Unique Research · Unique Interview · December 11, 2024 · by Qiuping
This Unique Interview brings together two AI founders with deep overseas training and industry experience: Yu Kaicheng (于开丞), co-founder and chief scientist of KMind / Ban Ge Yu Zhou (半个宇宙), and Jiang Yaokai (蒋耀锴), CEO of Functor Tech and founder of the Zion no-code platform. They discuss how large models and generative AI are remaking enterprise ecosystems — and how they may reshape how humans think.
Guests
Yu Kaicheng — Assistant professor and distinguished researcher at Westlake University; co-founder & chief scientist, KMind. BEng, University of Hong Kong; PhD in computing and communication science, EPFL; Qualcomm Innovation Fellowship (Europe; four awardees that year); 10+ papers in leading venues. Joined Alibaba DAMO Academy autonomous-driving lab via the Alibaba Star program (2021); Hangzhou overseas high-level talent plan (2022); Westlake University assistant professor / PhD advisor (2023), founding AutoLab; co-founded KMind as chief scientist.
Jiang Yaokai — CEO, Functor Tech; founder, Zion no-code platform. Early senior engineer at Medallia (Silicon Valley, later public B2B unicorn); MS, large-scale information systems, Carnegie Mellon; BA, computer science, University of Cambridge.
Self-introductions
Q: Please introduce yourselves briefly.
Jiang Yaokai: I’m Jiang Yaokai, CEO of Functor Tech. Our product is a full-site no-code development platform for founders without technical backgrounds to build software that fits their business model. On AI, we plan in two stages: first, an Agent builder so users can embed large-model capability into software, serve customers, and monetize; second, a collaborative Copilot that helps users optimize software design and development — lifting productivity and innovation.
Yu Kaicheng: I’m Yu Kaicheng, chief scientist and co-founder of Ban Ge Yu Zhou. Our vision is to make data create more value for individuals. From day one we believed the future may not need traditional software forms — through interaction with AI, many tasks that once required business software can be done more naturally and efficiently. We recently shipped kFind, focused on localized AI exploration: an exploration engine that personally summarizes and organizes the flood of internet information into more concise content. Core goal: empower everyone through data to discover and create more value — more convenient, intelligent life and work.
kOS — a personal computer OS
Q: Ban Ge Yu Zhou also launched kOS, a personal computer operating system. How should we understand it?
Yu: As large models handle fuzzier inputs, how people interact with systems changes fundamentally. As systems do more tasks, demand for coding may fall — a kind of “compression”: less code, more efficiency. From that we proposed an AI operating system. Traditional OS input leans on UI; the AI era shifts to dialogue — what we call a Language UI internally. The idea has gradually been accepted more widely.
kOS’s core is wrapping hardware and software resources into interfaces for users and apps. Behind that, two basic capabilities: a Task Dispatcher (priority, CPU time, balancing multi-task resources) and an Evaluator (monitor and assess execution — efficiency, resource use, overall performance). Plus classic I/O: input from video, sensors, other apps, or user actions — unified interfaces so apps need not know hardware detail.
What generative AI changes
Q: What new changes is generative AI bringing to work and life?
Yu: The deepest shift is compute that can parallelize and run high concurrency — saving time exponentially. That is AI’s core advantage. As tech advances, many social activities — including entertainment — can be done through AI: deep understanding of user needs, customized content, lower cost for professional content like film, and a lift to individual creativity so everyone can have the experience they want. That single-scene case extends to office and beyond. AI will leave a deep mark on this era.
Jiang: AI’s rise reminds me of Google’s. Before Google, the marginal cost of getting information was high — go to a library, rely on a librarian; holdings were limited. Google drove search cost to fractions of a cent — thousands of times cheaper — and unlocked business models that couldn’t work when information transfer was expensive. AI likewise lowers the cost of handling fuzzy information and flexible decisions. Most of us are limited outside our specialty — a strong programmer may be useless at design. AI raises the floor so people can reach a useful level across more domains.
Humans don’t process information as rigidly as computers; what we receive and emit is often fuzzy. Before QR codes, warehousing was tedious — humans had to judge product details machines couldn’t. AI’s generality makes fuzzy tasks — including intent recognition — feasible. In education, “teach according to the student” is ancient, but private tutors aren’t affordable for everyone. AI can watch growth, spot blind spots, and personalize plans — exciting because it can raise social efficiency.
