What It Takes to Build an AI That Actually Knows You

What It Takes to Build an AI That Actually Knows You

More

Summary

In this a16z conversation, Ken from First Person Labs discusses the company’s approach to building an AI that genuinely understands its user. The startup is creating a persistent model of a person, capturing memories, priorities, preferences, goals, values and intent, and giving the individual control over it.

The discussion centers on the idea of a portable “second” that travels between agents. Instead of re-explaining seat preferences or airport habits to every assistant, a user’s agents such as Codex or Grockbot can consult their second for context. The conversation describes how this model starts from baseline artifacts and gradually builds a richer understanding, using an Alfred-and-Batman analogy to illustrate delegation with trust.

A large portion of the interview addresses privacy. The speakers explain why zero data retention is only a policy promise, and how confidential inference on attested hardware such as Nvidia’s can offer technically verifiable guarantees. They also cover storage choices, including a local-only option and an encrypted cloud-hosted version similar to password managers. The conversation also touches on surprising development use cases, such as screenshotting an issue and letting a coding agent infer what to fix. It is a useful look at the emerging idea of portable personal context for AI agents.


📺 Source: a16z · Published November 18, 2026
🏷️ Format: Interview

1 Item

Channels