1. Overview

In the rapidly evolving landscape of artificial intelligence, few names carry as much weight as Mira Murati. As the former Chief Technology Officer of OpenAI, she was a central figure in the development of ChatGPT, DALL-E, and the GPT-4 series. Her departure from OpenAI in late 2024 sent shockwaves through the industry, sparking intense speculation about her next move. Today, on May 12, 2026, we are seeing the full realization of her vision with her new startup, Thinking Machines.

While many expected Murati to build another foundational model to compete directly with GPT-5 or Gemini, Thinking Machines has taken a more nuanced and potentially more disruptive path. The company’s core focus is not just on the intelligence of the model, but on the Interaction Model—the fundamental way in which humans and AI exchange information, intent, and feedback.

For the past several years, the "chatbox" has been the dominant paradigm for AI. We type a prompt, and the AI generates a response. Murati argues that this is a bottleneck. Thinking Machines aims to shatter this paradigm, moving toward a world where AI is not just a tool you talk to, but a cognitive partner that works alongside you in a fluid, multi-modal, and context-aware environment. As we explore in our AI Watch introductory post, the "now" of AI is defined by how these technologies integrate into our daily workflows, and Thinking Machines is at the forefront of this integration.

2. Details

The Shift from Chat to Interaction

The central thesis of Thinking Machines, as articulated in their technical blog, is that the current "turn-based" interaction of LLMs is a vestige of old computing habits. In their view, a true Interaction Model consists of three primary pillars: Fluidity, Intentionality, and Shared Context.

1. Fluidity and Real-Time Feedback

Current AI interactions are high-latency and discrete. You send a prompt, wait, and receive an output. Thinking Machines is developing interfaces where the AI begins to assist the moment you start a task. This requires massive improvements in LLM inference-compute optimization to ensure that the AI can process ambient data (voice, screen activity, and cursor movement) without lag.

Imagine a software engineer working on a complex refactoring project. Instead of asking the AI to "rewrite this function," the Thinking Machines interface observes the engineer's navigation through the codebase. It anticipates the need for specific boilerplate or security checks, offering "ghost-writing" suggestions that evolve as the engineer types. This is the evolution of the engineer from a "coder" to an "orchestrator," a theme we've tracked in our analysis of AI agent-era software development.

2. Intentionality and Proactive Agency

Most AI today is reactive. Thinking Machines is building models that understand intent rather than just instructions. By leveraging high-reasoning capabilities—similar to the breakthroughs seen in Gemini 3.1 Pro—their system can infer the ultimate goal of a user. If a user is preparing a financial report, the AI doesn't just format a table; it proactively looks for discrepancies in the underlying data and suggests visualizations that highlight the most critical trends.

3. Shared Context and Persistence

One of the biggest frustrations with current AI is the "amnesia" between sessions. Thinking Machines utilizes a persistent memory layer that builds a long-term map of a user's preferences, projects, and communication style. This isn't just about "saving history"; it's about building a shared context that allows for shorthand communication. The AI learns that when you say "the usual report," you mean a specific layout, data source, and tone.

Technical Infrastructure and Standardization

Thinking Machines isn't building in a vacuum. To achieve this level of interaction, they are leveraging emerging industry standards. The adoption of the Model Context Protocol (MCP), which has been accelerated by AWS's integration into SageMaker, allows Thinking Machines to connect their interaction layer to a vast array of enterprise data sources seamlessly. By using MCP, their "Interaction Model" can pull real-time data from CRM systems, code repositories, and communication tools like Slack or Teams, making the AI's assistance truly grounded in the user's actual environment.

The "Thinking Machines" Interface

Reports from early beta testers suggest that the Thinking Machines interface looks less like a chat window and more like a "canvas." It is a spatial environment where users can drag and drop ideas, and the AI populates the space with relevant data, drafts, and connections. It utilizes a "Human-in-the-Loop" design where the AI presents multiple paths of action, and the user steers the direction with minimal friction.

3. Discussion (Pros/Cons)

Pros

  • Cognitive Load Reduction: By moving away from precise prompting, Thinking Machines lowers the barrier to entry for complex tasks. Users can focus on high-level strategy while the AI handles the execution and context-switching.
  • Enhanced Productivity: The proactive nature of the interaction model eliminates the "blank page" problem. The AI provides a starting point based on the user's intent, significantly speeding up the creative and analytical process.
  • Personalization: Unlike generic LLMs, this model evolves with the user. The more you interact with it, the more it aligns with your specific workflow, becoming a truly bespoke digital assistant.
  • Integration with High-Reasoning Models: By focusing on the interaction layer, Thinking Machines can remain model-agnostic, utilizing the best-in-class reasoning models (like GPT-5 or Gemini 3.1) as the "engine" while they provide the "steering wheel."

Cons and Challenges

  • Privacy and Security: A model that monitors your intent and maintains a persistent memory of your work requires deep access to personal and corporate data. Thinking Machines will face significant scrutiny regarding how this data is encrypted and whether it is used for further model training.
  • Computational Cost: Real-time, fluid interaction requires constant inference. As discussed in our look at inference-compute optimization, the cost of maintaining this "always-on" intelligence could be prohibitive for individual users or small businesses.
  • The "Uncanny Valley" of Agency: If an AI becomes too proactive, it risks becoming intrusive or making incorrect assumptions that frustrate the user. Finding the balance between "helpful assistant" and "presumptive meddler" is a difficult UX challenge.
  • Market Competition: Tech giants like Apple (with Apple Intelligence) and Microsoft (with Copilot) are also moving toward system-wide integration. Thinking Machines must prove that their specialized interaction model is significantly superior to the "good enough" solutions built into the operating systems.

4. Conclusion

Mira Murati’s Thinking Machines represents a critical pivot in the AI industry. For the last three years, the race has been about scale—more parameters, more data, more compute. However, as we reach the limits of what raw scale can achieve, the focus is shifting toward utility and usability.

The "Interaction Model" is the next great frontier. By focusing on how we live and work with AI, rather than just how we query it, Murati is attempting to bridge the gap between AI as a novelty and AI as a fundamental cognitive utility. If successful, Thinking Machines won't just be another AI company; it will be the architect of the new human-computer interface for the 21st century.

At AI Watch, we believe that 2026 will be remembered as the year the "Chat Era" ended and the "Interaction Era" began. Whether Thinking Machines becomes the dominant player remains to be seen, but they have undeniably set the agenda for the future of AI development. Stay tuned as we continue to track the deployment of these models and their impact on the global workforce.

References