Mira Murati's Thinking Machines Lab Bets on Human-Centered AI, Not Replacement
The ex-OpenAI CTO's startup previews interaction models designed to keep humans in decision-making loops, contrasting with big labs pursuing autonomous superintelligence.
Last verified:
Mira Murati’s Case for AI That Keeps Humans in Control
Thinking Machines Lab, founded by ex-OpenAI CTO Mira Murati, is previewing a new class of AI models designed to embed human decision-making into every interaction, rather than automate humans out of the loop. According to Wired AI, the startup’s “interaction models” natively process continuous human communication—including pauses, tone shifts, and subject changes—allowing the system to adapt in real time based on clarification or new context. This positions Murati’s vision directly against the industry’s prevailing approach: autonomous agents trained to execute complex tasks, including software engineering, with minimal human involvement.
How Interaction Models Work Differently
The technical distinction is meaningful. Existing voice-mode AI systems typically follow a pipeline: capture audio, transcribe to text, pass the transcript to a language model for processing. According to Wired AI, Thinking Machines’ interaction models collapse that chain—they understand conversational meaning directly from audio and video streams, without the lossy intermediate step of transcription. Alexander Kirillov, a founding team member and multimodal AI specialist, frames this as enabling “customized and personalized” AI that perceives user context continuously.
The company demonstrated these capabilities in video previews, though the models remain unreleased. Tinker, the lab’s only public product to date—launched in October 2025 as an API for fine-tuning open-source models with proprietary data—hints at the company’s broader strategy: giving users and researchers the tools to adapt frontier models rather than depend on fixed, vendor-controlled systems.
A Contrarian Bet in a Race Toward Autonomy
Murati departed OpenAI in 2024 to co-found Thinking Machines with several prominent engineers and has raised billions to develop frontier AI. Yet her positioning stands apart from the major labs. According to Wired AI, OpenAI, Anthropic, and Google are pursuing superintelligence by scaling models to handle increasingly complex, autonomous work—writing entire applications from a text prompt. Murati argues this framing presents a false choice. “At some point we will have super-intelligent machines,” she tells Wired. “But we think that the best way to actually have many possible futures—good futures—is to keep humans in the loop.”
Other startups, including Humans&, are exploring similar collaborative frameworks. Some economists have also called for the AI industry to prioritize human empowerment over replacement.
Why This Matters
The interaction-model preview signals a structural disagreement about how frontier AI will be deployed. If adoption favors systems that require active human guidance and customization—rather than autonomous agents—the economic incentives shift: companies will compete on ease of human collaboration and model transparency, not pure task automation. This could reshape hiring and organizational design, at least for teams that adopt Murati’s tools.
However, the pitch is not yet proven at scale. Thinking Machines has released only one product (Tinker), and the interaction models remain in preview. The real test is whether the market will pay a premium for collaboration-first design when autonomous alternatives promise faster execution and lower operational cost.
Frequently Asked Questions
What are Thinking Machines Lab's interaction models?
They are AI models trained to understand continuous human communication through camera and microphone input, natively processing pauses, tone shifts, and interruptions rather than transcribing speech and treating it as text.
How does this differ from existing voice AI assistants?
Most voice assistants capture speech, transcribe it, and feed the text into a language model. Interaction models skip transcription and understand conversational nuance directly, allowing real-time adaptation.
What has Thinking Machines Lab already released?
The company launched Tinker in October 2025, an API tool for fine-tuning open-source models with custom data. The new interaction models have not yet been released publicly.