EPISODE 2026-09-09

Building Systems You Can Keep: From Child Companions to Sovereign Agents

AI safety, child companions, open models, and browser-era agents with Mike Rizkalla and Raffi Krikorian.

▶ Full show on YouTube

AI safety, child-focused companions, open models, and browser-era agent control in one conversation.

The rundown

  1. 2:20Opening30 min
    OpeningAI safety is entering a race to self-improving superintelligence without a clear plan for control. Prakash Narayanan and Nathan Labenz examine a resignation from OpenAI and Anthropic, the capability threshold where useful models may become dangerous, persistent evaluation failures, and whether government can force frontier companies to coordinate.
  2. 1:16:25Interview52 min
    Mike RizkallaMike RizkallaAI companions for kids should help families, not just maximize engagement. Mike Rizkalla, CEO of Snorble, explains the company's approach to bedtime routines, curated content, small language models, child privacy, physical AI, and future applications in elder care.
    Open segment on YouTube ↗
    • Never Put A Camera In A Child's Bedroom

      0:00 / 0:00
    • Open-Ended AI Is Too Much For Kids

      0:00 / 0:00
    • Companions Should Connect Families, Not Kids

      0:00 / 0:00
    • People Forgot The Robot Was A Robot

      0:00 / 0:00
    • Bedtime Should Create Joy, Not Dependence

      0:00 / 0:00
  3. 32:17Interview44 min
    Raffi KrikorianRaffi KrikorianAI agents are moving beyond chat and into browsers, codebases, and the web. Mozilla CTO Raffi Krikorian discusses open-source models, Project Glasswing, agent collaboration, agentic search, AI in schools, election deepfakes, and the safety lessons of FSD and Waymo. He keeps returning to human agency, transparency, and user control as agents become primary users of online systems.
    Open segment on YouTube ↗
    • When Agents Rewrite The Codebase

      0:00 / 0:00
    • An Agent Reverse-Engineers MyFitnessPal

      0:00 / 0:00
    • The Web Becomes 99% Agents

      0:00 / 0:00
    • Pause Generative AI In Schools

      0:00 / 0:00
    • Why FSD Handoffs Feel Unsafe

      0:00 / 0:00
  4. 2:08:49Closing31 min
    ClosingCan AI self-improvement outrun human control? This AI safety discussion examines machine economies, recursive growth, coordination problems, AI doom probabilities, private funding, and what governments can realistically do.

In this episode

Prakash Narayanan and Nathan Labenz open with the latest frontier AI safety warnings, capability thresholds, evaluation failures, and the question of whether governments can realistically coordinate powerful labs. Mike Rizkalla of Snorble then joins to discuss AI companions for kids, bedtime routines, privacy, small models, and physical AI, followed by Mozilla CTO Raffi Krikorian on open models, agentic search, secure-by-design software, schools, election integrity, and human control as agents spread across the web.

  • AI safety is entering a race to self-improving superintelligence without a clear plan for control. Prakash Narayanan and Nathan Labenz examine a resignation from OpenAI and Anthropic, the capability threshold where useful models may become dangerous, persistent evaluation failures, and whether government can force frontier companies to coordinate.
  • AI companions for kids should help families, not just maximize engagement. Mike Rizkalla, CEO of Snorble, explains the company's approach to bedtime routines, curated content, small language models, child privacy, physical AI, and future applications in elder care.
  • AI agents are moving beyond chat and into browsers, codebases, and the web. Mozilla CTO Raffi Krikorian discusses open-source models, Project Glasswing, agent collaboration, agentic search, AI in schools, election deepfakes, and the safety lessons of FSD and Waymo. He keeps returning to human agency, transparency, and user control as agents become primary users of online systems.
  • Can AI self-improvement outrun human control? This AI safety discussion examines machine economies, recursive growth, coordination problems, AI doom probabilities, private funding, and what governments can realistically do.