My deepest insight after two years of undergraduate study: cognitive compound interest

Looking back at the end of my sophomore year, the plans I had for the future when I entered college are almost completely different from the path before me today.

When I first started college, I was determined to pursue a purely academic path—direct PhD in North America, climbing up the ivory tower step by step. That was the most reasonable judgment I could make based on all the information I had at the time. Now, I am about to join ByteDance Seed Lab to do pretraining-related research, while holding multiple possibilities in academia, industry, and even early-stage entrepreneurship.

This was not something I planned from the start, nor did it come from sudden luck. I have always believed in one sentence: greatness cannot be planned. I am not saying I will achieve something great, but cognitive compound interest itself has this characteristic—once the compound interest effect truly kicks in, the final magnitude of returns is completely unpredictable from the starting point. You cannot reverse-engineer every step from a preconceived result, because you have no way of knowing how far the exponential growth will ultimately go.

This is my deepest insight from these two years: you do not need to meticulously plan every step; just identify long-term and correct things and persist in doing them, leaving the rest to compound interest. Where you stand today and the choices you can make are never the result of a single point of effort, but the end state that all your past cognition, experiences, and connections have snowballed into.

I like to use the Taylor expansion as an analogy. Your judgment about the future at any point in time is equivalent to expanding the function of life at that point. The higher the cognitive dimension you command and the more complete your information, the higher the order of derivatives you have, and the more accurately you can reconstruct the global curve. If you only have low-order information, you can at best make a linear prediction, which looks straight and clear, but once you go even a little further out, it deviates more and more from the real world.

When I first entered college, I only had that low-order information, so I could only see the direct PhD path. This was not shortsightedness on my part; anyone standing at that cognitive starting point could only make such a judgment. The true meaning of cognitive compound interest is that your principal is never fixed. Every new piece of knowledge you learn, every new person you meet, every experience that refreshes your cognition—they do not just bring you incremental returns, but directly merge into your cognitive principal, becoming the basis for your next judgment. With each additional order of information, the precision of your prediction about the future leaps up a level.

The starting point of all this was, frankly, a spur-of-the-moment decision at the beginning of freshman year.

In September 2024, right after I enrolled, my residential college was supporting student-initiated seminars, and without much thought I took the lead in setting up the school’s first deep learning seminar. At that time, I could not imagine how many things this would lead to. I just felt it was worth doing, so I did it.

During the process of running the seminar, I met two junior faculty members. One brought me into a research internship at our school, and the other opened up my later remote research collaboration. Also because the seminar gained some attention on campus, I got to know many seniors and peers from Xi’an Jiaotong University. Among them was a very impressive senior—he did his bachelor’s at our school, later went to Peking University for a PhD, and before finishing it he left to found an AI healthcare company. That is a story for later.

During that period, I developed a habit of regular output. At first it was the seminar, where I gave the talks myself, and I uploaded the recording of every session to Bilibili. Later it gradually turned into writing blog posts—writing up ideas I found interesting or insightful, along with new algorithms and new technical thinking, and publishing them. Many ideas did not warrant the effort of a formal paper but were genuinely worth sharing, and the blog became the best medium for them.

These outputs had no immediate returns at the time—no extra points, no awards, and they even took up a lot of time outside coursework. But I gradually found that as long as you keep writing seriously, these pieces become cognitive touchpoints you leave on the internet. Many blog posts sparked considerable discussion and helped many people in the industry get to know me. They were the first principal I invested in cognitive compound interest, quietly taking root.

From there, things started happening one link after another.

At the end of my freshman year, in July 2025, I went to Hangzhou to attend an academic conference on machine learning. For the first time I met in person the professor I had been collaborating with, and I also got in touch with researchers from Tsinghua and Shanghai University of Finance and Economics. It was then that I truly stepped into the AI academic circle. Almost simultaneously, because of the Bilibili recordings and continuous output on my personal website, a professor from Renmin University’s Gaoling School of Artificial Intelligence reached out to me, inviting me to collaborate on submitting to a top conference. Without much difficulty, I joined the Gaoling team and began doing real mechanistic research on large models.

