Jonathan Su on AI Research, Education and Happiness as a High-School Student

Zhang Xiaojun talks with high-school student Jonathan Su about his self-directed research and an attention paper accepted at ICML 2026. They discuss experimentation, AI use at school, and how young people experience both excitement and anxiety about technological change.

Su says he was surprised when he first saw demonstrations of natural language producing programs and prose. He had thought software could only follow rigid instructions. Curiosity led him to search for machine-learning courses online.

Andrew Ng’s courses introduced him to the field, and Andrej Karpathy’s videos on building a GPT and Transformer exposed him to implementation. Much was initially confusing, so he challenged himself to read and annotate one paper a day for 30 days.

Reading papers made him notice argument and presentation: why some papers feel clear while others are hard to follow. His primary interest is AI itself. To him, changing a model and observing what happens resembles studying a complex system.

He initially wanted to train the best language model at its size, but estimates of compute costs made direct scaling impractical. Starting from his own Transformer code, he implemented and modified ideas from papers; many experiments eventually led to a research idea.

He says he tried hundreds of modifications and eventually submitted to ICML. He found the reviewers’ comments professional and useful, and learned that research publication also involves communication and iteration.

Su recalls growing up in Shanghai before moving to Hong Kong for school. He appreciates the freedom his family gave him to explore. Mathematics and programming competitions included detours, but helped him understand what he enjoyed.

Chinese chess became another important interest. He learned from online resources and entered competitions, enjoying the thinking, patience and feedback of one-on-one play. Zhang notices that self-directed learning runs through many of his interests.

He explains his fascination with pretraining through an analogy: post-training shapes a body, while pretraining improves the material available to shape. This is his metaphor. He also acknowledges the expense and does not recommend that everyone train models without considering cost.

He thanks his family for supporting the experiments and recalls worrying before paying to scale one run. Even a successful experiment might not lead to acceptance, and compute costs could not be recovered. He decided to try and later saw results he considered meaningful.

He used public datasets and admits that his training code was still messy. In one run, an incorrect setting meant three models did not save checkpoints, wasting compute spending. He was upset; his family encouraged him to learn from the mistake.

His paper is titled “Attention Projection Mixing with Exogenous Anchors.” The [ICML proceedings](https://proceedings.mlr.press/v306/su26a.html) list Jonathan Su as author and describe ExoFormer, which learns anchor projections outside the layer stack and mixes them with attention projections.

The paper’s motivation is that reusing early-layer projections across deeper layers can create tension: the early layer must serve as a stable reference while performing its own computation. Its reported results should be understood within the tested models, data and baselines.

He recalls that moving from code and papers to repeated experiments was not a straight path. Some ideas stayed confusing until connections suddenly clicked. He enjoys trying things himself and seeing feedback.

To explain attention, he offers an intuition: a query represents what a token seeks, a key helps other positions find it, and a value carries information once found. He presents this as his own way of understanding the mechanism, not a complete mathematical definition.

At school he rarely finds peers who want to discuss the technical details, so he often works alone and asks AI tools or online resources for help. He does not see himself as someone who avoids people: he enjoys talking with friends.

At ICML, a chess social was a highlight: he could play while asking researchers from different places about their work. He describes it as one of his happiest experiences.

He also asks people whether AI could lead to human extinction. The small number of answers he mentions came from informal conversations, not an expert survey or risk estimate. He acknowledges that he cannot predict the future and is expressing a concern.

He worries that rapid progress could change work and research in ways that leave people uncertain about their purpose. He also imagines risks if powerful systems’ goals diverge from human interests. These are possibilities he is considering, not certain predictions.

He hopes AI continues to advance, but not so quickly that society cannot adapt. Part of his reason for studying AI is to participate in understanding the change, while recognizing that his own control over its trajectory is limited.

Chess and research provide feedback on different timescales. A game can end quickly, while a research idea needs experiments and waiting. He enjoys the shared process of forming an idea, testing it and learning from the result.

Looking ahead, he hopes to intern in a model research environment and is considering university. In his view, the internet and AI make knowledge easier to access, while university can still provide peers, discussion and connections.

On model and application companies, he offers a student’s observation: model developers supply underlying capabilities, while applications build around user needs. His views on competition and commercial prospects are personal judgments, not settled market conclusions.

He says most classmates he knows use AI for homework or independent learning, and he also helps his parents use it. AI is part of the everyday learning environment for his generation, but people use it differently.

His school allows devices in some learning settings, while exams must be completed independently. He says he uses AI for some homework and spends the saved time on other methods such as spaced repetition; he still does exercises he finds valuable himself.

Su thinks motivation affects what a tool yields: some students use saved time to learn more, while others may turn to other activities. He admits that short videos distract him too, making choices and time management difficult.

Zhang asks how AI may change education for young people. Su says a tool can explain material while also reducing some of the satisfaction that comes from difficult learning. He describes his own and classmates’ experiences, not a finding about all students.

He says students might once have imagined contributing through years of study in a subject. Rapid AI progress makes that path feel less certain to him, although no one can reliably predict whether or how it will change.

A classmate interested in medicine worries that the profession could look different by the time long training is complete. Su offers possibilities but acknowledges they are anxious speculation, not a reliable forecast about medicine.

He recalls feeling distressed for weeks after imagining negative AI futures. Conversation and reflection helped him recognize that the future is uncertain, and he chose to focus more on relationships, activities he enjoys and what he can do now.

He suggests that friends explore how AI can help while retaining activities they find meaningful and want to do themselves. They are free to accept or reject his advice, he says; continued conversation matters.

They discuss people turning to chatbots for emotional conversation. Su thinks these tools may appeal to those lacking companionship, while also raising safety and ethical questions. He hopes people will not overlook support from real relationships.

He describes how life feels different before and after AI became commonplace: there are more tools and more uncertainty. For him, convenience does not remove the value of spending time, playing and talking with other people.

After returning from ICML in Korea, he joined a volunteer teaching program in a rural community. He and his companions designed English and chess activities, using games and small rewards to encourage participation and enjoying the children’s responses.

He finds the daily rhythm different from AI conference discussions: preparing lessons, thinking about meals, walking and playing games. He deliberately sets AI aside for a while, enjoying a slower pace and talking with companions and children.

Asked what kind of person he hopes to become, he says he wants himself and the people he loves to be happy. Happiness is an important goal for him; his ideas about whether it is finite are personal philosophy, not established neuroscience.

He imagines a future with more widespread AI, robots and new ways of living, while acknowledging that the timeline could be entirely different. He hopes technology gives people more freedom to explore rather than producing the dystopia he fears.

In a closing rapid-fire round, he talks about favorite food, a familiar home and reading. He enjoys stories and values the depth of human conversations in which people share ideas and feelings.

Among influential research, he mentions Transformer papers and technical reports. Asked what he would bet on, he says he wants to believe people will continue finding goals worth pursuing and questions worth exploring.

Recent happy moments include both new model releases and time with fellow volunteers and children. The experiences come from different worlds, but both connect to curiosity and other people.

The conversation closes by returning to young people’s sense of purpose. Su says his current answer is to help himself and the people he cares about be happy; Zhang observes that every generation needs time to grow amid rapid technological change.