Another effect: some traditional skills matter less — like the record industry. When sound’s marginal cost fell, most pianists couldn’t make a living playing; only the top few; people bought records instead of paying for live performance. AI may do something similar: many fields need fewer people. Those who keep advancing and innovating remain necessary — as Lang Lang still records in the record era. AI may replace 80% or 90% of some work; in high-end domains that need innovation and top skill, human experts stay indispensable.
Yu: I don’t think that’s a problem if you build a self-iterating system. Like AlphaGo surpassing humans at Go — with enough time and resources, AI can match or exceed humans across fields. Over a long time scale, AI everywhere won’t even be the thing people fixate on. People may stop paying premiums only for flawless products — and pay for personalized quirks or distinctive “flaws” that mark human uniqueness. In a world where AI can play piano, people who still play may be valued for scarcity and human warmth.
Jiang: I agree. Some industries won’t vanish — acrobatics, bounded by physics and the human body. Machine capability may not matter; what’s fascinating is humans pushing their limits. I care more about that.
Yu: Paying for scarce experience, essentially.
Jiang: I talked with a friend today: sometimes I want a human butler not to save time but for the feeling of being served. I don’t care about machine service — I know it’s “beneath” me; I want an equal human to serve me.
Impact on software form
Q: How will AI change software’s form today?
Jiang: First, AI won’t replace software — it adds flexible decision-making on top of existing software. Form may change; replacement is not total. Second, on AI generating software: difficulty is universal — a point in search space. Harder problems, larger search space, longer time. That applies to humans and AI (neuron-like structures). I don’t believe everything will run on giant AI models: the world’s essence is effect. Fixed-code pipelines have far lower marginal cost than AI. Even as AI’s marginal cost falls, traditional software’s falls too — for some tasks, classic software stays more economical.
Is AI’s energy over-inflated?
Q: Is AI’s development matching your expectations? Is its “energy” too inflated — as the “Godfather of AI” criticized Sam Altman?
Yu: I think that criticism is right. I agree there are risks in some domains. I work in the field, but I believe AI has no emotion — it hasn’t reached the conditions for emergent intelligence. What I worry about is education: AI can shape children’s ideology without anyone noticing. After AI takes that role, full evaluation is rare. I’m concerned kids who spend long hours with AI may doubt humans and fully trust AI’s guidance. We may need to invest heavily in guiding children so they don’t blindly trust AI. Kids learn; they may idolize AI for its knowledge. Without guidance, that can become over-dependence or fervor — risky. I stay cautious. Still, AI is a powerful tool — an encyclopedia-like information source. Used as a carrier of information, it does great good. The key is how we manage and guide use so it helps education and daily life rather than becoming a risk source.
Jiang: A short add: society keeps gaining stronger forces; personal capability and energy density rise — from mechanical energy to driving a car with a fingertip; weapons from cold steel to firearms to nuclear. AI is the same — inevitable, with danger. Like Musk’s space work: putting things into orbit fast can build stations or enable space bombardment. We can’t abandon tech because it’s dangerous — humans will seek and master these forces. Tech isn’t the problem; people who use it are. Blocking information flow doesn’t stop pursuit. Rising destructive potential is a social fact. We choose not to use some forces — not because we can’t, but because we choose not to, guided by morals, law, and norms.
China vs. U.S. AI paths
Q: Both of you studied and worked overseas. How do you see China–U.S. AI route differences?
Jiang: Direction isn’t very different; speed in China is slower. Recently I booked Lexus service in the Bay Area by phone — an AI voice system, not digit/keyword IVR. It vaguely understood me and talked like a human. When I wanted service ASAP, it offered 9:30 next morning; I said morning didn’t work; it offered 1 p.m. the day after; I agreed; it booked. That experience of empowerment and efficiency felt clearly ahead of China. In product design we insist on gradualism — not “get fat in one bite.” We introduce AI in high-certainty scenes: write a few lines of code, even one line; in no-code, polish part of a page’s look or finish a data binding — smart-guess user needs. That value far exceeds one-shot generating a whole stack of mediocre content — which isn’t realistic yet anyway. Even a top engineer fails without clear requirements. AI and users need a clear collaboration channel. We start from small, clear tasks and explore best collaboration — that’s our core product idea.
