Hikaru · Personal experience
Original essay · A letter to new students
Timeline
September—November
When I first entered university, I explored and learned all kinds of things. I started learning C and Python, and tried student organizations, clubs, the robotics team, and embedded systems. I also happened to do badly on my midterms. Through this process, I began to understand the basic components of university life and met some of the people who are now walking alongside me.
December—January
I began learning about deep learning, MNIST, and YOLO, looking for all kinds of projects on Kaggle and reading papers. I could feel passion and excitement in it and started to think: this is what I want to do. I also met the senior student who guided me, met Jarvis once, and caught my first glimpse of the world of research.
February
Continuing from the exploration and paper reading of the previous semester, I carefully went through Li Mu's materials again. Then I began reading classic papers deeply, studying their authors, and made my first attempt to reproduce a time-series experiment. In this process, I began to understand how fields move and change, motivation and limitation, and the problems to solve.
March
I studied agents—ReAct, generative agents, and Voyager. I was interested in all of them and kept working on them, building small demos. Out of interest, I also studied CSAPP, computer communication, networking, and web knowledge, though not deeply.
April
I built AutoResearch and AutoCache, and studied diffusion and diffusion cache.
May
I decided to submit to a workshop, completed the workshop paper, decided to polish it for EMNLP, and submitted it to EMNLP. This was the first month in which I threw myself into a complete research cycle: LaTeX, figures, experiments, writing a paper, and telling a story. Everything was a first. It was exhausting and stressful, but I also grew a lot.
June
I studied reinforcement learning and world models, got started in the Lumina community, prepared my résumé, helped with a startup project, crammed for finals, and took my final exams.
July
I took finals, attended ICML, and worked on a rebuttal. I went to Shanghai to prepare for my first internship and attended the Evolvent AI after-party. Then it was study, study, study, study: from 10 a.m. to 11 p.m., six days a week. I also met many new friends.
August—Now
I am pushing forward an ICLR project, studying embodied AI, mathematics, CS336, architectures, code, and dataflow. Next semester, I expect to intern with a Tsinghua group while continuing an internship at a startup. I will also stay at the university to lead the next cohort of the AI Society.
A year later, I want to summarize what I have thought, seen, and felt. Over the past year, I could feel my world becoming larger and larger.
For the first eighteen years, my world was only about what score I could get on the Gaokao. In the first half of my nineteenth year, my world was still about how to get high grades, how to secure postgraduate recommendation, whether I should master a technical skill, and whether I should join a research group. I was still discussing with several senior students what I could do in my major, what competitions I should enter, and whether I should take on student work.
In the second half of my nineteenth year, I entered the honors program at the start of the semester. Under the guidance of several senior students who had a major influence on me, I joined a research group, did my own research, completed a first-author project on my own, took part in a company's startup journey, clarified my main research direction, went to Korea for a conference, attended an after-party, talked with senior students, and sought advice from senior researchers and researchers at universities and large companies. I spoke with founders, venture capitalists, and people in venture investment, learned how they think and what they care about, and learned from their insights. I also became president of the AI Society. For the first time, I caught a glimpse of how vast the world is and began thinking about where I stand within it.
Over these six months, I experienced more and more firsts: my first solo first-author project, my first time leading a team, my first internship, my first time asking for advice, my first conference, and my first paper presentation. The world I could see grew larger, and my horizons widened. I met senior students, PhDs, founders, and researchers, and asked them about research, philosophies of life, and entrepreneurship. Without realizing it, I began to hold and discipline myself to those standards.
I began thinking about what good taste is, how to make my experiments solid, how to write a paper beautifully, and what kind of research matters. Is your research merely strengthened by an LLM, or does it have real value? Are you parasite AI? Are you living inside the echo of AI hacks? What I can do is make sure that my motivation and method are my own.
I also came to recognize my level at this moment. The world suddenly opened wide, yet I am only a tiny grain in the vast sea. Although the past year is worth recognizing, compared with truly senior people, there is nothing special about me. I am only a junior and not-impressive researcher. Still, I do not want this to mean that I will be a nobody researcher. I sincerely hope to have my own place and contribution in this world: to do things I find meaningful, to do the best work in the best places, and to find places with gradients where I can train my parameters.
Through these experiences, I have also gradually changed and come to know myself better. The experiments I have done, the formulas I have studied, the people I have asked, the papers I have read, the formulas I have worked through, and the architectures I have taken apart have quietly shaped my understanding of this world and this field.
