance may be limited, losing more opportunities for autonomous exploration.
They further pondered: why did this student emphasize structure so much? What is structure? At that time, I found this particularly enjoyable, as it felt like I was witnessing Kimi's internal dialogue.
MCTS is structure, and the A* algorithm is too. These all limit the model's ability to think freely. They believed that OpenAI's PM-800K training method also had similar issues—it tells the model how to think in different situations through a well-formed reasoning dataset. This essentially sets a predetermined thinking path, limiting the model's own exploration capabilities.
They ultimately concluded: o1 did not restrict how the model thinks. This point is particularly important. The Kimi team thus decided not to adopt MCTS.
……
Based on this model, they used pure reinforcement learning for training, but the process itself was very simple. They used a fixed template during training.
Now everyone should be familiar with AI products, which can be understood as a system prompt. Specifically, the system prompt was set as "This is a conversation between a user and an assistant," where the user asks questions and the assistant responds. However, the assistant must first "think through the reasoning process in its mind" before providing the final answer. Additionally, the assistant must label the reasoning process within the think tag, while the answer is placed within the answer tag.
……
The R1 Zero path was correct, but then the DeepSeek team discovered some issues with R1 Zero. First, readability was poor. Second, there was often a problem of mixed language, similar to how expatriates in Shanghai might speak: "Maria, today's schedule is a bit full."
This issue is not only present in R1 Zero; most reasoning models exhibit this problem of mixed language.
Recently, there was a meme that some foreign netizens captured the thinking process of o3 and found that when o3 asked questions in English, it reasoned in Chinese. While we understand the true situation behind this phenomenon, many foreign netizens still captured it and @Sam Altman, asking if they were distilling DeepSeek R1. The wheel of karma turns.
The reason behind this is simple. When the model explores, for it, both Chinese and English are just tokens. It processes problems according to tokens during its thinking, regardless of whether humans can understand it. This is actually a language mixing issue.
Notes#
Training the AI's thinking chain is, to some extent, a reproduction of human logical reasoning patterns.
【Have we gone viral? A well-known "cloud storage director" on Douban made a film that cannot be publicly screened, yet has over 100,000 views】#
(Jiang Nengjie, vulnerable groups, independent producers, documentary) https://zhuanlan.zhihu.com/p/358299999
In 2016, I was fortunate to be selected as a "Ginkgo Partner." We had a Ginkgo overseas inspection project, and one of the partners, Feng Lu, was involved in services for people with mental disabilities. Coincidentally, we both participated in study tours in Japan and Australia. We visited and learned from the local disability institutions and experienced how urban construction cares for people with disabilities, deeply understanding their concern for vulnerable groups, especially in Japan. Of course, the attitude towards the weak is a mark of a country's civilization.
【The director on the cloud storage: "Let me walk the path of film"】#
(independent film, cloud storage director, film industry) https://m.thepaper.cn/baijiahao_7555277
Cheng Jiabin admitted to Southern Weekend reporters that independent directors in China are trapped in a negative cycle: due to a lack of funding, they cannot produce high-quality works, and because they are unwilling to compromise, they habitually choose to create works leaning towards self-expression, which often leads to poor outcomes. "Most independent directors make auteur films (referring to films with obvious personal style characteristics), they film their own lives, immersed in a somewhat excited state, but this prevents them from considering the audience's feelings."
Wang Si's films have won awards at two South Asian film festivals and have been nominated at several small film festivals in the UK and Italy. When he attended a film festival, he watched many films from various countries and felt that "world cinema is developing towards smaller scales"—some people film a stone in the sea; others film a person walking for three whole hours…
Wang Si once tested his videos on Bilibili, editing several videos in a day and posting them on different channels, discovering that what netizens liked were still entertainment content. He felt disappointed, "Those popular videos are achieved by degrading self-worth."
Thoughts#
Why Focus on 706 Youth Space: The Disappearance of Proximity and Public Life#
Some friends have asked me why I focus on entities like "706 Youth Space" and "1200BookShop." There are four reasons: to fill the void in private life; to meet more people; to access more information outside my perspective; to seek the lost "proximity" and public life.
The motivation stems from the boredom of personal life. Centered around home, within a 5-kilometer radius, I was surprised to find no interesting public activities or venues, as most have been privatized and made exclusive, such as KTVs, internet cafes, and billiard halls.
