I spent the past weekend at YC Startup School this summer. After some reflection, these are my main takeaways.
1. Stay curious about things even when it doesn’t seem relevant
- Approach things from first principles and chase answers to questions even if it doesn’t seem relevant.
- Reading fast-moving fields like tech requires an almost tactile sense.
- Jensen Huang compared it to how a surfer reads waves that look like chaos to everyone else.
- This only comes from staying curious enough for long enough.
2. Systems thinking is the most useful future skill, according to Jensen Huang
- Systems thinking is reasoning about how components interact and where the whole thing is constrained, rather than optimizing any single component.
- This matters because low-level work will be done agentically.
- The value moves up a level, to whoever can design and hold the whole system in their head.
3. Own a domain, not just a tool
- NVIDIA is big because they don’t just own chips, they own the entire domain of graphics, and now deep learning.
- Value is in owning the whole stack around a class of problems, not just the tool.
4. Physical AI is the next big market
- Jensen Huang expects the physical AI business (self-driving cars, etc.) to become NVIDIA’s next $100B business within ten years.
- A related idea is product overhang: capabilities today’s models already have that no product has elicited yet.
- Give the model slightly harder tasks than you think it can do.
5. AI is better than doctors and may replace them
- Vinod Khosla expects AI to almost completely replace doctors, with fees dropping significantly: he calls it “20% doctor included,” with humans doing the remaining 20%.
- Even current AI gives more reliable and consistent diagnoses than humans. If you ask 3 different doctors, you can get 3 different diagnoses.
- AI can pull from large corpuses reliably, so treatment isn’t based on one person’s opinion.
- Doctors can’t reliably keep track of the thousands of relevant medical articles coming out each year, even within their own specialty.
- AI doesn’t need to be perfect, just better than humans.
6. AI is the only way to scale medicine
- AI enables N=1 patient drug design, where AI systems design drugs for one specific patient, which Khosla argues can sidestep the usual FDA regulatory path.
- Mental healthcare is one of the earliest fields to see this (ex. Limbic).
7. “How hard can it be?”
- Even though things will be much harder than you think, you should still approach it asking “how hard can it be?”
- Never be scared to get started. You only need a vision, it’s okay if the first steps are unclear.
8. Be mildly helpful to a lot of people
- “Find a way to be like mildly helpful to a lot of people … it really goes a long way in terms of these things coming together later.” - Sam Altman
- Best way to maximize the chance one of these relationships becomes very useful in the future.
9. Less instruction is better
- Models perform much better with shorter, simpler instructions.
- Every few months, review all the tools and instructions: remove some, see if it does worse or not, then add them back piece by piece.