Andrew Ng YC AI Startup School Notes

The biggest opportunities in AI are often in startups because they’re willing to pursue ideas before they’re obviously valuable.
AI is more useful when integrated into workflows:
- Example: Instead of asking AI to write an entire essay, use it in steps:
  1. Research the topic.
  2. Draft a version.
  3. Critique itself.
  4. Revise iteratively.
- This agentic workflow is how humans work, and how AI can work best too.
From Single-Shot to Structured Intelligence:
- Old paradigm: prompt once → get output.
- New paradigm: task ****orchestration across tools, APIs, and documents.

Traditional AI: Asked to generate output in one go (e.g., an entire essay).
Emerging models: Iterative workflows, where AI participates step-by-step: researching, drafting, critiquing, revising, like collaborating with a human.
Emergence of agentic workflows where systems take structured actions across multiple tasks (e.g., write, research, revise).
These agentic workflows are the dividing line between functioning and non-functioning AI applications today.

Engineering used to be a bottleneck. Now:
- AI-assisted coding is 10x faster for prototypes and throwaway software.
- Production-quality software is 30–50% faster.
Cost of development has decreased, making fast iteration more feasible.
Some old one-way door decisions are turning to be two-way door decisions as it is easier to start from scratch due to faster iteration possibilities.
- For more on one-way & two-way door decisions: https://www.youtube.com/watch?v=rxsdOQa_QkM
Now, product management is the bottleneck:
- Historically: 1 PM per 4–7 engineers.
- Now: Some teams propose 2 PMs per engineer due to engineering speed outpacing product decision-making.