Yuzheng Sun · 立正
MAKE
WHAT
LASTS.
I’m Yuzheng Sun. I founded Superlinear Academy, where 30,000+ people learn to build with AI. Since 2020 I’ve hosted long conversations with 150+ guests, leading researchers and founders among them, and I bring what I learn back into teaching and real work.
- Cornell UniversityPhD, Economics
- AmazonEconomist
- MetaData Scientist
- Tencent IEGVice Director, Data & AI
Led a 30-person team - StatsigPrincipal Data Scientist · early team, later acquired by OpenAI
- Superlinear AcademyFounder, full time
AmazonNetflixLinkedInWalmartPinterestDoorDashTencentColumbia Universityand more
AI Engineer ShanghaiThe first AI Engineer conference in ChinaNov 5–6, 2026
Event site →“Yuzheng shares more than tools. He teaches a philosophy of thinking—the rarest and most valuable thing in the AI era. Even as a scientist, I found his course deeply illuminating.”
Liu JiaOn teachingChair Professor, Tsinghua University“Yuzheng has built an AI education community that is carefully curated, practical, and unusually useful for people doing real work.”
Wei ManfrediOn the communitySVP, AI & Architecture · IHG“Yuzheng distills years of product growth wisdom into actionable insight—helping data scientists surface decisive signals, PMs turn numbers into strategy, and founders find a repeatable path to compounding PMF.”
Vijaye RajiOn Growth Data Analytics PlaybookCTO of Applications, OpenAI“Yuzheng is a true AI-native pioneer. With his distinctive way of thinking, he helps more people achieve superlinear growth in the AI era.”
Dai YusenOn AI practicePartner, ZhenFund
We’re building a city.
Superlinear Academy is a free AI community. Engineers, product managers, researchers, founders, and people from every field bring what they build with AI here and give each other real feedback. Why a city? (essay in Chinese)
30,000+ members learn, share, and discuss here.
- Read in-depth essays →My co-founder Yage and I write on how AI works, how to build with it, and what it changes, past the news to the mechanism.
- Share what you build →900+ member projects, with tryouts, questions, and suggestions in the comments.
- Watch open trainings →AI sessions for teams at Tencent, Xiaohongshu, Meituan, and others, open to watch (in Chinese).
- Find collaborators →Members post collaborations and roles, and companies come here to hire AI talent.
Ask Lizheng
Stuck on something? Ask here first. An AI answers mainly from what I have written and said (mostly in Chinese), and every paragraph links back to its source.
Open full page →Long conversations with 150+ guests
Since 2020. Researchers, founders, and investors on what they saw, why they were willing to bet, and how reality rewrote the answer. Mostly in Chinese, some in English.

Yuandong TianFormer Research Director, Meta FAIR
Mike VernalPartner, Conviction · Former Sequoia partner & Facebook VP
Ethan EvansFormer Amazon VP · Executive coach
Howie XuChief AI & Innovation Officer, Gen
Wei ManfrediSVP, AI & Architecture, IHG Hotels & Resorts
Yang YingNeuroscience PhD · Serial entrepreneur
What I’ve worked out,
written down.
Three books, each answering one question.

Feels Like Progress
Fake work, fake learning, and what will last when AI can blow it all up. A book of ideas.
View on Amazon →
Growth Data Analytics Playbook
A practical guide to product-market fit, growth, and experimentation. With Mengying Li and Joe Kumar.
View on Amazon →
Real Stuff 真本事
A Chinese book on capability, agency, and income: what you can still do once the credential, title, and company name are taken away.
About the book →Write the call down.
Let time check it.
Right or wrong, every call stays where I made it, with the original record attached. Where enough time has passed, I add what happened.
- 2021.04
Where the last generation of AI hit its ceiling
Systems that learn mappings from labeled data can automate narrow settings well without gaining a general understanding of the world. Useful AI applications should be chosen from the mechanism and the task, not from benchmark spectacle.
Since thenChatGPT arrived in late 2022. In February 2023 I wrote that this ceiling had been broken; see the next entry.
- 2023.02Written before GPT-4
The five most important questions about ChatGPT
ChatGPT is a natural-language computer, close to the ideal way for people to call on data and compute.
Reasoned forward from the mechanism to much cheaper inference, memory and personalization, tool-connected workflows, direct delivery of results, and systems rebuilt around ChatGPT. Written in Chinese and widely shared there, then rewritten in English on Substack and given as a tech talk at Statsig.
Since thenInference got much cheaper, and ChatGPT gained long-term memory and tool use. I also got two things wrong, and they are in the essay’s postscript: I overestimated OpenAI, and I overestimated the ceiling of model training.
- 2025.01
In 2025, the biggest opportunity in AI is agents
Agentic AI is the most important near-term shift, with a window of roughly 10 to 18 months for early movers starting in early 2025.
Since thenClaude Code, Manus, and Codex all launched that year, and agents became the industry’s main line of work.
Original record - 2025.03
MCP has structural flaws
The hype around MCP exceeds what the protocol can reliably deliver; it has major structural problems.
Since thenMCP became the industry standard; OpenAI and Google adopted it the same year. That November, Anthropic wrote up the problem itself: with many tools connected, their definitions alone fill the context window.
Original record - 2026.03
Why OpenClaw is bound to fail, and why I still recommend trying it
The product category will not last, but trying it yourself is a good way to understand this round of capability change.
Since thenToo early to call.
Original record - 2026.09How to judge hype
Jev and the AI Narrative Trap
To tell whether a new technology is hype, find the real benchmark: its best actual alternative.
A new technology can win every comparison you are shown and still be the wrong thing to learn. Five levels of judgment get you out: understand the mechanism, test its limits, choose your own criteria, find the best real alternative, and decide whether the problem deserves your attention.
- 2026.11A prediction for what comes next
The next breakthrough for AI agents is the file system
The breakthrough general agents need next is in the file system.
Local agents are tied to one computer, and cloud computers tie your files to their machines. What agents need is one set of files that belongs to you, the same copy everywhere, read and written by you and by your agents. It may not arrive next year.
I’m wrong ifGeneral agents have reached most people and their files are still tied to a single machine.
Original record
What I’m thinking about now
Questions that keep coming up in teaching, interviews, and building. Most are English versions of essays first written in Chinese.
Behind AIE Shanghai: How a Small Team Quietly Absorbed Dozens of Rounds of Changes with AI
Being AI-native doesn’t mean using AI at every step. It means finding the one constraint that binds hardest, and using AI to widen it.
The Six Levels of AI Products
Knowing how to build things with AI coding does not mean knowing how to build AI systems.
Why Everyone Is Learning the Nouns, but the Real Gap Comes from the Verbs
Nouns are parts. Verbs are the ability to assemble.
Quality and AI
AI can only handle “goodness that has already been defined,” while you can feel “goodness that hasn’t yet been defined.”
Three ways into the city.
Start with the free community. For structured learning, take AI Builders; to keep up over the long run, join Stay Superlinear.
Start here
The city above. Come in and look around, then decide how far you want to go.
AI Builders
Three self-paced courses I co-teach with Yage: get started in one evening, spend about 40 hours on how AI really works, then build a project of your own.
Stay Superlinear
Deep News on the day's AI stories, a monthly live Q&A with Yage and me, and 12+ closed-door masterclasses a year.

AI training for teamsTeam seats, private cohorts, course customization.Fully custom programs$100,000+
With Yage at Amazon: Build useful things (with AI)
What lasts
starts small.
Pick one thing you can own from judgment to consequence, and that someone else is free to reject. Make it, then show it to real people.





