YC AI Startup Camp Day 2: Nadella, Andrew Ng, and Cursor CEO are all here.

The best use of AI is to improve iteration speed, rather than pursuing the "magic" of one-click generation.

Organizer: Founder Park

On the second day of YC AI Startup School, seven heavyweight guests were welcomed: Satya Nadella (CEO of Microsoft), Andrew Ng (founder of Deep Learning.AI), Chelsea Finn (co-founder of Physical Intelligence), Michael Truell (CEO & co-founder of Cursor), Dylan Field (CEO & co-founder of Figma), Andrej Karpathy (former AI director at Tesla), and Sriram Krishnan (senior policy advisor for AI at the White House).

Around topics such as AI technology and entrepreneurship, these big names shared many wonderful insights in their speeches, such as:

• Do not anthropomorphize AI. AI is not human; it is a tool. The next frontier is to give it memory, tools, and the ability to take action, but this is fundamentally different from human reasoning.
• In the future, intelligent agents will become the next generation of computers. This future depends not only on the precision of technology but also on user trust and seamless interaction experiences.
• Products that incorporate feedback loops, such as Agentic AI, greatly outperform those tools that can only complete tasks "once." Continuous interaction can optimize results, while iteration can lead to compounded performance improvements.
• The speed of building prototypes is now 10 times faster, and the efficiency of developing production-level software has increased by 30-50%. This advantage should be leveraged to reduce market risk through real-time user feedback.
• Code is no longer the scarce core asset it once was. With rapid prototyping tools and AI, code can be produced easily. What truly matters is the value that the code brings.
• Real-world data is irreplaceable. While synthetic and simulated data can be helpful, real data remains crucial, especially for complex visual and physical tasks.
• The best use of AI is to improve iteration speed, rather than pursuing the "magic" of one-click generation. Designers and product managers must now contribute to AI assessments.

In addition to Andrej Karpathy (for more details on Andrej Karpathy's presentation, see our article yesterday "Day One of the YC AI Startup Camp, Andrej Karpathy's Speech Went Viral") and Sriram Krishnan, we have organized the core points of the presentations from the other five guests.

Microsoft CEO: Satya Nadella

1. The compound effect of the platform: AI does not appear out of nowhere, but is built on decades of cloud infrastructure that has evolved to support large-scale model training. Each generation of the platform lays the groundwork for the emergence of the next generation.

2. The model is the infrastructure, and the product is the ecosystem: the underlying model is a type of infrastructure, similar to a new type of SQL database. The real product is not the model itself, but the entire ecosystem built around it: feedback loops, tool integration, and user interaction.

3. Economic impact is the benchmark: Satya's North Star metric for measuring the value of AI is: "Is it creating economic surplus?" If a technology cannot drive GDP growth, then it is not transformative.

4. The Boundary Between Computing Power and Intelligence: The level of intelligence will grow logarithmically with the investment in computing power. However, future breakthroughs will not come solely from scale, but from paradigm shifts, akin to the arrival of the next "moment of scale law."

5. Energy and Social Consensus: The large-scale development of AI will require more energy consumption and will also need to gain societal permission. To gain this permission, we must demonstrate that the real and positive social benefits brought by AI are sufficient to match its costs.

6. The real bottleneck of AI is change management: the obstacles to development in traditional industries are not technological, but are constrained by inherent workflows. True transformation requires rethinking how work is done, rather than just simply introducing AI.

7. Integration of Work Roles: Traditional roles such as design, front-end, and product are gradually merging on platforms like LinkedIn, giving rise to "full-stack" talent. AI is enabling more people to acquire interdisciplinary skills, thereby accelerating this trend.

8. Don't underestimate the value of repetitive work: In knowledge work, there is a significant amount of repetitive physical labor. The best application of AI is to eliminate this "invisible friction cost" and liberate human creativity.

9. Stay open to the future: Even Satya himself did not foresee the rapid advancements in "computing during testing" and "reinforcement learning" technologies. Do not assume that we have seen the final form of AI; it is likely that there will be more breakthroughs in the future.

10. Do not anthropomorphize AI: AI is not human. It is a tool. The next frontier is to give it memory, tools, and the ability to take action, but this is fundamentally different from human reasoning.

11. The Future of Development: AI will not replace developers, but will become their strong assistant. VSCode is a canvas for collaboration with AI. The core of software engineering will shift from writing code to system design and quality assurance.

12. Responsibility and trust are indispensable: The emergence of AI does not absolve humans of responsibility. Companies must still be legally responsible for the actions of their products. That is why privacy, security, and sovereignty must remain at the core.

13. Trust stems from practical value: Trust comes from practicality, not sweet talk. Satya pointed out that the chatbot deployed for Indian farmers is an example, emphasizing that visible assistance is the cornerstone of building trust.

