The mission of Thinking Machines is to build AI that extends human will and judgment. Most AI today is trained centrally and frozen, not shaped by the people it serves. We pursue strong, customizable models, tools for fine-tuning, interfaces for live collaboration, and open research to keep AI diverse and distributed.

Productive knowledge is tacit, local, and held by those doing the work, as Michael Polanyi and Friedrich Hayek observed. Centralized AI cannot aggregate this dispersed knowledge. Toyota's 2014 return of craftsmen to the line, led by Mitsuru Kawai, shows that mastery requires humans to teach machines. We build tools for organizations to cultivate unique knowledge, not replace it.

Human participation is a technical challenge: narrow text-box interaction limits feedback. We bet on native multimodal interaction models that scale with intelligence. Current benchmarks like METR's task-completion time horizons measure solo AI capability, not human-AI collaboration. We optimize for AI that makes users stronger long-term, aligning incentives with customization and ownership.