They Haven't Even Graduated, and They're Already Building Worlds with AI | 5Y Capital's The Pub Vol. 30 x AI Native Creators
What Is the AI Native Generation Building?
For this episode of 5Y Tavern, we invited three remarkably young AI-native creators — Jinhao Tu, who just finished high school and took first place in the AI track of Alibaba's Global Mathematics Competition preliminaries; Andy Liu, a senior at MIT about to join a top AI startup after graduation; and Chunyu Chen, who went from Tsinghua to founding a company in the Bay Area.
They discussed: How does AI shape their learning and creation? How has it become part of how they explore the world and build tools? Which products have caught their attention, and what problems deserve to be rebuilt with AI? What kind of people actually drive innovation, and how do you become one? Amid relaxed conversation, there's a generation of AI natives making clear-eyed judgments, practical methods for action, and — just maybe — a little ambition to change the world.
[Guests]
Yaopeng Xing, 5Y Capital (Host)
Chunyu Chen, Founder in the Bay Area
Andy Liu, Senior at MIT
Jinhao Tu, Incoming college freshman
Selected excerpts from the podcast:
First Contact with AI
Yaopeng Xing: Welcome to 5Y Tavern. We've invited three remarkable young people — some pioneering entrepreneurs in AI, others innovative technologists making interesting things happen in the AI era. Why don't we start with what you've been up to these past few years, and how AI has changed things for you.
Jinhao Tu: I first encountered AI in my freshman year of high school, right when GPT had just come out. I'd just joined the international program, so I started using GPT to help with studying — vocabulary, memory training, things like that. Later, when I was prepping for the TOEFL, I didn't want to keep bothering my teachers, so I tried writing prompts to get AI to do things like essay grading. It worked pretty well. I shared it with classmates, it spread gradually in a small circle, and I kept iterating from there.
In early 2024, I signed up for an AI competition held by Alibaba and got decent results. During the first semester of senior year, I was working on college applications. Sometimes I didn't fully grasp what was covered in class, but teachers would share notes, so I'd just feed those directly to AI. Occasionally it would go off-topic or beyond the scope, so I'd rewrite the prompts to make it list the knowledge points from the notes first before answering. The model itself hadn't really improved, but the experience got noticeably better. That's roughly what my exploration around AI applications has looked like these past few years.
Yaopeng Xing: Those following Richard (Jinhao Tu) may have noticed his Thinking Cloud project on GitHub, which gained quite a bit of attention. And during his junior and senior years, he was already doing prompt engineering and data prompt engineering work at large model companies. Chunyu, could you share your own journey — how you engaged with technology, found inspiration in it, and eventually decided to build a company based on new tech?
Chunyu Chen: My first encounter with large models actually dates back to 2020, when GPT-3 had just launched. I was taking a history of science course at Tsinghua University, and the professor asked us to explore the differences in tech development between Silicon Valley and Route 128. Right around then, GPT-3 had just gone live, so he assigned us to investigate it — ideally by trying it out firsthand. That was my first time on OpenAI's website. It still required an invite back then, and I ended up accessing the model through a terminal-based game — a text adventure based on D&D. D&D games already had tons of user-created workflows and prompts. To finish my history of science assignment, I gritted my teeth and "battled wits" with it. Before I knew it, my writing had veered off into a fantasy adventure about elves and dwarves. The whole process was incredibly fun.
At the time, I couldn't find any usable frontend at all — I was operating entirely through a terminal app. I didn't really know how to code then, so the whole experience was both unfamiliar and fresh.
I formally started my company in 2021, spent some time on games, then tried other projects, continuing through November 2022. By then I'd accumulated some users and raised funding.
Coincidentally, one day I was scrolling through Moments and saw a Tsinghua senior post a link: "chat.openai.com," with a note saying their team had recently finished a small project and welcome to try it out. He was one of ChatGPT's earliest developers. I clicked through and sensed this thing was no joke. Sure enough, by the Lunar New Year, the world had been completely transformed by it. That experience stands out as one of the most fascinating and pivotal moments of the past few years.
Yaopeng Xing: Andy, how have you experienced new large models and new technologies in recent years?
Andy Liu: I got into AI during my sophomore year of high school. I was preparing for informatics competitions, and our coach offered an elective called "Introduction to Artificial Intelligence." This was pre-GPT; we mainly studied deep neural networks, and I was immediately hooked. I was also applying to US schools around then. My grade director asked how I planned to write my personal statement. I told him I thought AI was incredibly powerful and would become huge. We ended up discussing for hours why AI could be so capable — I explained gradient descent, backpropagation, drew analogies to human learning. The funny thing was, a girl I liked had come to talk to him about her own statement, but seeing me deep in conversation about AI, she quietly sat down and listened. I talked for five hours until it got dark. Finally Director Li said: "Go write it." I was genuinely obsessed with neural networks back then.
