Here’s what Mira Murati’s AI company is up to

DeepTrendLab's Take on Here’s what Mira Murati’s AI company is up to

Mira Murati's Thinking Machines has articulated a fundamental critique of how modern AI systems interact with humans: they operate as sequential bottlenecks rather than collaborative partners. The company's core observation centers on the single-threaded nature of current large language models—systems that pause until input is complete, then generate uninterrupted output while remaining blind to the user's real-time context, reaction, or evolving intent. Rather than incremental improvements to existing architectures, Thinking Machines is positioning real-time, multimodal interactivity as a different category of AI entirely, one that can adapt continuously to human feedback across text, voice, video, and other modalities simultaneously.

The limitations Thinking Machines identifies are not new concerns, but they've become increasingly visible as models have grown more capable. The current turn-based interaction paradigm emerged naturally from transformer architecture—designed for predicting sequences from static input. This design has driven how every major AI application works today, from ChatGPT to Claude to Gemini. Users adapt to rigid constraints: they must fully formulate requests before sending, tolerate generation lag, and manually restart or rephrase when the model misinterprets their intent. The bandwidth constraint Murati identifies is real—humans cannot effectively communicate nuance, uncertainty, or dynamic preference shifts through fixed prompts, and models cannot adjust course mid-generation based on subtle signals they never receive.

If Thinking Machines can credibly deliver on this vision, the implications extend beyond user experience. Real-time interactivity would collapse the distance between what humans can express and what AI systems can understand, fundamentally expanding the types of work these systems could handle. Research collaboration, complex problem-solving, creative work, and technical debugging all suffer from the current paradigm—they require feedback loops and mid-stream adjustments that humans naturally expect but current systems cannot support. A shift here wouldn't just be a feature; it would represent a different category of AI application, enabling use cases that current systems structurally cannot serve well. The technical challenges are formidable, but the business and research incentives are equally massive.

The immediate impact would land hardest on developers and researchers building on top of AI systems. New interaction patterns would require rethinking how applications are architected—moving from static-input application design to continuous-signal processing. Enterprises deploying AI for knowledge work would suddenly have access to a fundamentally different tool, one that could adapt to domain-specific workflows rather than forcing workflows into a chatbot mold. Researchers exploring AI-human collaboration would gain new dimensions to explore. But the ripple effect extends to consumers too: applications built on this foundation would fundamentally change what people expect from AI assistants, creating pressure on incumbents to match capabilities or lose relevance.

This framing also represents a subtle but significant departure from the current competitive narrative, which has centered on model scale, reasoning capability, and cost efficiency. OpenAI, Anthropic, Google, and Meta are locked in a capability race along largely predictable dimensions—more parameters, better training data, faster inference. Thinking Machines is arguing that the game itself is misconfigured, that raw capability matters far less than interaction design. If the market agrees, it reframes what "better AI" actually means. A smaller model capable of true real-time collaboration could outperform a larger one constrained by sequential architecture. This could reshape how AI companies compete and which technical approaches attract funding and talent.

The key questions now are technical and temporal. Can true real-time interactivity across modalities actually be built without prohibitive latency or computational cost? What does the product actually look like in practice—how do humans and AI systems signal to each other continuously? And perhaps most importantly, when? Thinking Machines has articulated the problem sharply, but execution at scale remains unproven. If solved, this becomes a defining inflection point in how AI systems work. If not, it becomes a cautionary tale about the limits of architecture-level change against entrenched paradigms. Watch whether the company can move from critique to demonstration to deployment—and whether incumbents respond by attempting their own versions of real-time interaction or dismiss it as a niche concern.