The Talk Begins: AI Fund's Experience
Andrew Ng draws on his experience at AI Fund to discuss how to grow startups faster. He emphasizes that AI Fund co-founds an average of one startup per month, and that he has personally experienced every aspect of building a startup -- writing code, talking to customers, designing features, setting pricing, and more.
"We're not just watching other people build startups. We're actually in the trenches with founders, building startups ourselves."
The core theme of this talk is speed. Andrew says that one of the strongest predictors of startup success is execution speed. With recent AI advances enabling startups to move much faster, he wants to share the latest best practices from his direct experience.
The AI Stack and Where the Real Opportunity Lies
Andrew describes the AI stack as follows:
- Semiconductors (hardware)
- Cloud (hyperscalers)
- AI foundation models
- Application layer
While much of the media and social media focus on hardware or foundation models, he emphasizes that the biggest opportunity lies in the application layer.
"Almost by definition, the biggest opportunity has to be in the application layer. Because applications are what actually generate revenue."
The Rise of Agentic AI and Workflow Innovation
He identifies the emergence of agentic AI as the most important recent AI trend. Where previously AI was asked to produce results all at once, now AI can perform tasks iteratively across multiple steps.
- Example: Essay writing
- Draft an outline
- Conduct web research and gather materials
- Write a first draft
- Review and revise the draft
- Repeat
"Through agentic workflows, the model can think, research, and revise multiple times, producing better results."
The emergence of this agentic orchestration layer has made application development easier, and he reiterates that the application layer will remain the most valuable area going forward.
The Importance of Specific Ideas for Fast Execution
At AI Fund, they focus exclusively on specific ideas. Vague ideas cannot be executed quickly.
- Vague example: "Let's optimize healthcare assets with AI"
- Specific example: "Let's build software that lets hospital patients book MRI appointments online"
"Specific ideas bring speed. You can quickly tell whether it's a good idea or not."
Specific ideas give teams clear direction and enable rapid validation or quick confirmation of failure. He also notes that the intuition of domain experts who have spent long periods thinking about the field greatly aids fast decision-making.
"Data is important, but in startups, data collection can be slow. Expert intuition can enable faster decisions."
Fast Feedback Loops and the Role of AI Coding Assistants
The biggest risk for startups is failing to build what customers want. The solution requires fast feedback loops.
- Basic loop: Software development -> user feedback -> improvement -> repeat
The emergence of AI coding assistants has dramatically increased engineering speed and significantly reduced costs.
"The speed of building prototypes has increased more than 10x. It's now not unusual to completely rebuild the codebase three times in a month."
Even software architecture choices are no longer "irreversible decisions" as before -- they've become "two-way doors" that can be changed whenever needed.
"In the past, once you chose a tech stack, it was hard to change. Now it's possible to throw away the codebase and rebuild in a week."
Why Everyone Should Learn to Code
Thanks to AI coding assistants, coding is getting easier and easier. Andrew argues that every role should learn to code.
"Our CFO, HR lead, recruiters, and even front desk staff all know how to code. Because of that, everyone performs at a higher level."
The ability to precisely instruct a computer on what you want will be a core competency of the future, and learning to code is the first step.
Various Methods for Getting Product Feedback Quickly
Getting product feedback quickly is becoming increasingly important. Andrew introduces various feedback methods ranked by speed and accuracy:
- Fastest: Use the product yourself and decide based on intuition
- A bit slower: Get feedback from 3 team members
- Slower: Get feedback from 3-10 strangers (at cafes, hotel lobbies, etc.)
- Even slower: Deploy a prototype to 100+ testers
- Slowest: A/B testing
"A/B testing is discussed a lot in Silicon Valley, but it's actually one of the slowest feedback methods."
When gathering feedback, he advises that you should not just look at results, but continuously update your intuition to make faster and more accurate decisions.
Why Understanding AI Technology Increases Speed
Because AI is still an immature technology, teams that understand AI well can move much faster than those that don't.
"Making the right technical judgment about AI can solve a problem in days. Making the wrong decision can waste months."
Additionally, combining the various GenAI building blocks that have emerged over the past two years (prompting, workflows, evaluations, guardrails, RAG, embeddings, fine-tuning, etc.) enables building software that was previously unimaginable.
"Every time you learn a new building block, the combinations you can create grow exponentially."
Q&A: The Human Role in the AI Era, Overhyped AI Narratives, Startup Competitiveness, and More
1. What Should Humans Prepare for in the AI Era?
"In the future, the most powerful people will be those who can precisely tell computers what they want. People who can leverage AI will be far more powerful than those who can't."
2. Overhyped AI Narratives and Risks
Andrew criticizes the exaggerated risk narratives claiming AI will drive humanity to extinction, arguing these are amplified for the benefit of certain companies.
"The idea that AI is so powerful it could wipe out humanity is nonsense. These exaggerated narratives have only served to help certain companies with funding and influence expansion."
He emphasizes that responsibility matters more than safety:
"AI is neither safe nor dangerous -- it depends on how we use it. Using it responsibly is what matters."
3. Startup Competitiveness and Moats
"The most important thing is building a product that users truly want. Everything else can be figured out afterwards."
Moats (competitive advantages) often develop naturally over time, and right now there are far more unclaimed opportunities in the application layer than people realize.
4. Cumulative Effect and Flexibility of AI Building Blocks
When combining multiple building blocks, maintaining flexibility is crucial. For example, designing software so foundation models can be easily swapped means you can quickly switch when a better model appears.
"We often don't even know which model we're using -- when a new one is better, we just swap it in."
5. The Future of AI and Education
On how AI will transform education, Andrew notes there's still a lot of experimentation, and predicts it will become hyperpersonalized.
"There's still a lot of experimentation in education, and the final form isn't clear yet. For the next decade, we'll be mapping various workflows to AI."
6. AI's Social Responsibility and Inequality
Regarding concerns that AI could worsen inequality, Andrew emphasizes the importance of maintaining ethical standards.
"If you believe what you're building doesn't make the world better, don't build it. We've cancelled economically viable projects for ethical reasons."
He adds that helping everyone access and use AI is important.
7. Democratizing AI Education and the Importance of Open Source
To prevent AI knowledge from concentrating in the hands of a few, he emphasizes the importance of open source and open-weight models.
"If regulation and closed policies succeed, only a few companies will drive innovation, and everyone else will need permission to try anything new. We must protect open source."
Conclusion: Key Takeaways for Startup Success
Andrew emphasizes that speed is critically important for startup success, highlighting:
- Focus on specific ideas
- Develop rapidly using AI coding assistants
- Improve products quickly through diverse feedback loops
- Continuously learn AI technology and keep up with the latest trends
- Use AI responsibly
"Both speed and quality matter, but speed is absolutely critical."
He closes the talk with the message that everyone should learn AI, protect open source, and continue innovating responsibly.
Thank you!