Divya van Mahajan

Towards AI That Can Actually Interact

ai-daily-brief-podcast

Towards AI That Can Actually Interact: Interaction Models from Thinking Machines Lab

Overview

This episode of the AI Daily Brief (hosted by Nathaniel Whittemore, though he does not name himself in this episode) covers a new class of AI model called interaction models, introduced by Thinking Machines Lab (TML) — a startup founded by former OpenAI CTO Mira Murati. The core argument is that current AI systems are architecturally constrained by turn-based interaction, which creates a fundamental bottleneck in human-AI collaboration, and that TML’s new approach represents a potential paradigm shift in how humans and AI systems work together.

Source video: No URL provided. The episode aired on the AI Daily Brief podcast/channel on or around May 12, 2026.

Prerequisites

  • Familiarity with how current large language models (LLMs) and voice AI systems work, particularly the concept of turn-based dialogue
  • Basic understanding of the AI lab landscape (OpenAI, Anthropic, Google DeepMind, etc.)
  • Awareness of multimodal AI capabilities (text, audio, video)
  • General knowledge of AI agent frameworks and real-time audio APIs
  • Optional: familiarity with the history of human-computer interaction, including CLI vs. GUI paradigms

Main Points

1. Headlines: OpenAI Launches DeployCo, a Forward-Deployed AI Consulting Firm

  • OpenAI has officially launched the OpenAI Deployment Company (DeployCo), structured as a joint venture with 19 partners across consulting, private equity, and finance.
  • Initial investment: $4 billion at a $10 billion pre-money valuation; lead investor is TPG, with Advent International, Bain Capital, and Brookfield as co-lead partners.
  • DeployCo is built around the acquisition of engineering firm Tomorrow, providing ~150 staff experienced in AI deployment from day one.
  • Motivation from partners is partly to skip the queue for AI transformation support in their portfolio companies.
  • Key thesis: no matter how powerful models become, they will collide with institutional inertia; meaningful support structures are required for enterprises to close the capability overhang.
  • A parallel effort from Anthropic (as yet unnamed) is also underway; Goldman Sachs has backed both.

2. Headlines: Anthropic and OpenAI Crack Down on Unauthorized Secondary Market Stock Trading

  • Anthropic updated its documentation to explicitly void transfers of shares to SPVs (Special Purpose Vehicles) and called out specific firms by name offering unauthorized Anthropic stock access.
  • OpenAI issued a similar statement reaffirming that unauthorized transfers are legally void.
  • The announcement triggered a ~50% crash in Anthropic’s price on gray secondary markets.
  • Many retail investors are unknowingly holding layered financial abstractions — tokenized receipts for possible future exposure to SPVs that may or may not hold actual equity.
  • Broader concern: if major private companies systematically invalidate SPV structures, a reckoning in private markets may follow, especially around high-profile eventual IPOs like SpaceX.

3. Headlines: White House Walks Back FDA-Style AI Regulation

  • National Economic Council Chairman Kevin Hassett had floated an FDA-like approval process for AI models, triggering significant industry backlash.
  • Former AI czar David Sacks and Hassett himself both walked back the proposal; Hassett stated: “Nobody has an idea that we should do something like bring in a giant new bureaucracy to approve AIs.”
  • Current approach is described as direct collaboration between administration officials and AI labs to prevent extreme harm pre-release.

4. Headlines: Trump Tech Envoy to China — Jensen Huang Notably Absent

  • Trump is assembling a tech delegation for China trade talks including Elon Musk, Tim Cook, and Meta’s Dina Powell McCormick, plus semiconductor and finance executives.
  • Jensen Huang is absent despite expressing willingness to join; this may signal that NVIDIA’s AI chips are off the table in trade negotiations.
  • Zero export licenses for H200 GPUs to China have been approved by the Commerce Department despite earlier indications they would be.

5. Main Story: Thinking Machines Lab and the Concept of Interaction Models

Background on Thinking Machines Lab

  • Founded by Mira Murati (former OpenAI CTO) with a superteam of researchers poached from major labs.
  • Raised low billions in funding — significant but modest compared to frontier labs.
  • First product, Tinker (October prior year), was an RL-as-a-service platform for fine-tuning open-source models; received limited attention.
  • Notable co-founder departures (Barrett Zoff, Luke Metz) returned to OpenAI in January.

The Problem: The Collaboration Bottleneck

  • Current AI systems operate in a single-threaded, turn-based fashion: the model waits while the user speaks/types, then generates a response while perception freezes.
  • This creates a narrow channel for human-AI collaboration — users must batch their thoughts, cannot point at things, and must phrase requests like emails.
  • TML’s blog post frames this as: “The interface doesn’t leave room for us, so we adapt to the models.”
  • Analogy used throughout: current AI interaction is like trying to resolve a disagreement over email rather than in person.

The Proposed Solution: Interaction Models

  • Interaction models are trained from scratch around continuous, time-aware exchange rather than discrete turns.
  • Instead of flattening inputs and outputs into one ordered token sequence, the model processes parallel input and output streams split into 200-millisecond microturns.
  • The model is in constant two-way exchange — perceiving and responding simultaneously.
  • Architecture is a two-part system:
    1. A real-time interaction model that stays present with the user (listening, speaking, reacting)
    2. A background model that handles longer reasoning, browsing, tool use, and agentic work
  • The interaction model keeps talking/listening while the background model works, then weaves results into conversation when appropriate.