Later I got into MIT through my competition results. At freshman orientation, I told my group my goal was "to work for the welfare of robots" — I didn't see any fundamental difference between them and humans. By the end of my sophomore year, I'd just finished a natural language processing course. On the last day, the professor mentioned something new called Transformer that could apparently solve math problems. I got excited and wondered if I could build something with it.
MIT sends out tons of event emails, so I wanted to build a model that could parse them, automatically extract times and locations, and turn them into a clean calendar view. I recruited a classmate for the project. There was a competing team using regular expressions. I insisted large models were more powerful. Though they started earlier with more people, our two-person team finished first and launched ahead of them, getting 1,000 users to their 100 — their rule-based system basically couldn't run. Since then I've been a true believer in large models. I basically don't hand-write code for projects anymore.
Yaopeng Xing: From where we stand now, technology seems to evolve by the day. What's interesting is that you three are at different life stages. We're particularly curious: in your view, in the worlds you've observed, what has been the biggest change brought by technology in recent years?
Chunyu Chen: From what I've observed, the biggest shift in recent years is that AI can now complete the vast majority of work tasks that don't involve physical manipulation. People may still worry about incomplete toolchains or subpar user experiences, thinking truly human-replacing AI has to wait for AGI. But current models can already handle most white-collar jobs, and many other roles besides.
This doesn't mean AI is that powerful — rather, it reveals that much of our daily work simply isn't that difficult. Many positions amount to executing predetermined rules, repeating established workflows, all of which AI can now do, especially with multimodal capabilities. Of course, society still needs plenty of jobs, and companies are still hiring. But in the long run, there aren't actually that many truly "necessary" jobs. This may trigger a new transformation in production relations.
From a productivity standpoint, at least, AI already has the capacity to solve our basic needs — food, clothing, shelter. The more interesting question becomes: when these are no longer problems, what can humans still do? We'll enter a fundamentally different social stage.
Andy Liu: From my perspective, AI hasn't actually changed the world that much yet. Sure, the tech world is using AI to transform many things, but when I visit my parents and grandparents back home, I find that for most ordinary people, the changes brought by AI aren't really obvious. What people may stress about is just that AI completes tasks faster than humans.
For those of us who use AI, the change is immediate: one person can do much more. Take me — before, I might have only been a programmer. Now I can also write marketing copy, conduct interviews, organize and polish text. Things that used to take five people, one person plus AI can now handle, because AI is a strong generalist.
Jinhao Tu: I largely agree with Andy that AI's impact may still be limited for now. My observation, though, is that many people around me have a rather shallow understanding of AI models — for instance, thinking the 3.5 version can't code well. This is probably because they don't regularly engage with AI and aren't up to date on the latest developments, so they don't trust it or delegate tasks to it. It's really an information gap, depending on how frequently and deeply you interact with AI.
Change, Inspiration
Yaopeng Xing: We recently heard an interesting notion — that in the future, perhaps 1% of people might consume 99% of tokens, pushing scientific boundaries and the limits of intelligence. Meanwhile, the other 99% might live in "technology reserves," not working, no longer producing. This reflects intensifying resource inequality. What do you think of such a future? What kind of world would you want to live in?
Chunyu Chen: I think this structure has always existed historically. Before, a minority controlled most resources, capital, or labor; now it's just tokens instead — not much difference. The real change is: the efficiency with which a minority can mobilize resources has become much higher. Previously, organizing 2,000 people to collaborate would involve massive communication overhead. But now with AI, like Andy's MIT campus events calendar example, one person can send 2,000 prompts, each task executed to standard without disagreement. This level of efficiency was impossible before. So the issue isn't whether "99% of tokens" is too much, but that these tokens are becoming increasingly useful for that minority.
Andy Liu: It's true, as Chunyu says, that AI agents do save on communication costs compared to humans. But from a human happiness perspective, this isn't necessarily good. As Sapiens argues, the Agricultural Revolution boosted productivity without bringing greater happiness. AI is the same. On one hand inequality intensifies; on the other, the pace of change makes it hard to plan for the future. I think these changes are actually quite dangerous for humanity.
Yaopeng Xing: In your growth, have you encountered anyone particularly different who gave you important motivation or inspiration?
