The Vision Behind Keeping People Central to AI Development
Mira Murati stepped away from her role as chief technology officer of OpenAI in 2024, but her ambition to build artificial superintelligence never faded. What changed was her conviction about how to get there. Rather than pursuing machines that operate independently of human input, she founded Thinking Machines Lab with a specific bet: the most powerful AI systems will succeed because they keep people deeply involved in every step. This philosophy, often called “human in the loop” design, treats human judgment not as a bottleneck but as an essential ingredient. For anyone worried about AI replacing their career or making decisions without oversight, Murati’s approach offers a different path forward.
1. Designing Interaction Models That Grasp Messy Human Communication
Most voice assistants today follow a rigid pattern. You speak, the system transcribes your words, feeds that text into a language model, and returns a reply. Thinking Machines Lab rejected that pipeline. Instead, the company built interaction models trained to process continuous, unstructured human speech through a camera and microphone without first converting everything to text.
Why Pauses and Interruptions Matter
These models do something that traditional systems cannot. They interpret the meaning behind a pause, a shift in tone, or a sudden interruption. When you stop mid-sentence to rethink your point, the model recognizes that hesitation as meaningful rather than treating it as an error. When you interrupt yourself to clarify something, the model follows the new direction immediately. Alexander Kirillov, a founding team member and multimodal AI expert at Thinking Machines, explained that the turns in conversation are determined by the model itself rather than by a separate, less intelligent scheduling system. This capability fundamentally changes how natural collaboration feels.
A Real Scenario for the Reluctant Developer
Imagine a software developer who spends hours each day pair-programming with an AI coding assistant. Current tools require precise prompts. If the developer changes their mind mid-request, they often need to start over. With interaction models that understand messy communication, the developer could say something like, “I need a function that sorts this list by date—actually, no, let me filter it first by status—and then sort.” The model follows the real-time shift without confusion. That kind of fluid interaction preserves the developer’s creative control rather than forcing them into a rigid input format.
2. Keeping Human Customization at the Core of Frontier Models
Murati argues that the best path toward superintelligence does not involve removing people from the equation. Instead, she envisions a future where individuals customize their own frontier AI models to reflect their personal values and workflows. This stands in stark contrast to the approach taken by most large AI labs today, where one massive model serves millions of users with minimal variation.
Customization as a Safeguard
A more optimistic approach, Murati suggests, is to let people tailor their AI tools the same way they tune a musical instrument or adjust the settings on a camera. When you can fine-tune a model to match your ethical boundaries, your industry terminology, or your communication style, you retain authority over how the technology behaves. For a small business owner, this could mean training an AI assistant to understand their specific inventory system, their preferred tone with customers, and their unique return policies. The model does not replace the owner’s judgment; it amplifies it.
The Teacher Facing Classroom Integration
Consider a teacher trying to bring AI into the classroom without losing personal interaction. A generic model might offer lesson plans that ignore the teacher’s pedagogical philosophy or the specific needs of their students. A customizable frontier model, by contrast, could be adjusted to emphasize Socratic questioning, incorporate local history examples, or respect the teacher’s grading preferences. The teacher remains the decision-maker. The AI acts as an adaptable assistant rather than an autonomous instructor.
3. Building AI That Adapts in Real Time to Human Behavior
Traditional AI systems process input in discrete chunks. You submit a query, wait for a response, then submit the next query. Thinking Machines Lab’s interaction models break that cycle. They perceive what a person is doing continuously and remain ready to reply, search for information, or activate tools at any moment. Kirillov described this as a model that is constantly present and constantly aware.
Adapting to Clarifications on the Fly
When a person clarifies a point or changes the subject mid-conversation, these models adjust without requiring a reset. This is something that none of today’s other models can actually do, according to Kirillov. The result feels less like operating a machine and more like working alongside a colleague who pays close attention. For users who find existing AI frustrating because it misunderstands context or loses the thread of a conversation, this represents a meaningful improvement.
Hypothetical Scenario for a Busy Parent
A parent planning a family trip might start by asking for flight options, then pivot to hotel recommendations near a specific landmark, then circle back to flights with a different date. A conventional chatbot would probably lose track or require the parent to repeat themselves. An interaction model that adapts on the fly follows the natural zigzag of human thinking. The parent saves time and frustration, and the AI proves genuinely helpful instead of exhausting.
4. Prioritizing Human Collaboration Over Full Automation
Many companies racing toward superintelligence emphasize automation. They build systems that write code, generate reports, and even manage software projects with minimal human involvement. Murati’s vision moves in the opposite direction. She wants AI to collaborate with people rather than replace them. Human-in-the-loop AI could preserve jobs rather than eliminate them, because it positions the technology as a partner that handles repetitive tasks while leaving creative and strategic decisions in human hands.
What This Means for Job Security
A 2023 study from the McKinsey Global Institute estimated that about 30 percent of work activities could be automated by 2030, but the same report stressed that most jobs would change rather than vanish. Murati’s philosophy aligns with that nuanced view. By designing systems that require ongoing human involvement, she effectively builds a safeguard against mass displacement. The model handles what it does best—processing vast amounts of information quickly—while the human handles what they do best: setting goals, making value judgments, and applying context.
The Developer Who Wants to Stay Creative
For the software developer worried about automation, this distinction matters. If AI writes entire applications from a single prompt, the developer’s role shrinks to typing prompts and reviewing output. But if AI acts as a collaborative tool that understands intent and adapts to the developer’s decisions, the developer retains ownership of the architecture, the design choices, and the final product. Murati’s approach suggests that the future of work is not about humans versus AI but about humans working with AI.