Whether agentic or embodied, my world model keeps scaling up. Develop and grow freely. Treat research as an integral rather than chasing a fixed KPI; in the end, different paths will converge and things will fall into place. Through research, understand the world better, understand models better, and understand my own thoughts better. I hope ideas can work. I hope to participate in larger and better endeavors, not waste time, love what I do, and do something in this world.
So from here on, I will continue to ask for advice, study deeply, stay curious, keep seeking knowledge, think more, ask more questions, dig deeper, do my own work, share my own views, ask more people for advice, participate in better work, go to good places, and do good things. Everything I need is already within me; I need a stable inner core.
Whether the tide rises or falls, I want to recognize the vastness of the world and also my own smallness, move forward with conviction, and complete my own recursive self-improvement.
All in all,keep passion / keep going / work hard / be self-motivated / have agency.If, months from now, I can look back and smile, then I should have become a better version of myself.
Unfortunately, you can hardly escape academic metrics
Before anything begins, it is unfortunate that I cannot truly escape coursework and the postgraduate-recommendation process either, and I know how much inflation and harm this system can create. This is structural. My only hope is that, as the survival guide says, you can exchange the least effort for the highest grades, become familiar with the rules, use the rules, take some paths beyond the ordinary, and show your real ability in actual exams and presentations. Let the narrative of meritocracy participate in your life as little as possible.
People I have asked for advice
There is my guide: building a worldview of research and business, interacting with people, cognition, and agency. There are important mentors: a deep understanding of agentic systems, good paper-writing and rebuttal skills. There are also senior students who taught me what good work is—contribution, importance, and impact—who reminded me not to become absorbed in age narratives, to put my energy into important things, to explore the question of who I am, to judge the importance of research, and to understand industry.
There is a founder, with whom I discussed socialization and development that comes naturally when conditions are ready.
There is a senior student from Kimi who shared a path of growth with me: be seen, sell yourself, share more, and have the corresponding ability; pretraining is a discipline of its own, and you should focus on problems that are truly worth studying. Friends in my own year also taught me to just do it, think outside the box, and stay hungry for knowledge. The guidance that Selen and Kabi gave me on research and technical matters also benefited me greatly.
Be motivated: your life is TTT, not always pretraining
There are many trajectories in this world, and you have to walk your own. Your path in life is not confined by a roadmap prescribed by school. There is no fixed standard template, no rule for “what you should do,” and no standard linear development: chase grades in your first year, competitions in your second, research in your third, offers in your fourth; do research as a graduate student, publish papers as a PhD student, work or start a company when you grow up, raise a child, and have that child repeat the cycle. That is not it.
University is not a fixed game to clear. Leaving objective constraints aside, if it is something you can actively reach for, there is no “I need to be ready” or “I am not good enough yet.” There is no discourse about academic culture, unfairness, or “you are impressive because you do research.” There is no rule that you need grades before you are allowed to do research, or a small group before a large one. There is no need to have a dream school but wait until you have improved yourself before daring to fight one final, decisive battle.
You want to do it and explore it, so do it directly. Let preparation and growth happen in parallel. Have the courage to take the first step, find a guide, show yourself, and let yourself be seen and linked. Work hard and keep going. Go to the best places, do what you most want to do, and find your answer. Do not wait to finish your pretraining in this fixed environment. Run your own test-time learning and recursive self-improvement. What matters is who you are, what you want to do, and how you will make it happen. No one can decide for you.
In the words of my guide
“So I am telling you: there are no rules, no fixed forms or natural laws, nothing that says you must proceed step by step and follow convention. Everything is a vast sea where fish may leap and a high sky where birds may fly. Where there is no road, I force one into being. I am telling you that a way of playing is design; it is art. Look up at the moon. Being bright is useless, yet even if it is useless, it still shines.”
Age narratives, the cost of time, and places with gradients
The transformation of university into high school, meritocracy, and social anxiety has produced a narrative of genius. We do not deny that their growth comes from diligence, high energy, and agency, which enable them to encounter and seize opportunities; fate plays a role as well. But all kinds of publicity try to maximize peer pressure, making people feel that they must achieve certain things in their first year or by age twenty. It pushes people toward the peak of Mount Stupid and tells them they alone are supreme: if you did not join a group in your first year or win a national award and land an internship in your second, you are already behind.
We can acknowledge the time advantage of starting early and admire their motivation and agency, but we should neither belittle ourselves nor become self-satisfied. To quote Professor Saining's interview: “Do not care too much about success or failure at every moment. Do not count gains and losses at every point estimate. Research is an integral; all evaluations come together, and they determine whether you are a good researcher.” What matters is what you ultimately did and what new knowledge, new contributions, and important problems you left behind.
