Reconstructing Personal OS in the AGI Era: How Your SFT Dataset Determines Your Class

AI’s era is fundamentally rewriting the logic of personal growth. As the marginal cost of knowledge acquisition approaches zero, the traditional theory of ‘skill moats’ has become obsolete. This article delves into the principles of large model training, sharply pointing out that the environment serves as an individual’s ‘supervised fine-tuning dataset.’ Through various real-life examples, it reveals how to reconstruct one’s social circles and growth paths like training an AI model. In this explosion of computing power, mastering a ‘cloud-native’ growth mindset may be the ultimate algorithm for transcending social classes.

Recently, I learned about some disheartening phenomena: students who learned design a few years ago and were elevated to high-paying positions are now back to square one, re-enrolling to learn new tools.

Many believe that having returning students is a good thing, but I see it as a failure of education. Why do students, once taught with great effort, slowly get eliminated from society once they deviate from their original paths?

As a frontline data engineer and product manager for large models, I increasingly feel that this is not just a matter of ’not trying hard enough’; it is a problem of incompatibility between the underlying system architecture and the era’s version.

If we view ‘personal growth’ as a multimodal AI agent operating in the AGI era, the past system upgrade method was ‘handwritten rules’—learning a specific skill. However, in today’s explosive growth of computing power and model capabilities, this path has become entirely ineffective.

Today, we will not discuss abstract concepts but will focus on the underlying logic of AI product management and data training, discussing how ordinary people can reconstruct their ‘operating systems’ in this era.

1. The Commodification of Knowledge: The Failure of the ‘Rule Engine’

In the past, we firmly believed that the gap between people was due to ‘knowledge differences.’

We thought there was a secret, a textbook, or a master in the world, and once we mastered that key, our lives would take off. For instance, if you were the only one in the industry who could explain the PS curve tool through ‘shadows’ or create rounded corners using edge detection before the rounded corner tool existed, you had an unbeatable moat.

But that is a story from the classical internet era. Today, Figma versions are updated rapidly, and AI plugins are emerging endlessly. On platforms like Bilibili, you can learn any language, fashion tips, and top-notch technical tutorials for free. More cruelly, all existing human data has already been input into AI, meaning that, in principle, AI can perform better than any individual in analyzing books, principles, complex documents, and historical events.

Product Insight: When the marginal cost of acquiring knowledge drops to zero, ‘specific skills’ are no longer scarce resources. Just like how manual 3D point cloud labeling for autonomous driving will ultimately be ruthlessly replaced by efficient automated algorithms, merely investing time in ‘acquiring certain knowledge’ can no longer yield a satisfactory ROI (return on investment) for class mobility.

2. Computing Power Distribution and System Bugs: Why Do We Lack Motivation?

Since knowledge is readily available, why don’t people go learn?

I once saw a delivery worker passionately speaking English during breaks, saying, “My future is not a dream.” I’ve also encountered many ride-hailing drivers who could easily listen to podcasts or study for certifications while working but instead choose to deliver orders day in and day out, spending late nights tipping female streamers.

Why? Because that delivery worker is seen as a ‘freak’ by ordinary people.

Underlying Logic Breakdown: The vast majority of people will not change for years, even decades, because the human brain has an innate ’energy-saving mechanism.’ The brain prefers to execute tasks in the most certain, simplest way and is reluctant to engage the most energy-consuming organs.

More critically, the backlash of the network environment. Humans are social nodes; if your surroundings are filled with colleagues who binge-watch shows and gossip after work, your attempts to improve (like watching Elon Musk’s interviews or studying cutting-edge technology) will be rejected by the system. Colleagues may perceive you as a ‘striver’ and treat you as an outsider.

This environment can lead to immense guilt and self-doubt during midnight reflections. Whether through malicious ridicule or well-meaning attempts to pull you back into the fold, the essence is that the system tries to erase your ‘anomalous data’ and forcibly pull you back to the original mean.

3. Core Architecture Reconstruction: Treating ‘Class’ as a High-Quality SFT Dataset

Since knowledge becomes outdated easily and personal willpower is highly unreliable, how should we design the core of our personal product?

The answer is: autonomously construct your class.

The common understanding of class is sudden wealth (like making 10 million from a demolition, winning the lottery, or being included in a big shot’s group). However, from a data engineering perspective, this is merely incidental ‘pre-training data.’

The true class is a new group you autonomously build that can sustain your current income and striving state. Your environment is your personal SFT (Supervised Fine-Tuning) dataset.

