Why Notion’s biggest AI breakthrough came from simplifying everything

Initially, Notion AI engineers utilized intricate code generation and complex data modeling for their LLMs and agentic AI. However, they soon discovered that simpler approaches yielded better results. They shifted to using straightforward prompts, human-readable formats, and minimal abstraction, specifically opting for markdown. This "rewiring" significantly improved model performance, leading to the successful release of Notion's V3 software. A key feature of V3 is customizable AI agents, which have become Notion's most popular AI tool to date. Ryan Nystrom, Notion's AI engineering lead, emphasizes describing prompts as if explaining them to a person. He stresses the value of using markdown to represent Notion pages, enabling easier model interaction and manipulation of text. Instead of over-complicating context, they found a sweet spot for token limits, avoiding information overload. Notion prioritizes a curated set of tools, resisting the temptation to add endless features for the models to consider. The team advocates for a spartan approach, emphasizing plain language, and using APIs as intended. This strategy resulted in a highly effective AI-powered product.