Lex on AI Interaction Models

Summary of Lex on AI Interaction Models

by Lex Fridman Podcast Fan

13m•May 26, 2026
Listen to Episode

Overview of Lex on AI Interaction Models

This episode discusses a new AI voice interaction approach from Thinking Machines Lab—a “full duplex” system designed to listen and speak simultaneously, much more like a human conversation than today’s typical voice assistants. The hosts frame it as a potential breakthrough for AI receptionists, customer support, and other real-time voice applications where even small latency makes interactions feel robotic. They also discuss the business angle around Mira Murati’s credibility, OpenAI ties, and whether larger labs like OpenAI or Google will quickly copy or surpass the idea.

What Thinking Machines Is Building

“Full duplex” voice interaction

  • Traditional AI voice agents use a sequential loop:
    1. User speaks
    2. System transcribes
    3. Model reasons
    4. System speaks back
  • Thinking Machines’ approach aims to let the model process incoming speech while simultaneously generating its own response.
  • The goal is to remove the awkward pause that makes current voice assistants feel unnatural.

Why this matters

  • The hosts compare it to real human conversation:
    • People interrupt each other
    • People respond mid-thought
    • Conversations are fluid and overlapping
  • Current assistants often:
    • Cut off too early
    • Freeze when interrupted
    • Misread short pauses as the end of a sentence

Key Problems This Could Solve

Latency and awkward turn-taking

  • Even “fast” AI assistants still have a noticeable delay between the end of a user’s speech and the start of the response.
  • This delay creates a robotic feel, especially in customer-facing use cases.

Better conversational flow

  • The new model could make AI:
    • More natural in live conversations
    • Better at handling interruptions
    • More suitable for fast-moving discussions, negotiations, and support calls

Reduced frustration in support systems

  • A major complaint with current chatbots and phone trees is that they:
    • Ask repetitive questions
    • Route users to humans anyway
    • Waste time without solving the issue
  • If the AI can actually resolve problems end-to-end, users may care less whether it is human.

Business and Market Implications

Why Mira Murati’s name matters

  • The hosts note that Thinking Machines has raised major funding largely on the strength of Mira Murati’s reputation and OpenAI pedigree.
  • Murati is described as having a significant OpenAI stake, meaning she already has substantial personal wealth and credibility.
  • That makes investors more willing to back ambitious, early-stage work.

Potential product strategy

  • The episode speculates that Thinking Machines could:
    • Build its own product
    • License the technology
    • Partner with voice platforms like ElevenLabs
  • If the tech is difficult enough, partnerships could be likely.
  • If it’s easy to replicate, larger players may simply copy it.

Competitive pressure

  • The hosts repeatedly question whether this is a durable moat:
    • OpenAI and Google could potentially build similar systems
    • AI features are often reverse-engineered quickly
  • The main risk is timing: even a great product could be overtaken if it launches too slowly.

Adoption Outlook

What needs to happen for broad adoption

  • The tech must feel:
    • Fast
    • Natural
    • Reliable
    • Capable of actually solving user problems
  • The hosts argue that businesses will adopt AI receptionists if they:
    • Can take real action
    • Have proper permissions
    • Reduce support delays from hours/days to immediate resolution

Framing matters

  • A key insight from the discussion: people may resist “replacing humans,” but they are more open to:
    • Doing more than they could before
    • Fixing backlogs
    • Answering missed calls
    • Handling overflow support
  • In other words, the best pitch is not “replace the receptionist,” but “eliminate a bottleneck.”

Main Takeaways

  • Full duplex voice AI could be a meaningful step toward human-like conversation.
  • The biggest win is likely in customer support, AI receptionists, and live voice agents.
  • Latency and interruption handling are still major barriers to making voice AI feel natural.
  • Mira Murati’s credibility helps explain the massive funding behind Thinking Machines.
  • The biggest unknown is whether the company can ship fast enough before OpenAI, Google, or others replicate the idea.
  • If the system can actually solve user problems with the right permissions, users may care far less about whether the agent is human.

Notable Insight

The episode’s core idea is that people don’t just want AI to sound good — they want it to feel conversational and actually solve the problem.

Mentioned Companies and Technologies

  • Thinking Machines Lab
  • OpenAI
  • Google
  • ElevenLabs
  • Claude / similar collaborative AI workflows
  • AI voice receptionists
  • Full duplex speech systems