Two years of undergrad: my deepest takeaway is cognitive compounding

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 in front of me today.

When I started university, I was set on a pure academic track: a direct PhD in North America, climbing step by step up the ivory tower. 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 conduct research on pretraining, while also having encountered various possibilities across academia, industry, and even early-stage startups.

This outcome was not something I planned from the start, nor can it be chalked up to simple luck. I have always believed in the saying: greatness cannot be planned. Not that I am destined to achieve something grand, but cognitive compounding inherently works this way—once the compounding truly kicks in, the final magnitude of return is simply impossible to foresee from where you stand at the beginning. You cannot reverse-engineer every step from a preconceived outcome, because at the starting point, you have no idea how far things will unfold.

My deepest takeaway from these two years is just that: you do not need to obsess over locking down every future step in advance. As long as you commit to things that are valuable over the long run, leave the rest to time. What choices you can make and where you stand today are never the result of a single burst of effort; they are the stacked outcome of past cognition, experience, and external connections compounding like a rolling snowball.

I like to use the Taylor expansion as an analogy. Judging the future at any given moment is like expanding the function of a life at that specific point. The richer your cognitive dimensions and the fuller your information, the higher the order of derivatives you effectively possess, and the closer your reconstructed curve aligns with reality. If you only hold low-order information, the best you can do is a linear extrapolation—it may look neat and straight, but as you move just a bit further down the road, it diverges drastically from the real world.

As an incoming freshman, I had only that limited prior, so viewing a direct PhD as the only path felt completely natural. That was not a lack of vision on my part; almost anyone at that starting point would have reached a similar conclusion. The essence of cognitive compounding is that your principal is never fixed. Every piece of new knowledge you acquire, every remarkable peer you meet, and every experience that resets your view does more than pay a one-off return. It directly merges into your underlying cognition, becoming the new baseline for your next judgment. With every additional dimension of useful information, the precision of your future judgment leaps to a new level.

Looking back, the starting point was a decision I made right after entering university.

In September 2024, as a freshman, my residential college encouraged student-initiated academic activities. Going with the momentum, I spearheaded the university’s first deep learning seminar. At the time, I had no expectations of where it would lead; I simply thought it was interesting and worth doing, so I went ahead with it.

Through running the seminar, I met two junior faculty members on campus. One brought me into an on-campus research internship, and the other helped open up subsequent remote research collaborations. Because the seminar drew some attention within the university, I also connected with many outstanding peers and seniors at Xi’an Jiaotong University. Among them was an especially thoughtful alumnus: after graduating from XJTU, he pursued a PhD at Peking University, then left during his studies to found an AI + healthcare startup. That story comes later.

During that period, I gradually built a habit of consistent public output. At first, I uploaded recordings of my own seminar talks to Bilibili. Later, this naturally evolved into writing technical blogs, turning algorithmic principles, frontier thoughts, and mathematical derivations that fascinated me into written posts. Many ideas did not necessarily need to be turned into formal academic papers, yet they had genuine value for sharing; a blog was the ideal medium.

At the time, this output brought no immediate payoff: it did not count toward GPA, nor did it earn scholarships, and it took up substantial extracurricular time. But over time, I realized that if you keep writing seriously and consistently, these writings become durable professional credentials and thoughtful deposits on the open web. Quite a few posts sparked discussions in technical circles and helped industry practitioners take notice of me. They formed the very first deposit into my cognitive principal, quietly putting down roots.

After that, things began to chain together, one after another.

At the end of my freshman year, in July 2025, I traveled to Hangzhou for a machine learning conference. There, I met the collaborator I had previously only worked with online, and connected with researchers from Tsinghua University, Shanghai University of Finance and Economics, and other institutions—marking my real entry into the academic community. Almost concurrently, drawn by my Bilibili recordings and the technical writing on my website, a faculty member from Renmin University’s Gaoling School of Artificial Intelligence reached out, inviting me to collaborate on top-tier conference submissions. I smoothly joined the Gaoling research team and began investigating large language model mechanisms.

By that point, I had begun to realize that many opportunities are not things you desperately chase. They are natural feedback that arrives once your earlier groundwork accumulates past a critical threshold.

