The Eyes Are The (Context) Window to The Soul: How Windsurf Gets to Know You — Sam Fertig, Windsurf

Introduction to Windsurf and Context 00:00

  • The speaker, Sam Fertig, a Principal Forward Deployed Engineer at Windsurf, introduces the concept of inferring information from superficial observations.
  • He demonstrates how framing (e.g., "devil" vs. "angel" interpretations) can change perceptions of simple facts.
  • The presentation emphasizes that hard evidence, combined with visible observations, provides a more complete understanding.
  • Windsurf offers an AI coding toolkit available as an editor or an IDE plugin, aiming to help users code faster and better.
  • Key takeaways so far: surface-level observations are insufficient, educated guesses can be made via heuristics and hard evidence, and Windsurf assists with coding.

The Challenge of Relevant Code Generation 05:29

  • The main problem in the AI coding space is not generating code, as simple UIs and models can easily produce answers from prompts.
  • The difficulty lies in generating code that is specific and relevant to an individual user, fitting into existing large codebases, adhering to organizational policies, matching personal preferences, and being future-proof.
  • The "magic" behind Windsurf's relevant suggestions comes from its understanding of context.

Windsurf's Context Philosophy 07:01

  • Windsurf's approach to context is based on two pillars: "what context" and "how much context."
  • "What context" is categorized into "heuristics" (user behavior and intent, like code around the cursor, open files, clipboard content, terminal activity) and "hard evidence" (the environment and codebase, such as documentation, rules, agent memories, and repository content).
  • The formula for relevant output is: user prompt + state of your codebase + user state.
  • "How much context" focuses on optimizing what context is included in the LLM call, rather than just increasing the context window size.
  • This optimization improves results and helps address latency, which is crucial for good AI coding agents.
  • Windsurf excels at finding relevant context quickly, especially in large codebases, due to its background in GPU optimization (originally Exaf Function).
  • Windsurf utilizes various methods and connectors for context finding, including embedding search, memories, rules, custom workspaces, and plain text search.

Data Privacy and Handling 12:35

  • Windsurf processes information only within the user's editor.
  • Data sent to Windsurf servers is stateless, meaning it's a pass-through transaction.
  • Windsurf does not store or train on user data.