Andrew Ng: Building Faster with AI

Ng argues that the largest startup opportunities are usually in applications, which can generate revenue to pay for infrastructure, cloud, and foundation models. He highlights agentic workflows as a way to improve results: models can research, draft, critique, and revise instead of producing an answer in one pass. AI Fund has used these workflows on tasks including compliance documents, medical diagnosis, and legal reasoning.

Ng says a product idea should be specific enough for an engineer to build. “Optimize healthcare assets” is too vague; letting patients book MRI slots online is concrete and can be tested quickly. He recommends committing to one clear hypothesis at a time, using subject matter expertise and customer knowledge to make decisions, then pivoting when evidence warrants it.

Ng distinguishes production code from quick prototypes: he estimates AI may make production coding 30 to 50 percent faster, while prototypes can be at least 10 times faster. He encourages building many low-cost prototypes, while making code secure and scalable before shipping it. As engineering speeds up, product feedback can become the bottleneck; his feedback options range from personal judgment and a few testers to larger tests and A/B experiments.

Ng says AI coding tools are changing quickly, and teams may benefit from keeping up with the latest generation. Cheaper engineering can make codebases and architecture decisions easier to change. He argues that people in every role can benefit from learning to code or directing AI tools, and says product managers with coding ability and engineers with product instincts can be especially effective.

Ng says understanding AI can prevent teams from spending months pursuing the wrong technical approach; choices such as prompting, fine-tuning, and workflows can lead to very different results. He describes tools like retrieval augmented generation, evaluations, guardrails, and embeddings as building blocks that can be combined into new products. In the Q&A, he advises prioritizing products users love, evaluating social effects, and stopping projects when the team believes they would make people worse off.