14. From Voice to Intelligent Agents: Microsoft's AI journey began with voice technology in 1995. Today, its strategic focus has shifted to fully functional "intelligent agents" that integrate voice, vision, and ubiquitous environmental computing devices.

15. Intelligent agents are the computers of the future: Satya's long-term vision is that "intelligent agents will become the next generation of computers." This future depends not only on the precision of technology but also on user trust and seamless interaction experiences.

16. Insights on Leadership: His advice is to start from the most grassroots positions, but to hold the most ambitious aspirations. Learn how to build a team, not just develop a product.

17. The person Satya is looking for: He values people who can simplify complexity and bring clarity; inspire team vitality and unite people; and are willing to solve complex problems under challenging constraints.

18. Favorite interview question: "Tell me about a problem you didn’t know how to solve and how you solved it." He hopes to see the candidate's curiosity, adaptability, and perseverance from this.

19. The Potential of Quantum Computing: The next disruptive technology may come from the quantum realm. Microsoft is focusing on the development of "error-correcting qubits," a technology that may enable us to simulate the natural world with unparalleled precision.

20. Advice for young people: Don't wait for others' permission. Go build tools that truly empower people. He often reflects, "What can we create to help others create?"

21. Favorite products: VSCode and Excel - because they empower people with superpowers.

Founder of Deep Learning.AI: Andrew Ng

1. Execution speed determines success or failure: The best indicator of whether a startup can succeed is the speed of building, testing, and iterating. Speed brings a compounding effect of learning, and AI makes this effect grow exponentially.

2. Most opportunities are at the application layer: Currently, the greatest gains do not come from building new models, but from applying existing models to valuable, user-facing scenarios. This is where founders should focus.

3. Agentic AI is superior to "one-off" tools: products that include feedback loops, such as Agentic AI, perform far better than those that can only complete tasks "one-off." Continuous interaction can optimize results, while iteration can lead to compounded performance improvements.

4. "Orchestration Layer" is on the rise: An emerging intermediary layer is forming between foundational models and applications: agent-based orchestration. This layer can support complex multi-step tasks across tools and data sources.

5. The more specific the idea, the faster the execution: The best way to take quick action is to start with a specific idea, one that is detailed enough for engineers to begin building immediately. A good specific idea often comes from domain experts with an almost intuitive clarity.

6. Beware of the trap of "grand narratives": Abstract goals like "AI-enabled healthcare" sound ambitious but often lead to slow execution. What truly brings efficiency are specific tools like "MRI appointment automation."

7. Be brave to adjust your direction, provided that you take the right first step: If early data shows that your idea is not feasible, a concrete initial plan will make it easier for you to pivot. Clearly understanding what you are testing will allow you to quickly pivot to another direction after a failure.

8. Utilize feedback loops to mitigate risks: The speed of prototyping has now increased by 10 times, and the efficiency of developing production-level software has improved by 30-50%. This advantage should be leveraged to reduce market risk through real-time user feedback.

9. Try more, instead of pursuing perfection: Don't try to perfect your first version. Build 20 rough prototypes and see which one sticks. The speed of learning is more important than polishing.

10. Act quickly and take responsibility: Andrew Ng reinterpreted the classic Silicon Valley motto: do not "move fast and break things," but rather "act quickly and take responsibility." A sense of responsibility is the cornerstone of building trust.

11. Code is losing its scarcity value: Code is no longer a core asset with scarcity like it was in the past. With rapid prototyping tools and AI, code can be easily produced. What truly matters is the value that the code delivers.

12. The technical architecture is reversible: In the past, choosing an architecture was a one-way decision. Now it is a two-way door, and the cost of changing architectures has significantly decreased. This flexibility encourages bolder attempts and faster experimentation.

13. Everyone should learn programming: The argument of "don't learn programming" is a kind of misinformation. When people transitioned from assembly language to high-level languages, there were similar concerns. AI is lowering the barriers to programming, and in the future, more jobs will require programming skills.

14. Domain knowledge enhances AI: A deep understanding of specific fields enables you to utilize AI more effectively. Art historians can create better image prompts. Doctors can shape better health AI. Founders should combine domain knowledge with AI literacy.

15. The product manager is now the bottleneck: Currently, the new constraint is not engineering, but product management. A certain team led by Andrew Ng even suggested adjusting the ratio of product managers to engineers to 2:1 to accelerate feedback and decision-making processes.

16. Engineers need product thinking: engineers with product intuition act faster and develop better products. Technical skills alone are not enough; developers also need to have a deep understanding of user needs.

Seeking friends' opinions

18. Deep knowledge of AI remains a competitive moat: AI literacy has yet to become widespread. Those who truly understand the principles of AI technology still have a significant advantage—they can innovate in smarter, more efficient, and more autonomous ways.