Andy Liu: I had a competition teammate who was really something. His creed was "continuous optimization" — every day, seize one thing and polish it to the extreme, then move to the next. In high school, he was obsessed with Hearthstone, skipped classes to optimize his deck. Later in college, he optimized driving to the extreme too, every parking maneuver had to be perfect. At first I thought this was pointless — what's the benefit of optimizing games and driving? But I eventually realized that the act of continuous optimization itself trains your thinking efficiency. He spent six months optimizing Hearthstone, three months optimizing driving, then moved on to optimizing research and other more useful things. He even developed his own methodology for how to advance research, prepare background material, break down tasks. He's the classic example of someone on "exponential growth."
Jinhao Tu: I wasn't really inspired by a specific person; it was more conceptual influence. Starting freshman year, I listened to lots of podcasts and interviews, especially internal company talks — interviews or CEO speeches. One interview last year left a deep impression. The CEO said that AI isn't programmed, but trained, grown. That perspective really struck me and shaped my understanding of AI's essence and my imagination of its future.
Yaopeng Xing: Chunyu, has anyone been particularly inspiring in your entrepreneurial journey?
Chunyu Chen: For me, it's hard to name one single "most" influential person, because different people inspired me at different stages and in different directions. For instance, I seriously studied Allen Zhang's product thinking and learned a great deal about product from him. He emphasizes that human needs are actually simple, but making good products is extremely hard. His line "everything I say is wrong" stuck with me deeply. On our current product team, we have people with psychology and anthropology backgrounds who often discuss needs from very abstract angles, spending lots of time on a fundamental question: what do users actually care about?
The founders of Cursor also influenced me. They made me realize that building something impressive isn't as hard as we imagine. At first I was hung up on whether I should build some "practice" products first and gradually transition. But I came to understand that if you know what you want to build, you can just go do it, even if you hit walls.
Then there was my early angel investor who told me directly: your vision is too small, stop building little tools, go do something big. At the time I was working on a group chat tool, and his words woke me up. He made me realize: the difficulty of doing small things versus big things is about the same, but big things are more worth doing. I was barely in my twenties then. I thought about how I could build a small B2B SaaS, make money, live comfortably. But if I wanted to truly build something breakthrough without constraints, I might need to try now. So for me, people I've met along the way have constantly refreshed my definition of "what's worth doing."
Yaopeng Xing: I was just going to ask Andy — you're at MIT now, surrounded by tech people and young founding teams. What have you learned from interacting with them?
Andy Liu: I think their defining characteristic is — they're extremely clear about their goals. Many people do too many scattered things and end up doing none of them well. The impression they give me is that every day they grind on one core direction, not like typical college students just "living life," but actively pushing forward what they truly want to do. That focus is quite inspiring to me.
Chunyu Chen: I really resonate with this. Focus doesn't mean stubbornly sticking to one path. The key isn't never making mistakes, but not juggling too many balls at once — know what you most want to do right now, temporarily set other things aside, and your force will concentrate.
Andy Liu: Yes, there's an essay about the difference between "persistence" and "stubbornness." Persistent people lock onto a big goal but remain flexible on the path; stubborn people won't even change their methods. For example, if you want to build a great tool product but realize the direction is wrong, you should pivot. But the goal doesn't waver — that's persistence.
"I Ask ChatGPT About Everything"
Yaopeng Xing: Returning to the technology you're currently exploring. Could you share what AI products or models you've been following most closely and find most interesting? And your own experience using them?
Andy Liu: The two products I care about most right now are Cursor and ChatGPT. Cursor because I write code daily, and I barely write it manually anymore. At MIT we even have a weekly Sunday event called "Sunday" where everyone has to use AI to write code — no manual typing allowed. The other is ChatGPT, mainly because I've been working on an AI therapy project and researching users' emotional pain points. I found that people who chat well with ChatGPT don't have obvious pain points — they genuinely get emotional relief.
I ask ChatGPT about everything now. I have a habit of recording everything in my calendar. Who I had lunch with, what we talked about, next steps — I ask it about everything. Like, I danced with a girl, do you think she likes me? I ask ChatGPT. I'm in a bad mood today, why, what do you think? It turns out ChatGPT is really powerful as a psychological coach. What used to take me two hours to process emotionally, it resolves in thirty minutes.
