Pika's Demi Guo launches 'AI Selves' — a personalized always-on avatar that learns and monetizes your expertise
Feb 26, 2026 with Demi Guo
Key Points
- Pika co-founder Demi Guo launches AI Selves, a personalized avatar trained on user data and voice that operates continuously as a scalable proxy for expertise monetization.
- The avatar learns through three inputs: imported content libraries, real-time correction during interactions, and identity itself, though bootstrapping sufficient self-documentation remains a practical constraint.
- Guo positions AI Selves as operating at baseline quality ten times faster than the original person, enabling experts to charge for access without direct availability.
Summary
Demi Guo, co-founder and CEO of Pika, launched AI Selves, a product that creates personalized, always-on AI avatars trained on user data, voice, and knowledge.
The core value proposition is availability. An AI Self can help a user's family debug a tech problem, join a group trip the user can't attend, or field questions from people seeking access to the user's expertise. Guo positions it as a living extension of a person, not a static chatbot.
Training
The avatar learns from three sources. Users can import existing content such as podcast archives or writing to establish a baseline. The avatar learns continuously as users correct and guide it during real interactions, gradually absorbing taste and judgment. Identity itself carries weight too. A family member talking to an AI Self may not need perfect accuracy, just enough recognizable voice to feel like real access.
Expertise as a product
The commercial use case is selling access to expertise. A skilled marketer could let others consult their AI Self and charge for that access. Guo argues the AI version operates at comparable quality but ten times faster, making it a scalable proxy for knowledge work.
The product is early. Most people don't generate enough self-documentation to seed a convincing replica, and Guo doesn't fully resolve that bootstrapping problem. The continuous-correction loop is meant to close the gap over time, but the initial data wall remains real.