These are some questions I've asked myself and my answers to them. My answers may change as I change.
These are some questions I've asked myself and my answers to them. My answers may change as I change.
A1. Read a top-tier paper and criticize it.
Pick a paper from a top-tier venue and read it critically. Find its weaknesses, unclear points, or things you would have done differently. If the paper seems too perfect, find another one.
(updated 2026-08-14)
A2. Treat your paper like a child's drawing.
Your paper may look messy compared with a masterpiece, but treat your paper like a child's drawing. Become a parent of the child and ask yourself, "what is lovely about this drawing, and how can I help others see it?"
(updated 2026-08-14)
A3. Improve just one thing at a time.
Pick one thing you want to improve: clarity, logical flow, figures, punctuation, experimental design, framing, or anything else. Then make that your goal for the next paper. "Even if my next paper is rejected, I want the reviewers not to criticize my paper for this same thing again."
(updated 2026-08-14)
A1. learn to surf the waves.
Think of the ups and downs as surfing in the ocean. At first, you may be overwhelmed by the violent waves. It just takes time and practice until you learn to ride them. Someday, you may enjoy them like Poseidon.
(updated 2026-08-14)
A2. Fully welcome small, unexpected successes.
Accept that it is completely natural for research to be full of failures and only occasional successes. So, break success into smaller pieces and enjoy each small success as much as you can. Small successes rarely visit you on the day you expect them. They often show up on an ordinary day when you least expect them. Let's kindly welcome our unexpected guests whenever they arrive.
(updated 2026-08-14)
A3. Do a hard workout.
Negative emotions make you think negatively, and negative thinking narrows your perspective. The two can amplify each other and create a vicious cycle. To break the loop, use your body. Do a hard workout: hard enough that, for a while, you cannot think about anything except your breathing.
(updated 2026-08-14)
A1. Look far backward.
Ask yourself why you started what you are doing now. The reason probably wasn't to win a competition.
(updated 2026-08-14)
A2. Look far forward.
Imagine the person you want to be at the end of your life. That person probably isn't simply someone who won this competition you're concerning now.
(updated 2026-08-14)
A3. Be responsible for yourself at this moment.
The only person we are allowed to control is ourselves, and the only time we are allowed to control is this moment. I don't believe in God. But if there is a God, I imagine that God limited the range of what we can control in order not to burden us with responsibility for everything else. So, just be responsible for yourself, at this moment.
(updated 2026-08-14)
A1. Efficiently searching over combinations of discrete building blocks.
Nature is built from relatively simple units, such as atoms, amino acids, but achieves enormous diversity through their combinations. Many important scientific problems are therefore combinatorial. However, much of modern AI's success relies on gradient-based optimization in continuous spaces.
(updated 2026-08-14)
A2. Learning to destroy its own knowledge.
Science progresses not only by accumulating knowledge, but also by finding new evidence that challenges and sometimes destroys what we previously believed to be true. Current AI, however, learns primarily from past data and is optimized to reproduce patterns supported by that data.
(updated 2026-08-14)
A3. Integrating principles across disciplines.
Nature does not divide itself into physics, chemistry, biology, and other disciplines. These divisions are ours. Different phenomena often share underlying principles across spatial scales, temporal scales, and scientific fields. Yet today's scientific AI models are largely developed within individual disciplines.
(updated 2026-08-14)