By this time, I had a vague feeling that many opportunities are not something you beg for; they are the feedback from everything you have done before, suddenly coming back to you at some point.

What truly caused a qualitative change in my cognition was the month I spent on site at Gaoling over the 2025 winter break.

The most special thing about Beijing is its density. Information density, talent density, and experience density are all incredibly high. Even with the same active effort to absorb, on campus you might run into a conversation that refreshes your cognition only once a month, but in Beijing you meet people from different backgrounds and different positions in the industry every day. It is as if the same cognitive principal earns an interest rate several times higher here, so the compounding immediately picks up speed.

Before going, I wrote a blog post explaining scaling laws from a mathematical perspective and posted it on Xiaohongshu. It was seen by the technical lead of ByteDance Seed Lab’s Pre-Train group, who directly contacted me and invited me to join. At that time, I had already decided to go to Gaoling, so I scheduled the internship for this summer.

During the month at Gaoling, I met senior students in the group and, through networking, got to know friends from Fudan and the quant fund Ubiquant. When I was chatting with the friend from Ubiquant, he said that if I could land total compensation on the order of a million yuan after graduation, there would be nothing wrong with spending two years in industry first, and it would not be too late to go back for a PhD later if I wanted.

That remark made everything click. Before that, I had assumed that the path of life was a one-way street, and that only studying straight through to a PhD counted as the right choice. But that day, I suddenly realized that is not the case at all. The path is never unique. A lot of the most cutting-edge AI research now happens in industry itself. Entering the industry first to accumulate frontline experience, and then returning to academia, might actually give you a clearer view.

This is what I call higher-order derivative cognitive information. When you see one more dimension, the shape of the entire life function changes completely in your eyes.

After that, I became even more certain of one thing: don’t cling to a fixed plan; update your cognition frequently and adjust your judgment in real time. Especially in the AI industry, the landscape shifts roughly every three months. If you still use last year’s cognition to make today’s decisions, you are making decisions about a world that no longer exists.

From then on, things rolled faster and faster.

In April this year, Dong Kehan’s team reached out to me through Xiaohongshu to talk about early-stage support. I declined the first time because I wasn’t ready. But ten days later, they came back, and I realized this was an opportunity to access a higher-density circle, so I agreed to have a deeper conversation.

After meeting them in person, I decided to join. For no other reason than this: I knew clearly that working alongside a group of people at the very frontier of the industry, absorbing their insights, judgments, and betting logic face to face, gives you a rate of growth that going it alone cannot match. Inwardly it means an extremely fast rise in personal cognition; outwardly it means getting within reach of genuinely changing the world. There’s no reason to miss such a chance.

Through that connection, I got in touch with the youngest and most central practitioners at companies like MiniMax, DeepSeek, Moonshot AI, and Tencent Hunyuan, and even had the opportunity to talk face to face with senior people who had DeepMind and Anthropic backgrounds. Every exchange adds to my cognitive principal, making my judgment of the industry a bit more accurate.

In April, I also met in Beijing the senior from Xi’an Jiaotong University whom I had met when I organized a seminar in my freshman year. The AI healthcare company he founded has now raised over 100 million yuan at a valuation far above 1 billion, aiming to use AI to connect the entire chain of life sciences, from upstream basic research to downstream protein synthesis and drug development.

We talked for over an hour, and I asked him whether, at this point in time, doing a PhD is still a sensible choice. My assumption at the time was that a PhD is fine as long as you find the right advisor, and that the pitfalls come mainly from bad advisors. But he turned the question back on me: even if you find a good advisor, won’t there still be pitfalls?

This question stumped me and refreshed my cognition again. I suddenly realized my previous judgment was too simple and black-and-white. When I actually heard this answer from a frontline entrepreneur, combined with all the industry information I had been exposed to, I became even more determined to go into industry first and intern at Seed this summer. Many things have no absolute right or wrong, only what suits the current self, and this judgment must be based on as much and as new cognition as possible.

You see, coming full circle, my connection to this senior who influenced me so deeply traces back to that seminar I started on a whim when I entered as a freshman. This is the most magical part of compound interest: every seed you plant now may bloom in unexpected ways much later.