Here is an excellent ‘A/B test’ case:

• Group A (Mainstream Choice): Join a big company, pursue a certain background and identity, and spend evenings skiing, singing, or binge-watching shows.
• Group B (Non-Consensus Choice): Join a lesser-known, low-paying startup, working hard and burning the midnight oil to learn cutting-edge technology.

Three years later, the differences are astonishing. Those in Group A may fall behind due to industry changes and be forced to find new jobs to maintain their class, while someone in Group B, who once couldn’t find a job paying 15,000 a month, persevered in a high-pressure team, preparing lessons late at night, flying to teach on weekends, and absorbing the latest AI papers. They successfully transitioned to a top internet company with a management role, earning nearly 500,000 a year, leading graduates from Tsinghua and Peking University.

Why could she endure such a ’non-human existence’? Because her surrounding team environment was like that. When everyone around you is striving, your efforts won’t be wasted, just like during the college entrance examination when everyone suffers together, you don’t feel the pain.

Class even possesses network repair capabilities. When the founder of Qvod returned to the CEO position after prison, it was because the brothers he had fought alongside had grown during those years, directly lifting him back up.

It turns out that investing in a wealthy background is not as valuable as investing in an environment that provides sustained motivation.

4. Practical RLHF for Personal Large Models: How to Growth Hack Your Circle?

Having clarified that ’environment’ is the core product strength, how can we autonomously deploy our personal environment like configuring a cloud server? You need to treat your attention, energy, and resources as if you were a VC investing.

1. Identify Nodes of Inevitable Growth and Actively ‘Leech’ Off

When building an investment portfolio, we look for future growth targets. The workplace is no different. Look around at your colleagues; if they are all failing, why stay there? If there’s a geek in the company who studies cutting-edge stuff every day, even if they are annoying and don’t include you, you should study what qualities they possess and elevate yourself to their level to engage with them. You must find groups that are guaranteed to grow and collaborate with them to drive your motivation forward.

2. Purchase OOD (Out-Of-Distribution) Data to Reshape Your Effort Threshold

If a large model is only fed homogeneous data, it will overfit (Overfitting), and the same goes for people. You need to intake high-variance data to broaden your horizons. Our team once spent tens of thousands to fly to Dali to visit a teacher named Dabin. He didn’t teach any specific editing or traffic knowledge, but he showed us that a person older than me, who could be blown over by the wind, could live stream for eight consecutive hours! Before that, I thought two hours of live streaming was the limit. This kind of investment has a high ROI; it directly shatters your self-imposed limits and raises your tolerance threshold for ’effort.’ Seeing someone do it and believing you can too is worth more than learning any knowledge.

3. Use ‘Useless Classes’ to Filter High-Quality User Groups

How can you quickly gain access to a high-quality social environment? Enroll in an offline French class, a cooking class, or even a ballet class. You’re not there to learn a skill. Because in the world of exhausted adults who still spend money to learn these things at night, there’s no negativity; they are all full of hope for the future. What you’re buying is not the class but a ticket to enter this ‘high-potential user pool.’ Even if you only sit there playing games, you can absorb the positive feedback from those who believe in tomorrow and strive for improvement.

4. RLHF Agile Iteration: Get Feedback Relentlessly and Pragmatically

When you try to enter a new environment, you will inevitably face resistance. At this point, you must adopt an RLHF (Reinforcement Learning from Human Feedback) strategy: keep interviewing and getting criticized by interviewers. After being criticized, record the issues, research thoroughly using AI, revise your resume, and apply again. After seven or eight iterations, you will have memorized all the interviewers’ critiques, and you will be able to secure a job. This is about abandoning emotional waste and aggressively capturing boundary data, ultimately achieving perfect fitting of the model to the new environment.

Conclusion: Three Years is Enough to Bridge Five Classes

How long does it take for people to bridge four or five classes? In fact, three years is sufficient.

The three years from middle school to high school can separate individuals into top universities and vocational schools. Similarly, three years in the workplace can eliminate someone who enjoys comfort while elevating someone who works hard in a startup to an executive position. The exams never stop; they just don’t reveal the scores after you start working.

In today’s technological explosion, learning a specific software or obtaining a certification is no longer important. The true core competitive advantage of a product is transforming oneself from a ‘standalone execution program’ into a ‘cloud-native distributed system.’

Stop searching for shortcuts in existing knowledge; invest your energy in finding and building an ecosystem that pushes you forward. This is the only algorithm for ordinary people to achieve exponential growth in the AGI era.