Al Engineering 101 with Chip Huyen (Nvidia, Stanford, Netflix)

Chip Huyen, a machine learning engineer and author of the book AI Engineering, who has worked at NVIDIA, Netflix, and taught at Stanford, joins Lenny Rachitsky to explain core AI concepts and practical lessons for building AI products. The conversation covers pre-training versus post-training and fine-tuning, reinforcement learning from human feedback, evals, retrieval-augmented generation, measuring AI productivity in companies, and how organizations and product work may change.

01Key learnings

• Focus on what actually improves AI apps, such as talking to users, improving data, writing better prompts, and optimizing workflows, rather than chasing every new tool or constant news.
• Be cautious about committing to newly released technologies that have not been widely tested, since switching away from them later can be costly.
• Fine-tuning and post-training can shape model behavior a lot, and many teams now focus effort there because base pre-training data is largely similar across labs.
• Evals should be treated as an investment: build them most thoroughly for high-stakes or core use cases, and measure each step of multi-step systems rather than only end-to-end outputs.
• In RAG systems, improving data preparation such as chunk design, added context, and metadata often yields bigger quality gains than agonizing over which vector database to use.
• Measuring AI productivity is hard; consider controlled comparisons across performance tiers, and recognize that senior engineers' value increasingly lies in system thinking and reviewing AI-generated code.
• To find AI product ideas, pay attention to your own recurring frustrations over a week and consider whether a small tool could address them.

02Frameworks mentioned

Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.