In early June, I confirmed my summer internship with ByteDance Seed. They told me there happened to be a closed-door talk in the next couple of days and asked whether I was in Beijing. I was still in Xi’an then, and without a second thought I booked a flight and a hotel.

I calculated clearly in my mind that being able to hear the most cutting-edge theories in the industry on site and connect with core team members would bring returns many times the cost of the flight and hotel.

That trip to Beijing far exceeded my expectations. I heard about research progress first-hand, talked with Ziming Liu for a long time, and met the core team of Seed. Then we had dinner together, and the internship was basically settled. (I still went through the standard interview and onboarding process.)

This is the most tangible form of cognitive compound interest: if you actively move toward high-density places and heavily invest in high-value information and connections, opportunities will gradually turn from probabilistic events into certainties.

I posted a tweet about joining ByteDance, which quickly got over 20,000 views, and more than a hundred senior people in the industry followed me, many of whom I had been following one-way for a long time. More interestingly, a few days after signing the offer, a recruiter from Moonshot AI reached out, saying they had seen my blog post “Re-listening to Yang Zhilin: Bet on Scaling, First Principles, and Long-termism” and thought it was insightful, and wanted to chat.

That blog post was just my casual notes from listening to a podcast; when I wrote it, I never imagined it would be read by people at the company in question. But this is the characteristic of cognitive assets: once you write it down and publish it, it stays there, continuously transmitting your signal outward. It doesn’t expire or become invalid; it will always bring matching opportunities to you at some unexpected moment.

Even though I had confirmed my offer, I still scheduled a coffee chat with them. Many people think that once something is settled, there’s no need to explore other opportunities. But I don’t see it that way. Every exchange with top practitioners is not to immediately get an offer, but to absorb new information and expand cognitive boundaries. These things eventually become your principal and pay off at some future point.

By this point it should be clear that none of the paths I am on today were planned back in my freshman year. I simply became convinced that AGI is the direction most worth investing in over the next decade, and along that direction, I do what’s worth doing, meet people worth knowing, and seize opportunities that enhance cognition. I never locked down a specific path.

There’s always this phrase online: what will you do in four or five years if you do this now?

Every time I see such a question, I find it strange. What the industry will look like in four or five years, what opportunities and risks will exist, cannot be accurately predicted by us today. Using your current limited cognition to lock in the whole path for the person you will be four or five years from now is the most irresponsible thing you can do about the future. The world is dynamic, and your cognition should be dynamic too. A truly reliable plan is never about calculating every step, but about being extremely firm on the big direction, keeping the specific path flexible, always using the highest-order cognition you currently have to make the most up-to-date judgment, and adjusting at any time.

Many people understand compound interest as saving enough principal and then lying back to collect interest. But cognitive compound interest is not like that. It requires you to stay open, actively extract experience from the environment, absorb quickly, evaluate quickly, and iterate quickly. Hesitate for a month and you lose one round of compounding. When you’re young, the last thing to fear is trial and error; the cost of trial and error is always lower than the cost of missing out. If you want to do something, just do it; don’t hesitate.

Of course, all of this has a premise: what you show to the outside world (blog, pub, insight, bet) must match your true level. Socializing and output are not about bragging or packaging, not about trading false information for connections. Only when what you say, what you do, and what you output are aligned with your actual ability can these connections be solid and truly start to roll. Connections gained only through packaging are ultimately castles in the air and cannot participate in the real compound interest cycle.

The curve of cognitive compound interest is always flat and long in the early stage. You may do many things and not see any obvious change for a long time, even feeling like you are standing still. But as long as you keep investing principal, keep absorbing, outputting, and connecting, one day you will reach that inflection point, and then everything will start to accelerate toward you.

Two years is just a beginning. Fortunately, now that I understand this curve, I no longer rush for answers.

Take it slow; cognition will do the math for you.

Citation

If you need to cite this article, please refer to:

@article{zou2026cognitive-compounding,
  title={本科两年，我最深的感悟：认知复利},
  author={Zou, Jiaxuan},
  journal={Jiaxuan's Blog},
  year={2026},
  url={https://jiaxuanzou0714.github.io/blog/2026/cognitive-compound-interest/}
}