Building AI Startups at Agent Speed: The YC Startup School Talk

In his Y Combinator talk, Andrew Ng argues that AI coding assistants make building cheaper and faster, so the slower part becomes deciding what to build and learning whether customers want it. He bases this on his experience at AI Fund, a venture studio he describes as co-founding about one startup a month. These are his own observations, not a measured result.

His first step is a concrete idea: one specified in enough detail that an engineer can build it. "Use AI to optimize healthcare assets" is too vague, because different engineers would build different things. Letting hospital patients book MRI slots online is concrete. It may prove right or wrong, but a team can find out quickly, while a vague idea mostly collects agreement.

He then lists ways to get feedback, from faster and less accurate to slower and more accurate: judge the product yourself, ask a few friends, ask strangers, send prototypes to around 100 testers, and run A/B tests. He says teams should also study the results to sharpen their instincts about users, so later decisions come faster.