It is difficult not to feel anxious. East Asia is a generation marked by meritocracy and comparison. As your field of view expands, it magnifies your pressure. You discover trajectories like these in the world: your peers have Seed, Galbot, national awards, papers, and meritocratic achievements. You are also on platforms that can maximize your anxiety and make you feel unimpressive.
As long as you keep expanding your horizons, you will always meet people who are better at coding, better at mathematics, stronger at research, able to start companies, able to manage people, or have greater reputations. Again, put your attention on the integral. Ask them for advice and stand alongside them. This is something you can only slowly understand and demystify. I hope we can all have a stable inner core, find our own coordinate system, throw ourselves into the place with the steepest gradient, keep growing, and slowly find our own place.
Research is not that simple
My earlier EMNLP submission received my first rejection since I began doing research. It was expected, so I felt calm and did not have any particularly strong emotional reaction—especially since the PC gave such extraordinary comments.
What I can feel is regret. Apart from the bad luck of encountering that reviewer and AC on my first writing attempt, the main problems were insufficient experience in pushing the work forward and writing in May. The experimental design, data, and writing were all too immature, and both execution and solidness were problems. Of course, it feels somewhat unfair to judge those shortcomings again three months later, but it was indeed a good lesson: I learned the price of not doing these things well.
Of course, it is hard not to keep thinking: if only I had started one day earlier, if only I had done this or that, perhaps there would have been a chance for something else. But what happened has happened. Reflect on it deeply. Research is the ability to execute; you need to handle and manage the entire process. Avoid being hacked by AI, and make the paper solid.
Likewise, good work and good contributions are worth thinking about. Try to let go of academic arrogance and algorithmic-module supremacy. In an era when papers are devalued and acceptance can feel like a lottery, as a junior, if your observation and motivation make sense, if you make the work solid, try to solve important problems, bring new knowledge, and explain it to everyone—while occasionally doing infrastructure, engineering, solidly staking out an area, and doing open source as a place to stand—you do not need to be so harsh on yourself. Whether you succeed or fail, becoming familiar with a cycle, remembering this baseline, and requiring yourself to produce better work can all make you a better researcher.
Did you really learn it?
Compared with people in the pre-AI era, our greatest advantage and disadvantage are the same: the cost of learning is extremely low. An agent, as a teacher and executor, can patiently answer according to your needs and questions. It seems that for any difficult knowledge, theory, or code, instead of sitting through forty hours of online courses and struggling through documentation, you can simply have AI explain it in smooth language. For an experiment, it can run a demo; everyone sees the loss go down and handwritten-digit recognition work, and everyone is delighted. It is a sweet prison.
It instantly makes you feel as though you are standing on the peak of Mount Stupid, supreme above everything, and creates the illusion that you have learned, understood, and mastered it.
Learning cannot lose information. This does not mean morally condemning yourself with “if I cannot implement it from scratch or do not know interview trivia, then I am not solid and do not deserve to do research”—that can also be overfit. Think more, ask more, communicate more, dive into things, and express more of your own ideas.
Take TRPO as an example. AI may tell you that it is an RL algorithm that uses KL constraints for stable training, but that its optimization objective is too expensive to compute, so PPO replaces it with clipping, and then GRPO removes the critic and performs more rollouts. That is different from fully exploring why this is so: working through sₜ and sₜ₊₁, how the policy samples and is approximated, how J and A are constrained, and then examining what properties scaling laws and stability actually reveal. Reproduce it for real, know exactly how it works and how it breaks, and develop your own observations.
You may feel that you have seen attention, consumed food already chewed by AI, and read a paper. But when you truly derive the mathematics and architecture in detail and really read it, you discover that hyperparameters, architectures, initialization, and similar details all contain subtleties and theory. Only then can your learning gradually deepen.
Beyond this, what you need to cultivate is not only research taste. You need many abilities: leading people, organizing, how you think about problems, and how you see things. All of these matter. I also need to stand up and speak more, creating more signal for who I am. Do not let yourself become someone who can be generated. Ask AI more and talk with AI more—but ask humans more and talk with humans more as well.
There are still things I have not figured out
I also have my own hobbies and a longing for things that are useless and free. I hope to open a tokusatsu shop, get along well with everyone, find aesthetics, and have balance in life. I am also fortunate to have met very good friends, both in research and outside research.
I hope that when I spend time with important people, my emotions will not be so easily affected.
I hope to have a complete inner core, build my own metrics, and do better work.
I feel that I cannot easily let go. If I cannot solve a problem, the whole day feels wrong.
I want more solid mathematical foundations and coding abilities of all kinds.
I still hope this can serve as my checkpoint. I hope to keep filling these gaps, and when I look back, smile and let them pass.