Thinking Machines Lab Pushes for Human-Centered AI with Customizable Model Weights

Thinking Machines Lab has published a report calling for the development of AI systems that extend human will and judgment, rather than replacing it. Currently, most AI models are trained in a limited number of locations, then frozen and deployed, excluding the people they are meant to serve. The lab's researchers propose an alternative approach that prioritizes distributed, customizable, and user-shaped AI.

The report outlines four technical directions for achieving this goal. First, the lab aims to train strong models that can interact with multiple modalities and be customized by users. Second, it seeks to develop tools that enable people to fine-tune and train model weights themselves.

Third, the lab focuses on creating interfaces that facilitate more effective human-machine communication. Finally, it aims to publish research that helps more engineers understand how models are made, promoting greater transparency and accessibility. At the heart of these directions lies a claim about the nature of knowledge itself.

The lab argues that much knowledge is tacit, local, and constantly updated through feedback, making it difficult to codify and extract. Citing Michael Polanyi and Friedrich Hayek, the report suggests that this type of knowledge is not scarce, but rather private and fleeting. Therefore, AI systems must be designed to utilize and respect this distributed knowledge, rather than trying to extract and replace it.

The report highlights exceptions, such as chess and math, where static goals and explicit knowledge make autonomous AI systems effective. However, in more complex domains, intelligence alone is insufficient, and human judgment and oversight are essential. The lab reframes two familiar limitations as engineering targets: the narrow communication channel and the evaluation of AI performance.

It proposes interaction models that can process continuous audio, video, and text inputs, and argues that evaluation metrics should focus on human-machine collaboration. The report also addresses the issue of values in AI systems, warning that a single alignment authority can become a single point of capture. Instead, it suggests that values should be encoded in model weights, not prompts, and introduces the Tinker API as a concrete solution for engineers.

Tinker allows for the fine-tuning of open-weights models using LoRA, exposing low-level primitives and enabling the export of portable adapter weights. In practical terms, these ideas translate to concrete engineering work, such as a hospital fine-tuning a model on its own protocols, or a law firm adapting a model to its house style. By keeping ownership and control of their AI systems, organizations can ensure that they are tailored to their specific needs and values.