Demonstrated Capabilities

  • Simultaneous translation: begins translating speech from one language to another while the speaker is still talking (similar to live human interpreters).
  • Dialogue management: tracks whether the speaker is thinking, yielding, self-correcting, or inviting a response — no hardcoded system, adapts to context.
  • Visual proactivity / interjection: e.g., noticing when a researcher starts slouching and reminding her to correct posture without being asked.
  • Real-time professional softening: researcher speaks candidly; the model reformulates the statement into a more socially appropriate version simultaneously.
  • Background search during live conversation: model searches the web while conversing, incorporating results naturally — demonstrated with a conversation about The Devil Wears Prada 2, a film not in its training data.

New Benchmarks Introduced

  • TimeSpeak: tests whether the model can initiate speech at user-specified times with correct content (e.g., breathing reminders every four seconds).
  • QSpeak: tests whether the model speaks at the appropriate moment with semantically correct responses (e.g., code-switching correction).
  • The necessity of inventing new benchmarks signals that the capability set is genuinely novel.

6. Framing: The GUI Moment for AI

  • TML co-founder John Schulman argues that interactivity and collaboration capabilities are under-emphasized relative to intelligence and autonomy because they’re harder to evaluate.
  • TML researcher Claire Birch draws an explicit parallel to the transition from CLI to GUI: the GUI democratized computing by removing the requirement for users to “think like the computer.”
  • Her argument: current chat interfaces are still surprisingly CLI-like — they reward verbal fluency, careful prompting, and procedural skill, not natural human communication.
  • The GUI moment for AI will be: “when the user no longer has to think like the computer, or like the AI, or like the prompt engineer, in order to access the machine’s capabilities.”
  • Interaction models are positioned as a step toward that moment.

7. Community Reaction and Broader Implications

  • TML’s messaging was notably cohesive — all team members communicated a consistent narrative about increasing human-AI bandwidth and keeping humans as “main characters.”
  • Professor Ethan Mollick noted demos leaned toward fun/novelty use cases and called for more demonstrations of high-value applications (meetings, education, training).
  • Developer Nick Dobos framed it as a tech demo for a developer/VC audience, not a consumer product demo.
  • Community consensus: this likely will not stay unique for long — frontier labs iterate quickly on each other’s abstractions.
  • OpenAI’s GPT Real-Time 2 model was already being demonstrated doing similar background-agent work (e.g., updating a Kanban board during a stand-up meeting).
  • Speculative future: “Software engineering is 100% meetings and your AI note-taker orchestrates all your coding agents in the background.”

Key Concepts

  • Interaction Model: A new class of AI model trained from scratch to handle continuous real-time interaction natively, rather than as a layer added onto a turn-based system.
  • Turn-based Model: The dominant current architecture in which a model waits for user input to finish, generates a response, then waits again — creating discrete, sequential dialogue turns.
  • Microturn (200ms): The unit of time-aligned processing in TML’s interaction model, allowing parallel input and output streams rather than sequential ones.
  • Collaboration Bottleneck: TML’s term for the structural limitation imposed by turn-based interfaces that prevents users from naturally interrupting, pointing, correcting, or interacting fluidly with AI.
  • Visual Proactivity: The ability of a model to initiate speech or action based on changes in the visual environment without waiting for an audio prompt.
  • Background Model: The second component of TML’s two-part system, handling long-form reasoning, tool use, browsing, and agentic tasks while the interaction model maintains the live user-facing conversation.
  • TimeSpeak: TML’s benchmark measuring whether a model can initiate speech at user-specified times with correct content.
  • QSpeak: TML’s benchmark measuring whether a model speaks at the contextually appropriate moment with semantically correct responses.
  • Unlock Index / Unlock Score: A proposed alternative to traditional benchmarks that measures how many and what types of new use cases a model enables, rather than performance on existing tasks.
  • DeployCo (OpenAI Deployment Company): OpenAI’s newly launched forward-deployed engineering joint venture to help enterprises implement AI at scale.
  • SPV (Special Purpose Vehicle): A legal holding entity used in finance; in the AI context, unauthorized SPVs have been used to offer exposure to private company stock outside official transfer approval processes.
  • Capability Overhang: The gap between what AI models can theoretically do and what enterprises are actually extracting from them due to implementation and adoption challenges.
  • GUI Moment: Claire Birch’s metaphor for a future inflection point where AI interaction becomes as naturally accessible as a graphical interface was relative to the command line.

Summary

The central argument of this episode is that Thinking Machines Lab has identified a genuine architectural limitation in current AI systems — their turn-based, single-threaded nature — and has built a new class of model, the interaction model, designed from the ground up to engage in continuous, time-aware, multimodal interaction. By processing parallel input and output streams in 200-millisecond microturns and pairing a real-time user-facing model with a background reasoning and agentic model, TML’s system can do things no existing commercial real-time API can: visually proactive interjection, simultaneous translation, real-time contextual correction, and background search woven naturally into live conversation. The broader philosophical framing — articulated through the CLI-to-GUI analogy — is that current chat interfaces still require users to adapt to the machine, and that the next paradigm shift will occur when AI interaction becomes as natural and fluid as human-to-human communication. While the host and community acknowledge this capability gap may not remain unique to TML for long given how rapidly frontier labs iterate, the announcement is treated as a meaningful signal that persistent, real-time, multimodal interaction represents an emerging and important category of AI capability with significant implications for enterprise, education, and everyday use.