Jinhao Tu: What's impressed me most recently is still OpenAI's GPT-4o (o3). Whether in model capability or product utility, it's the most stunning to me. One key ability: actively calling tools during its thinking process. Early models had function calling, but often could only "call once and be done," struggling to continue thinking and adjusting after using a tool. But o3 can not only search, cross-check, crop images, but make continuous decisions and loop through a complete chain on its own, like a true agent. For example, someone previously asked it to identify where a photo was taken, and it kept zooming in on details, searching and comparing — this capability, without complex prompt engineering, is almost unprecedented in a model. Previously this kind of effect usually required manually built workflows; now it can do it autonomously.
Internally at OpenAI they seem to call it the "deep research" model, because it integrates memory, image creation, web search, Python, and a whole suite of tools. While it can't do true deep research yet, it can quickly produce an accurate, focused response in five minutes, which is extremely practical for me. In terms of utility right now, I haven't seen any product surpass ChatGPT — especially strong agent models like GPT-4o and o4 mini.
Chunyu Chen: I'm actually quite平常心 about various products, but I can share some recent experiences I've found worthwhile. For example, Menus. In the first month alone we assigned it over twenty tasks, and it really felt like having an intern. Though it had "intern problems" — occasionally misunderstanding requirements, producing bugs, hallucinating — overall it helped a lot.
I think a general-purpose agent is meaningful. Before, if I wanted to make a survey, I'd have to write the content myself, open Google Forms step by step, fill it out, then send it. With ChatGPT, I could generate bilingual versions first, but still had to manually copy-paste at the end. Menus is different — I just tell it: help me generate a survey, build the frontend and backend, and make a data dashboard. It does it all in one go. Not perfect, but the key for me is saving time.
I'm also using Notion AI now, mainly their newly launched email feature. I find the experience noticeably better than Superhuman. Superhuman is quite popular in the Bay Area, but its shortcuts are weird, the interface design too personalized, and much of the basic UI/UX doesn't even pass muster. Notion AI is more practical — email categorization, document processing, these AI features are quite polished, and even gave me some product inspiration. What I'm particularly interested in next is Figma AI. It's deeply integrated with design tools, and what you design is essentially the actual frontend. I think this is a exemplary case of deep fusion between AI products and native tools, very worth watching.
If You Had Unlimited Resources,
What Problem Would You Tackle?
Yaopeng Xing: If you had unlimited resources to build an AI product, what problem in the world would you most want to solve?
Andy Liu: What I'd most want to build is an AI personal coach that has full context of my life, like a "second me" that helps me make decisions. I habitually record everything that happens each day — what happened each hour, who I had lunch with, what we talked about. When making decisions, I feed all relevant documents about a person or situation to AI for reference. But this is time-consuming, and many details don't get recorded. If there were an AI that could integrate all my life data, combined with all internet information and massive compute, it might understand me better than I understand myself. That would be an incredibly useful product.
Chunyu Chen: I'd probably answer in two parts. On the application level, even with unlimited resources, I'd still be doing what I'm doing now, at the same pace. Because money alone doesn't solve problems — maintaining rhythm, finding the right people, staying focused on solving problems matters more. What we're trying to build now is a new kind of communication tool: on one hand helping users better organize incoming information without getting interrupted; on the other, making you more efficient when sending information — like using agents to automatically schedule meetings, pick locations, with features still iterating.
But looking further out, what I'd want to research more is interpretability in complex systems. Nervous systems, cellular systems, even large language models — they're all complex, but I believe there may be some common underlying structure. I'd want to fund the Santa Fe Institute, or just do the research myself, to explore the fundamental logic of AI, life, and information.
Jinhao Tu: I'm more focused on everyday usability. Like the personalized AI mentioned earlier, I'm quite optimistic about it. Though its impact isn't huge yet, the design logic is right, and it can support many sudden small ideas to quickly materialize.
I've also been trying ChatGPT's memory function recently. Overall it's good, but there are real issues. For example, when it remembers conversation details, it sometimes confuses information from different contexts, even "hallucinating" things I never said. I tried exporting what it remembered to check, and found many places didn't match what I actually said. But directionally, I still endorse this product form — just room for improvement in implementation.
Yaopeng Xing: You all just mentioned several important traits in model evolution: unlimited memory, smarter human-computer interaction, and interpretability of complex systems. So what do you think of AGI? If AGI truly emerges, what do you think it fundamentally needs to break through?
Chunyu Chen: I discussed this with a junior once, and he used a very illuminating coordinate system to explain model capabilities: one axis is context length, the other is task difficulty. Current large models can stretch context very long, almost inputting all the world's information, but task difficulty remains limited. The problems we can express in natural language actually have a complexity ceiling — some problems, like the Riemann hypothesis or the Navier-Stokes equations, are difficult even to "clearly formulate."
