Mira Murati, known for her track record in artificial intelligence, has presented alongside Thinking Machines Lab a vision that challenges the current paradigm of centralized and frozen models. Their proposal: a distributed, customizable AI shaped by users, where the tacit and local knowledge of each organization can be embedded directly into model weights. This approach not only changes the technical architecture but also redefines who controls and aligns artificial intelligence.
The core idea is that most valuable knowledge in companies and institutions is tacit, constantly updated through feedback, and cannot be extracted or captured in a static database. A chef perfecting a recipe, a doctor adapting protocols, or a lawyer refining their style—all possess know-how that traditional AI, trained once and then frozen, cannot reflect. Thinking Machines Lab argues that AI must be distributed to manage that distributed knowledge, not to extract or replace it.
To achieve this, the lab has defined four technical directions. First, train multimodal models capable of continuous interaction and customization. Second, develop tools like Tinker that allow technical teams to fine-tune model weights using techniques like LoRA (Low-Rank Adaptation). Third, build interfaces that widen the human-machine communication channel, moving from text turns to micro-turns of about 200 milliseconds that integrate audio, video, and text in real time. Fourth, publish open research so more engineers understand how models are built.
This paradigm shift has profound implications for companies seeking to adopt AI with sovereignty. Instead of renting a fixed model that can barely be modified through superficial prompts, organizations can own exportable weight adapters, trained on their own data and reflecting their values and processes. For example, a hospital could fine-tune a model with its internal protocols while keeping data and weights on-premises. A law firm could adapt the model to its style and update it whenever internal guidelines change. A support team could correct the model mid-interaction, guiding it with continuous micro-turns.
From a technical standpoint, Thinking Machines Lab argues that alignment should not be a single centralized capture point. If all power to decide what values a model incorporates resides in one lab, it creates a risk of capture or uniform bias. Therefore, they propose that values be encoded in weights, not prompts. Prompts change the surface layer of behavior, but the model's deep habits remain fixed. In contrast, fine-tuned weights carry the imprint of the organization that trained them.
This vision directly connects to the need for adequate technical infrastructure. Companies wishing to implement this type of distributed AI require custom software development platforms that integrate open models, secure storage systems, and cloud orchestration capabilities. This is where companies like Q2BSTUDIO bring their expertise, offering services ranging from language model customization to integration with cloud environments like AWS or Azure, ensuring that data and adapters remain under client control.
Furthermore, security becomes a fundamental pillar. By owning the model weights and training data, the organization must protect itself against unauthorized access and potential data leaks. Cybersecurity services are essential to audit the fine-tuning pipeline, secure inference environments, and ensure that personalized AI does not expose vulnerabilities. Similarly, business intelligence (BI) and tools like Power BI benefit from models that understand the company's own language, enabling contextualized reports and analysis without relying on generic assistants.
AI agents, one of the most promising trends, also fit into this philosophy. Instead of autonomous agents acting without supervision, Thinking Machines Lab advocates for collaborative agents that interact in micro-turns with humans, correcting course in real time. This enables complex tasks where human and artificial intelligence complement each other, such as legal document review or image-assisted diagnosis. The combination of local fine-tuning with multimodal interfaces opens the door to systems that learn and adapt to the specific context of each team.
In summary, the proposal from Mira Murati and Thinking Machines Lab represents a shift toward a more democratic and useful AI. It is not about building larger models, but about distributing them so that each organization can cultivate its own knowledge without losing control. To make this vision practical, an ecosystem of development tools, secure cloud platforms, and specialized consulting is needed. Companies like Q2BSTUDIO are ready to accompany organizations on this path, offering custom software solutions, cloud integration, and data protection, so that personalized AI becomes an operational reality and not just a theoretical concept.