Young people aged 20-30 are a contradictory group, in a transitional state, continuing the vitality, innocence, and confusion of adolescence while also inheriting the melancholy and anxiety of adult society; unable to integrate into the middle-aged community, yet unable to retreat to the childlike society dominated by students; partially liberated from the family life that dominates individuals, yet troubled by the "emptiness and loneliness" of personal life. One could argue that youth is a restless transitional phase in life.
A friend recommended a book to me, "Science and Dignity." Coincidentally, I found a book by Xiao Suo Wei titled "Desire and Dignity: Class, Gender, and Intimacy in Transitional China" on zlibrary, which mentions in the preface: "In any case, since the 21st century, the uncertainty and insecurity brought by the risk society to modern people have been increasing day by day, increasingly permeating people's ambivalent attitudes towards the ancient institution of family. Especially for East Asian societies in the so-called 'compressed modernity,' on the one hand, families have become fragmented under the irreversible trends of aging and declining birth rates, while on the other hand, the sense of dependence on family by both the state and individuals seems to be increasing. In this state, the myth of the modern nuclear family appears to be powerless in responding to the changing society, while individuals drifting between family and society, holding the banner of autonomy, face a sense of homelessness, not knowing where to settle themselves."
This passage awakened me: youth, both homeless and caught between private and public life, yearns for dialogue with others and society, longing for a "home."
Every society's public life is a collective composition of "private," "semi-public," and "fully public" spaces and activities. However, the combination of rising commercial rents, the expansion of internet socialization, and the authorities' strict prevention of gatherings has led to most physical spaces being absorbed by capital, becoming traps of consumerism—commercial streets, supermarkets, amusement facilities, and other consumer venues have emerged. Thus, I awkwardly discovered that activities where people could gather offline to chat have mostly turned into consumer scenarios, such as dining, karaoke, drinking, and internet use. Otherwise, it’s just about going out shopping, visiting parks, exhibitions, or museums…
In terms of physical spaces, Xiang Biao mentioned that "proximity" has disappeared, and even the public life supported by each "proximity" has vanished.
Various "706 spaces" have addressed the demand for "semi-public and semi-private" offline activity venues. Online information media and social platforms are all monitored and regulated, making it difficult to achieve results when discussing public issues: either everyone actively avoids public topics and dares not speak; or if they "talk too much," they are asked to leave by the platform or community managers. Topics that cannot be discussed online can be openly discussed offline, even passionately debated. However, a new problem arises: we lack venues that satisfy the nature of "semi-public and semi-private"—parks, libraries, etc., are public spaces, uncontrolled and devoid of privacy; who would dare invite not-so-familiar netizens or strangers directly to their home? If you want to hold a public event, you can only seek support from entities like "706 spaces," which operate in a semi-commercial or even fully commercial manner (like Guangzhou's "Mujiji Tea House").
Middle Class and the Global Political Economic Landscape#
"We know that the global political and economic landscape has undergone profound changes over the past few decades. Some Chinese scholars discuss the 'rise of the East and decline of the West,' while others propose different views. From the data, the rise of the East and decline of the West has indeed occurred. For example, compared to the 1960s, the GDP of the United States has actually decreased by 50%, while China has transformed from a relatively weak developing country into the world's second-largest economy. If we compare the rise of the East and decline of the West across Asia with traditional North America or Europe, it becomes even more apparent. According to World Bank research, by 2030, four of the world's five largest middle-class countries will be in Asia, including China, India, Japan, and Indonesia, with the only one outside Asia being the United States. This also indicates that the entire global political and economic landscape, especially the economic landscape, has undergone fundamental changes." (https://www.hnzk.gov.cn/zhikuyanjiu/19024.html)
Consumer power and technological power are the two main powers today. Whoever has a larger population market with consumer capacity holds consumer power. Whoever controls key links in the high-tech industrial chain holds technological power. These two powers can change hands.
Based on the distribution of the middle class, languages such as Chinese, Japanese, Hindi, and Indonesian will (already) become secondary instrumental languages besides English. Africa is still in its infancy, and I expect to enjoy the benefits of Sino-African exchanges only after I turn 40; among them, various languages are mixed, with some still using French, so learning French can also be considered. History has repeatedly proven that the degree of political organization is the most powerful force in human society; Rwanda has a promising future, while the Democratic Republic of the Congo may fall into long-term chaos.
When China and Indonesia develop, it means that East Asia and Southeast Asia will also follow suit, the Middle East will remain relatively stable, and Asia (the Chinese cultural sphere) will truly become another pole outside of Europe and America. This will mark the "rise of the East and decline of the West."