So I think true AGI isn't about making existing linguistic logic stronger, but about constructing entirely new concepts and tools to solve problems humans can't even yet express. Only at that stage would it surpass the limits of human intelligence. Much current AGI discourse is more conceptual hype, still far from true "civilizational transformation."
Yaopeng Xing: That reminds me of something quite mysterious Sam Altman said: the significance of next-generation technology may not lie in the Industrial Revolution, but more in the Renaissance. Andy, what do you think?
Andy Liu: I also lean in this direction. If AGI is just a tool, then we're merely using new tools to satisfy old desires. But if it can genuinely influence how we think, bring cultural-level change, then that's really something to look forward to.
Jinhao Tu: I largely share this view. Ultimately it depends on whether it can truly impact our daily lives. No matter how far AGI develops, what matters is whether it can help us live better, more meaningful lives.
Questions Matter More Than Answers
Yaopeng Xing: I've raised many questions today; I believe this has also been a good opportunity for mutual discovery, and I hope you can connect more deeply going forward. Now let's move to an open round — can each of you pose one or two questions?
Chunyu Chen: I want to ask two questions. First, do you think AI, at a cognitive level, could help you better understand yourself? Second, at an action level, could it help you become better, closer to who you want to be, through small nudges — like recognizing a photo you posted, analyzing a behavior pattern?
Andy Liu: I've been trying something with AI recently. I have a calendar, and I rate each time block's mood from 1 to 5. Then I use the data to find patterns — which moments scored high, and why? I found this is a self-understanding tool that can be automated with AI, telling me what I actually like and dislike.
Another headache is converting cognition into action. I made a primitive attempt: I have a document called Andy's Life Principles, listing various insights that can guide my actions. Each morning when I review my schedule, I pick a few to remind myself to execute. But I wonder, if AI could continuously read my behavior, understand my principles, like a real-time earpiece assistant telling me "what to do now," that would make things much easier.
Jinhao Tu: I quite agree with Andy. For me, I also hope AI can motivate me to do things I know I should do but don't immediately act on. But current models can usually only give feedback in a given moment; they lack strong situational awareness and sustained action guidance. So I think there's still considerable room here.
Andy Liu: I want to ask: what abilities do you think people should cultivate in this era?
Chunyu Chen: I've discussed this with some clients recently. Many people's capabilities already lag behind models like GPT-4o. So what kind of people do organizations actually need? I think what matters most isn't professional expertise, but initiative — knowing what you want to do, being willing to do it. AI is just a tool; what matters is whether you have drive, whether you're someone worth collaborating with.
Going forward, people will frequently switch between different roles, so what matters more than skills is your character: whether you're motivated, responsible, whether you can skillfully use tools. At a deeper level: do you know what kind of person you want to become? The barrier to technology is lowering, but if you can figure this out, nothing can truly stop you.
Andy Liu: I think he put it really well — you need to know what you want. Now with AI, often you just need to know what you're dissatisfied with, ask it how to improve, and it tells you the method; you just follow through.
So it's no longer about execution capability, but about whether you can clearly express: where am I, where do I want to go. I recall a line from the Bible: "Ask, and you'll get an answer. Knock, and the door will open." With AI, it increasingly feels like this. If you know how to ask a good question, or just ask many good questions, you more easily reach your goal.
Jinhao Tu: I especially agree. For me, being good at asking questions is indeed important — whether asking models, people around me, or myself. It's genuinely an important trait.
Yaopeng Xing: If you could say one thing to your future self ten years from now, what would you say?
Chunyu Chen: If I said one thing to my future self ten years from now, I'd probably lean toward "I hope you've already started acting on your greatest curiosity."
Jinhao Tu: Though it might sound cliché, I think it's still "follow your heart." I've heard it many times, and intuitively I still trust my first judgment, so I hope I can keep following my intuition and my own judgment in the future. Because a ten-year span is really long; it's hard to pinpoint specifically.
Andy Liu: I'd say "I hope you enjoy the journey." Because with AI, life may increasingly become process-oriented rather than outcome-oriented. Because if AGI arrives, everyone's outcomes will be pretty similar, so you might as well enjoy the process.
5Y Capital seeks out, supports, and inspires lone entrepreneurs, providing everything from spiritual to operational support. We believe that if the world begins to believe in the crazy ones, things will get interesting.
BEIJING · SHANGHAI · SHENZHEN · HONG KONG