5. Opening Fine-Tuning Tools Like Tinker for Individual Control
Thinking Machines Lab released its first product, Tinker, in October 2025. Tinker is an API that lets researchers and engineers fine-tune open source frontier models using their own custom data. This tool gives individuals and small teams the ability to shape AI behavior without needing a massive budget or a dedicated machine learning department.
Democratizing Model Customization
Before Tinker, fine-tuning a frontier model required significant technical expertise and computational resources. Even then, the process was opaque. Tinker lowers the barrier. A researcher studying rare diseases could feed the model a collection of specialist journals. A small e-commerce business could train it on customer service transcripts. A local journalist could tune it to follow their editorial guidelines. The result is AI that reflects the user’s domain knowledge and values rather than a one-size-fits-all baseline.
Why Fine-Tuning Strengthens the Human Role
When you fine-tune a model, you are essentially writing a manual for how it should behave in your context. That act of teaching is deeply human. It requires the user to think carefully about what they want, what they do not want, and what edge cases might arise. This process keeps the human firmly in the decision loop. The model does not learn in isolation; it learns from human examples and corrections. Murati has described Tinker as the first bet on human collaboration, and it represents a concrete way for people to exercise control over powerful technology.
You may also enjoy reading: Cardiovascular Tech Career Guide: Steps to Start.
6. Ensuring AI Captures Intent and Values, Not Just Commands
Murati has stated that the ultimate direction of this work is amplifying people’s own preferences and values, with AI actually understanding intent and predicting intent. Most current systems respond to explicit commands. They do what you say, not what you mean. By training models to grasp intention, Thinking Machines Lab aims to close that gap.
Reading Between the Lines
Understanding intent means recognizing when a user is being sarcastic, when they are asking a rhetorical question, or when they are expressing uncertainty. It means predicting what a user will need next based on the trajectory of the conversation. For example, if a user asks, “What’s the weather like this weekend?” and then pauses, a model that understands intent might follow up with, “Would you like me to suggest activities based on the forecast?” instead of waiting for another explicit command.
A Practical Example from the Workplace
A project manager might tell an AI assistant, “I need the Q3 numbers, but prioritize anything that looks concerning.” A command-based system would simply retrieve the numbers. An intent-aware system would recognize that the manager wants a flagged summary highlighting risks, not a raw spreadsheet. The manager stays in control of what matters—making decisions about the risks—while the AI handles the sorting and prioritization. That division of labor keeps the human role substantive rather than clerical.
7. Funding a Vision That Places Human Judgment Alongside Superintelligence
Thinking Machines Lab has raised billions of dollars to build frontier AI. That financial backing allows Murati and her team to pursue a long-term vision that does not sacrifice human involvement for speed. While other labs rush to deploy fully autonomous systems, Thinking Machines is investing in infrastructure that keeps people central.
The Contrast with Big AI Labs
OpenAI, Anthropic, and Google are all developing large models that can perform increasingly complex tasks with minimal human input. Anthropic has discussed the concept of “constitutional AI” where models self-govern according to a set of principles. Google’s DeepMind has pursued systems that learn from their own simulations. These approaches minimize reliance on human feedback once training is complete. Murati’s approach explicitly rejects that trajectory. She believes that even superintelligent machines should remain accountable to human judgment.
The Bigger Picture for Society
If the only AI systems that reach superintelligence are those designed to operate without humans, society faces a future where critical decisions about healthcare, finance, education, and governance are made by machines that no one truly oversees. Murati’s bet is that keeping humans in the loop is not just ethical but practical. Systems that collaborate with people can earn trust, adapt to changing social norms, and incorporate diverse perspectives. That is a future worth building toward.
What Makes This Approach Different from the Rest of the Industry
Other labs, including Humans&, also aim to develop AI systems that prioritize human collaboration. Some prominent economists have called for researchers to build technology focused on empowerment rather than replacement. But Thinking Machines Lab is unique in its combination of massive funding, a founding team with deep expertise in multimodal AI, and a product strategy that starts with fine-tuning tools before releasing general interaction models.
The Role of Multimodal Understanding
Alexander Kirillov’s expertise in multimodal AI—models that handle audio, video, and text simultaneously—is central to the lab’s approach. By training models to process all these signals together, Thinking Machines builds AI that can read a room, so to speak. It sees your expression, hears your tone, and reads your words in context. That holistic perception is what makes the human-in-the-loop model feel natural rather than forced. It also sets a high bar for competitors who treat voice input as just another text channel.
The Long Road Ahead
So far, the interaction models have not been released publicly. The demonstrations prove the concept works in controlled settings, but real-world deployment will reveal new challenges. Tinker, the fine-tuning API, is available today, and it gives early adopters a way to start customizing models right now. Murati has framed these early steps as the first bet on human collaboration. The full vision—superintelligence guided by ongoing human involvement—will take years to realize. But the direction is clear, and it offers a compelling alternative to the automation-first mindset dominating the industry.
A Final Thought on the Human Future of AI
Mira Murati’s work at Thinking Machines Lab matters because it challenges a default assumption: that smarter machines must mean less human involvement. By designing interaction models that understand how people actually communicate, by opening fine-tuning tools like Tinker, and by insisting that human judgment remain central even at superintelligent levels, she is building a different kind of future. Whether you are a developer worried about your career, a small business owner seeking tools that fit your workflow, or a teacher trying to keep education personal, the idea of staying in the loop offers reassurance. The most powerful AI may still be ahead of us, but how we shape its development is a choice we get to make today.