What truly triggered a cognitive leap was the month I spent on site at Gaoling over the 2025 winter break.

Beijing’s most distinctive feature is its extraordinary density—density of talent, density of information, and density of practical experience. In a typical campus environment, you might wait weeks or months for a single conversation that genuinely broadens your horizon; in Beijing, however, you can interact almost daily with people from diverse backgrounds working on the industry’s front lines. It is as if the same cognitive principal experiences a significantly amplified information rate and feedback frequency, noticeably accelerating the pace of cognitive iteration.

Before heading to Beijing, I wrote a blog post examining Scaling Laws from a mathematical perspective and shared it on social media. The technical lead of ByteDance Seed Lab’s Pre-Train group happened to see it and directly reached out with an internship invitation. Because I had already committed to the on-site visit at Gaoling, we agreed to schedule the internship for the summer of 2026.

During my month at Gaoling, I met senior lab members and made friends from Fudan University and quantitative investment firms like Ubiquant. While chatting with a friend from Ubiquant, he shared a very pragmatic perspective: if upon graduation you can secure highly competitive compensation and frontier resources in industry, spending a couple of years in industry first is entirely viable, and you can always decide later whether to return to academia for a PhD.

That conversation was a major revelation. Before then, I had subconsciously treated a direct PhD as the only legitimate one-way street, assuming that climbing all the way through the ivory tower was the only way to stay on the right track. But that exchange made me realize that real paths are never singular. Today, much of the frontier compute, engineering, and large model exploration in AI is heavily concentrated on the industry front lines. Entering the field early to accumulate hands-on experience with large-scale cluster training and empirical systems before mapping out long-term research plans can actually give you a much broader perspective.

This is precisely what I mean by higher-order information reshaping one’s judgment. Add one real dimension, and the shape of the entire path transforms before your eyes.

From then on, one thing became increasingly clear: never cling rigidly to an unchanging blueprint. Instead, frequently update your cognition with new incoming information. Especially in a field moving as fast as AI, where technical paradigms and industry landscapes refresh every few months, trying to make today’s decisions using a framework from a year ago is no different from notching the boat to find the sword.

Afterward, developments accelerated even further.

In April 2026, Dong Kehan’s team reached out via social media to discuss potential early-stage collaboration. I was initially hesitant, feeling I was not fully prepared. But when they reached out again a few days later, I recognized it as a rare window to observe up close how an early-stage deep-tech team makes decisions, and decided to dive into serious discussions.

After talking in person, I decided to join. Above all, I knew that working alongside a group of explorers at the technological frontier—absorbing their technical judgments, research tastes, and betting logic firsthand—would offer a growth velocity impossible to achieve by pondering alone. When such an opportunity presents itself, there is simply no reason to let it pass.

Through this connection, I got to know young, core researchers from teams at MiniMax, DeepSeek, Moonshot AI, and Tencent Hunyuan, and had opportunities for face-to-face exchanges with seniors with backgrounds from DeepMind and Anthropic. Every deep discussion enriched my grasp of the industry and technical nuances, making my subsequent judgments much more grounded.

Also in April in Beijing, I met again with the XJTU alumnus I had met through the freshman seminar. The AI + healthcare company he founded had raised hundreds of millions of yuan, aiming to use AI to connect the entire life sciences pipeline—from basic research to protein synthesis and drug development.

We talked for over an hour. I asked him whether pursuing a PhD was still a sound choice in the current environment. My initial assumption was that finding an outstanding advisor would eliminate most friction, and that setbacks usually stemmed from poor advisor matching or wrong directions. But he asked me back: even with a great advisor, are there not other dimensions of structural friction?

That question stayed with me for a long time. I realized my previous framing had been overly simplistic; the inherent cycles and real-world constraints of the academic establishment exist objectively. Combining that insight with what I had observed in the industry, I became even more resolute about interning at ByteDance Seed Lab first to accumulate real-world industry experience. Many choices have no abstract, absolute hierarchy; the key lies in whether it fits who you are right now. And making that judgment requires having access to rich, fresh, and grounded information.

