Thinking Machines Lab challenges centralised AI with human-centric distributed model
In a new blog post, Thinking Machines Lab outlines a technical and philosophical shift away from single-locus AI alignment, citing historical economic theory and industrial precedents to support a future where artificial intelligence extends rather than replaces human judgment.
Thinking Machines Lab has published a detailed manifesto titled "The Future Worth Building Is Human," outlining its mission to develop artificial intelligence that extends human will and judgment rather than replacing it. The organisation argues against the prevailing industry trend of centralised, autonomous AI systems, which it claims crowd people out and create a single locus of power. Instead, it proposes a distributed approach where AI tools are tailored to local, tacit knowledge and allow for continuous, live collaboration between humans and machines.
The post draws heavily on the works of Michael Polanyi and Friedrich Hayek to support the view that productive knowledge is inherently local and that central planning fails. Polanyi’s 1966 work, *The Tacit Dimension*, and Hayek’s 1945 essay, *The Use of Knowledge in Society*, are cited to argue that knowledge is often unarticulated, fleeting, and held privately by those who acquire it through their work. Thinking Machines Lab contends that attempting to aggregate this dispersed knowledge for a centralised intelligence ignores the nature of how value is created in complex environments.
To illustrate the practical application of this philosophy, the organisation references Toyota’s 2014 initiative under Mitsuru Kawai, which brought expert craftsmen back to the production line to teach machines. This example underscores the lab’s belief that the production of knowledge and the application of intelligence lift each other, rather than acting as substitutes. The lab aims to develop models that are diverse, owned by users, and shaped by the specific values and goals of the organisations that use them, rather than offering a standardised snapshot of knowledge.
Thinking Machines Lab also criticises current industry practices, such as using previous flagship models to generate training data for new ones, describing this as a loop that suppresses creativity and diversity. The organisation argues that evaluation metrics like METR’s Task-Completion Time Horizons measure only autonomous capability, not the collaborative potential of human-AI teams. It advocates for a long-term bet on live, multimodal interaction models that can handle the richness of human intent, moving beyond the bottleneck of narrow text-based communication channels.
The lab warns that centralised alignment creates a "single locus of power" that can be captured, posing a risk to individual sovereignty and corporate profit. By allowing core model behaviour to change significantly with prompts, a malleable centralised model becomes vulnerable to repeated attacks. Instead, Thinking Machines Lab envisions an ecosystem of AIs raised in different places, disagreeing, competing, and learning from each other, ensuring that the power to shape a model remains with the people making the choices.