19. Hype ≠ Truth: Be wary of narratives that sound impressive but are primarily used for fundraising or elevating status. Terms like AGI, extinction, and infinite intelligence are often signals of hype rather than signals of impact.

20. Security is about usage, not the technology itself: The concept of "AI safety" is often misunderstood. AI is like electricity or fire; it is neither good nor bad in itself, depending on how it is applied. Security is about usage, not the tool itself.

21. The only important thing is whether users love it: there is no need to get overly caught up in model costs or performance benchmarks. The only question that needs to be addressed is: are you creating a product that users truly love and are willing to continue using?

22. Education AI is still in the exploratory phase: Companies like Kira Learning are conducting extensive experiments, but the ultimate form of AI in the field of education remains unclear. We are still in the early stages of transformation.

23. Beware of "doomsday theories" and "regulatory capture": excessive fear of AI is being used to justify regulations that protect existing businesses. Be skeptical of "AI safety" narratives that favor those in power.

Physical Intelligence Co-creation: Chelsea Finn

1. Robotics requires a full-stack mindset: you can't just add robotics to an existing company. You need to build the entire tech stack from scratch — data, models, deployment.

2. Data quality outweighs quantity: Massive datasets from the industry, YouTube, or simulated environments often lack diversity and authenticity. Accurate, high-quality data is more important than scale.

3. Best Mode: Pre-training + Fine-tuning: First, pre-train on a broad dataset, and then fine-tune using about 1000 high-quality, context-consistent samples. This method can significantly enhance robot performance.

4. General-purpose robots will surpass specialized ones: Universal models that can cross different tasks and hardware platforms (such as third-party robots) are proving to be more successful than systems built for specific purposes.

5. Real-world data is irreplaceable: While synthetic and simulated data can be helpful, real data remains crucial, especially for complex visual and physical tasks.

6. Having too many resources can be counterproductive: excessive funding or complicating things too much can slow down progress. Clarity of the problem and focused execution are the most important.

Cursor CEO & Co-founder: Michael Truell

1. Start early and build continuously: Even when partners dropped out, Michael continued to code. Early viral spread (a clone of Flappy Bird) helped him build confidence and skills.

2. Rapid validation, even in unfamiliar domains: Their team built a programming assistant in the field of mechanical engineering without prior experience. Their creed is "learning by doing."

3. Differentiated positioning, no need to fear giants: They once hesitated about whether to compete with GitHub Copilot, but later realized that few companies aim for "full-process development automation." This positioning opened up the market for them.

4. From code to release, quick action: They only took 3 months from the first line of code to public release. Rapid iteration helped them quickly calibrate the product direction.

5. Focus over complexity: They decisively abandoned the plan to simultaneously develop an IDE (Integrated Development Environment) and AI tools. By concentrating on the AI functions themselves, they achieved faster development.

6. Distribution can start from a tweet: The early user growth of Cursor stemmed from a tweet by the co-founder on social media. Before the official market promotion, word of mouth had already become the main driving force.

7. The compound effect of execution: In 2024, Cursor's annual recurring revenue grew from $1 million to $100 million within a year, achieving a 10% weekly compound growth driven by product improvements and user demand.

8. Best advice, follow your curiosity: forget about those things done to embellish your resume. Michael's main advice is: do what interests you with smart people.

Figma CEO & Co-founder: Dylan Field

1. Find a co-founder who can inspire you: Dylan's motivation comes from working with his co-founder Evan Wallace, "Every week feels like creating the future."

2. Start early and learn by doing: Dylan started his entrepreneurial projects when he was 19 and still in college. The failures of early projects like the "meme generator" ultimately honed the great idea that became Figma.

3. Rapid release, faster feedback: They contacted early users via email for quick iterations and insisted on charging from the very beginning. Feedback is a continuous driving force for product evolution.

4. Break down the long-term roadmap into short-term sprints: Decomposing a grand vision into smaller parts is key to ensuring speed and execution.

5. Product-market fit may take years: Figma took five years to receive a decisive signal: Microsoft stated that if Figma didn’t start charging, they would have to terminate the partnership.

6. Design is the new differentiating factor: he believes that due to the rise of AI, design is becoming increasingly important. Figma is also responding to this trend by launching a series of new products such as Draw, Buzz, Sites, and Make.

7. Accelerating Prototype Design with AI: The best use of AI is to enhance iteration speed rather than pursuing the "magic" of one-click generation. Designers and product managers must now contribute to AI assessments.

8. Embrace rejection instead of avoiding it: Dylan learned to face criticism and feedback with composure through his childhood performance experiences. He believes that being rejected is part of the journey to success.

9. Human connections are always at the core: a warning not to replace interpersonal relationships with AI. When asked about the meaning of life, he replied: "Explore consciousness, keep learning, share love."