- Published on
Mira Murati’s Thinking Machines Lab Makes The Technical Case For Human-Centered AI Built On Customizable Model Weights
- Authors
- Name
- DeepHorus
Mira Murati’s Thinking Machines Lab has published a report advocating for a fundamental shift in artificial intelligence development toward human-centered systems built on customizable model weights. The report argues that most current AI models are trained centrally and then frozen, which excludes the people they are meant to serve. Instead, the lab envisions AI that is distributed, customizable, and shaped by its users to extend human will and judgment.
The lab outlines four technical directions to achieve this vision. First, it aims to train strong models with multimodal interaction and customizability. Second, it builds tools that allow users to fine-tune and train model weights themselves. Third, it develops interfaces that widen the human-to-machine communication channel. Fourth, it publishes research to help engineers understand how models are made. Together, these directions move both knowledge and alignment closer to users.
Underpinning these directions is a claim about the nature of knowledge. The report cites philosophers Michael Polanyi and Friedrich Hayek to argue that much know-how is tacit, local, and constantly updated through feedback, making it impossible to encode in a static database. Therefore, the lab contends, AI must be distributed to leverage distributed knowledge, helping organizations cultivate rather than extract expertise. The only exceptions are closed domains like chess and math, where goals are static and knowledge is explicit.
The report also identifies two technical bottlenecks as engineering targets. The first is the communication channel—currently a small text box with long waits—which the lab’s interaction models address through continuous audio, video, and text input with 200-millisecond micro-turns. The second is evaluation: benchmarks like METR’s measure solo model performance, ignoring what humans and machines achieve together. The lab argues that alignment should be encoded in model weights, not prompts, to avoid a single point of capture.
To operationalize this philosophy, Thinking Machines Lab introduces the Tinker API, which fine-tunes open-weights models such as Llama and Qwen using LoRA (Low-Rank Adaptation). Tinker exposes low-level primitives, allows export of portable adapter weights, and follows a minimal supervised learning loop. This enables engineers to customize models for specific domains while keeping data and adapters in-house.
The report contrasts its distributed, customizable approach with today’s centralized frozen AI paradigm. Practical use cases include a hospital fine-tuning a model on its own protocols or a law firm adapting a model to its house style and retraining it as guidance changes. By placing model weights and fine-tuning tools in users’ hands, Thinking Machines Lab aims to foster AI that truly serves human needs, promising a future where AI systems are continuously shaped by the communities they support.