Looking back, the initial connection to this alumnus who influenced me so deeply still traced back to that seminar I started as a freshman. That is the magic of accumulation: casual decisions made in the present can reverberate in unexpected, profound ways long into the future.

In early June 2026, while finalizing summer internship details with ByteDance Seed Lab, the team mentioned a closed-door technical session taking place in Beijing and asked if I would be around. I was in Xi’an at the time, and booked round-trip flights and lodging almost without hesitation.

Hearing frontier pretraining ideas in person and discussing them directly with core team members yielded value that vastly outweighed the immediate travel costs.

That trip to Beijing exceeded all expectations. I heard firsthand unpublished research progress, had extensive discussions with Ziming Liu, and met core researchers at Seed. We then shared a meal together, and the internship was essentially confirmed. (Naturally, I still went through the full standard interview and onboarding procedures afterward.)

This is cognitive compounding at its most concrete: when you proactively step into high-density talent circles and take part in high-value technical discussions, what initially appeared to be random chance gradually converges into clear certainty.

After my joining Seed was announced, it drew attention from peers and seniors on social media, many of whom were people I had followed from afar for a long time. Even more interestingly, just days after completing onboarding, Moonshot AI’s recruiting team reached out, mentioning they had read my earlier blog post, “Re-listening to Yang Zhilin: Bet on Scaling, First Principles, and Long-Termism,” found the analysis insightful, and wanted to connect.

That post had originally been nothing more than quick reflections jotted down after listening to a podcast; when writing it, I never imagined it would reach the team in question. But that is the nature of cognitive assets: when you write them down earnestly and put them out publicly, they remain as durable signals, continually projecting your technical taste and depth of thought. They do not expire or degrade; at the right moment, they will bring unexpected, high-quality connections right to your doorstep.

Even though my internship was already set, I still attended the conversation. Engaging with top practitioners is never about rushing to grab another offer; it is about absorbing frontier industry perspectives and stress-testing your own thinking. That input will ultimately become part of your cognitive reserve for future decisions.

At this point, it should be clear that the path I walk today could never have been pre-planned as a freshman. I simply became convinced that artificial intelligence is the macro direction most worth dedicating oneself to over the next decade. Along that direction, whenever I encountered something worth doing, I did it; whenever I met people worth learning from, I connected with them; and whenever a chance to expand my cognition arose, I seized it—all while keeping my specific execution path open.

People online often ask: if you make this choice now, what will you do four or five years down the line?

Whenever I see questions like that, I feel that the technological landscape and industry shifts four or five years out are fundamentally impossible for us to predict with precision today. Using our current, limited cognition to lock down every step for our future selves is simply unrealistic. The external world evolves dynamically, and our personal cognition must recalibrate in tandem. A truly robust plan is one with an unwavering macro direction and an agile, adaptable path—always using your fullest and highest-order cognition to make the latest judgments.

Many understand compounding as passively waiting for returns to accumulate. But cognitive compounding is not passive. It demands that you remain consistently open and perceptive, actively immersing yourself in high-density environments to absorb rapidly, verify quickly, and iterate frequently. In the early stages of learning and growth, when your cognitive base is relatively small, the marginal return of every effective cognitive update is at its highest, and the cost of trying is far lower than the opportunity cost of hesitation. If you have a good idea, go execute it; do not waste time in pointless indecision.

Of course, all of this rests on one fundamental premise: what you put out publicly must match your genuine technical accumulation. Public records and peer exchanges are never about empty self-promotion or social posturing. Only when your analysis, derivations, and technical practices are grounded in solid competence and deep thinking can the connections you forge be stable and lasting. Superficial packaging cannot sustain long-term professional trust, let alone enter a healthy compounding loop.

The curve of cognitive compounding is almost always flat and protracted in the early days. You may do many things without seeing noticeable feedback for long stretches, and might even feel you are making slow progress. But as long as you keep depositing into your principal—continuously taking in, putting out, and connecting with high-caliber peers—the breakthrough past the plateau will inevitably arrive, and the accumulated momentum will suddenly manifest in full force.

Two years of college life is only the beginning. Now that I understand the rhythm of this curve, I am no longer in a rush for instant answers.

Take your time; cognition will do the accounting.

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/}
}