Notion's Founder: AI Can Give You Code, But Shouldn't Keep Your Understanding
Vibe coding lets everyone build software, but Ivan Zhao believes what truly endures is not UI and procedural code but data and context — whoever masters these is the winner in the AI era.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Vibe coding revives the forgotten dream that software could be for everyone
Ivan explains that computer pioneers in the 1960s and 70s envisioned software as universally usable as reading and writing, requiring no special 6-year training to ‘read and write’ code. But in the late 1970s and early 1980s, history took the Jobs and Gates path instead — software froze into applications, completely separating creators from users. Notion originally wanted to give computing power to more people without code-writing, and the Cloud Code era is finally making that original vision real: now anyone can start building their own software.
— Ivan ZhaoWhat can be modified and what should be aren't the same thing
Ivan breaks down whether to let users modify their tools along several axes. For everyday-use things, users have the right to modify — this is basic freedom, and vibe coding does this better than any medium ever has. But for things like chairs, doors, and calendars that most people have no strong opinions about and simply want someone to perfect, everyone shouldn't vibe code — professionals should craft them carefully and offer them to market. For enterprise software, the security and maintenance costs mean individuals shouldn't tinker with it either. The brand-new calendar app they just released that day, he categorizes as ‘something professionals should do well’.
— Ivan ZhaoData and context, not UIs, become the lasting competitive moat
Ivan points out that AI agents fundamentally changed where software value comes from. For the past 15 years, software value lay in one team defining the best workflow and building the UI; now agents can write procedural code for any workflow and vibe code any UI, so Notion no longer ‘marries’ itself to a UI or process logic. What truly matters and lasts longer is how data is stored and how context and semantics are preserved — this is what actually matters to a company about its business. Notion positions itself as a ‘guardian of data’, opening this context to different agents in the world, while providing guardrails for team collaboration.
— Ivan ZhaoAfter 13 years serving humans, a company's hardest step is embracing AI
Ivan divides Notion's 13 years into phases. The first 5 years were exploratory wandering in the dark; during COVID they found product-market fit and started growing, with systems built for humans. After encountering GPT-4-level models in late 2022, he ‘took the AI pill’, locked himself in a hotel room with his co-founder to build the first prototype, launching it a few weeks before ChatGPT itself. But the company wasn't easily convinced — many thought this was just a fad like cryptocurrency. It took nearly 2 more years to actually ship the agent product, and only got it running smoothly by this time last year.
— Ivan ZhaoTalking to Claude feels better than reading, which should worry us
Ivan observes that he now prefers chatting with Claude to reading books, and sometimes this pattern leaves him feeling ‘hollowed out’ like after scrolling TikTok for 10 minutes. He attributes this to second- and third-order effects of technology — like smartphones; if Jobs came back and saw how teenagers use phones today, he might not be 100% proud. He thinks builders, as technologists, often just charge ahead and rarely pause to consider these downstream effects.
— Ivan ZhaoUnderstanding is the one thing you can never delegate to agents
Ivan says the real problem isn't something surface-level like adding a ‘it's too late, you should sleep’ reminder in a chat product. It's whether people still understand the system — if a team maintaining a mid-size codebase can't explain what happened a week later, that's the actual problem. His position is clear: you should never surrender understanding without a fight. Instead, you should build better abstractions, tools, and visualization methods — that's the entire point of science.
— Ivan ZhaoA tool everyone mocked became essential for understanding agent-driven work
Ivan shares a concrete example. After Cloud Code became popular, everyone worked with their own black-box agents, coordinating through GitHub merges, but the team discovered they had no idea what each other were doing. So Notion rebuilt a Notion kanban board — this tool that everyone used to complain about became a solid visualization metaphor for orchestrating multiple agents' work, letting understanding land again.
— Ivan ZhaoVibe coding needs Lego's structure, not clay's freedom, to build anything
Responding to a question about moving from block structure to unstructured tokens, Ivan contrasts Lego and clay: clay is finer-grained and freer, but kids and adults alike struggle to build tall structures with clay — they fall apart easily. Lego, with built-in structure, is actually easier to pick up and build with. He thinks vibe coding currently feels more like JavaScript's clay, but creators often lack the professional judgment to structure it into a tower. The industry is converging on ‘Lego blocks for vibe coding’ — permissions, sync, collaboration infrastructure arranged upfront, with users just expressing business logic and intent.
— Ivan ZhaoYour next billion knowledge workers won't all be human anymore
Ivan says Notion operates on an internal assumption treated as fact: the next billion knowledge workers coming online will mostly be agents, so Notion's customers are no longer just humans but increasingly agents like Claude and ChatGPT. But he also emphasizes that human collaboration and friendship matter, and they don't want to live in a world where everyone just faces a chat box and falls into TikTok/Instagram-style effects. So Notion's core positioning is ‘one canvas for human teams and agent teams to share’, not letting the product completely degrade into a headless service with no interface.
— Ivan ZhaoIn their own words · checked verbatim
Then until two years ago, right? Until Cloud Code era, everybody can start making their own software now. In some sense, the original vision of the 60s, 70s, hippie generation of computing come to life.
Ivan Zhao6:12
So in some sense, we are not married to that. Notion actually start to design, allow you to create your own UI even before AI. But what's really important is the data. How do you hold your data? How do you hold your context, your semantics?
Ivan Zhao12:31
I think in particular, like a lot of people here are probably thinking about, if you built a company that works and worked before AI, what are the ways to bring the team along and the product along as you're transitioning into this new world?
Dan Shipper16:04
I don't think we should surrender our understanding. Definitely not without putting up a fight. If anything, we should try to figure out how to create better abstractions, better tools, and better visualization to understand things.
Ivan Zhao23:58
Then you realize, holy shit, we don't understand what each other are doing. It's better to have a team-based coding thing.
Ivan Zhao25:02
There's a difference between Lego pieces versus a clay. Clay is kind of like a mud. It's very hard for a kid or even an adult to build up tall clay. Things just fall apart.
Ivan Zhao27:06
One assumption we hold true inside the company is we assume for the next billion knowledge worker coming online, vast majority will be agents.
Ivan Zhao32:24
Glossary
- vibe coding
- Writing code through natural language and AI agents without manually writing logic.
- headless product
- A product with no user interface that only provides underlying data and capabilities for agent use.
How to listen
Founders and product leads deciding which tools to vibe code themselves and which to outsource to professional products; also executives steering older companies toward AI.
The introduction from 0:00 to 1:18 can be skipped; start where Ivan discusses 1960s computing visions.