Yu: Look from two angles: scientific research tech vs. product and application tech. Jiang’s car-service example may lag in China not for lack of ability to build it, but because some organizations won’t cede decision rights. When I worked at a Bay Area company I saw startups supporting each other’s products — a clear mutual-use culture. Domestic large firms are more cautious choosing products — culture and decision process, not pure tech limits. What I worry about more is GPU supply and performance — a real factor in tech progress.
How hard are those limits for domestic AI?
Q: Do these limitations challenge domestic AI development heavily?
Yu: Personally, the challenge shows more at product definition — clear definition should come before those challenges. Whether larger tech gaps appear later is uncertain. With huge clusters (100k or 1M cards), today’s approaches might be surpassed or obsolete — but for now it’s manageable. Even including OpenAI’s o1, we haven’t seen a dramatic breakthrough yet.
Redefining human–machine division of labor
Q: After enterprises introduce AI automation, how should they redefine human–machine collaboration and reinvent org management?
Jiang: We’ve tried a lot. For VIP customers we’re testing: when a customer sends one sentence, AI auto-searches related context and materials and shows them to help the reply. Still debugging, still needs humans — but customer-service efficiency is already up. Emphasize: simulating human process is the easiest way to show AI value. Give AI the tools it needs; design its workflow from human workflow — fastest, most direct value. A common mistake: treating AI as identical to humans. Humans aren’t just an information-processing “brain” — we have dynamically updating long-term memory. AI memory is static; it doesn’t actively learn or update. So either design a “hippocampus”-like memory update, or use it more precisely and fixedly — treat it as an intern who never learns. Powerful, but not self-learning — that use pattern can still work well.
Yu: We should actively embrace AI, even if many still resist — understandable. Tools aren’t ideal yet, especially precision. We talk about “AI precision”: I still work in academia and write tedious reports. I tested popular AI writing tools on a polished passage of mine. Some left it alone — recognizing it was already fine. Others were overconfident — swapped words, changed structure, broke logic and flow. Current precision isn’t high enough; when it rises to a point it plateaus and can’t finish the “last mile.” Sometimes I abandon those tools.
Will AI alienate people?
Q: Should we worry that people will ultimately be alienated by AI?
Yu: That’s why I worry about education. Adults can filter; children’s judgment is still forming — more vulnerable to bad information environments. Cognitive psychology: frequent interaction with the text around you shapes you. Dramatic, exaggerated headlines raise psychological thresholds and change sensitivity to information.
Jiang: I wouldn’t over-worry — everything gets verified by reality feedback: environment and external factors. Even if AI advice occasionally drifts, as long as we get reality feedback we stay tethered — like talking to diverse people, or simply standing outside and watching. Randomness itself helps maintain contact with reality. If you live only inside AI dialogue, you can detach and look “weird.” That positive feedback isn’t unlimited. As long as people live in an uncertain real world full of randomness — and touch real random events — I agree they won’t develop forever in one direction.
Industrial Revolution → internet → large models — same essence?
Q: From the Industrial Revolution to the internet to today’s large models and generative AI — is the essence of these three stages the same?
Jiang: AI isn’t fully replacing human work — it substitutes some tasks. In the internet era, logistics still existed, but information made delivery more efficient and routing more precise — more effort on planning, physical logistics still necessary. Likewise AI assists rather than fully replaces. As information gets easier, AI matters more in decisions and lowers decision cost — subtle, but people spend more time and energy on decisions. New tech often brings overinvestment. Capital floods AI; returns may not be immediate; patience isn’t infinite. Long-term, the internet did change the world. After 2000’s bust, companies died and stocks crashed — but overinvestment became infrastructure, like fiber. AI may rhyme: many AI companies may fail; giant training may not sustain. Yet overinvestment leaves technical inheritance as future foundation. The shift won’t be as fast as many imagine — AI won’t remake the world in a few years.
Yu: One line: keep building something useful.
Beyond chat — how founders should invent
Q: ChatGPT made many equate AI with dialogue. How should founders explore more innovative, lasting AI applications — and better influence how people think?
Jiang: Don’t treat AI as the core or only focus. Start from solving a real user problem; introduce AI because it helps. AI is a tool — like a database isn’t a product’s selling point, AI shouldn’t be either. Design the solution from the problem’s nature; don’t change the product’s essence just because you added AI. Pick the right tools and interaction for the need — don’t force an “AI experience.” Over-emphasizing AI can put the cart before the horse.
Author: Qiuping · Unique Research, December 2024
— END —