Mira Murati’s Thinking Machines Lab Makes The Technical Case For Human-Centered AI Built On Customizable Model Weights

Thinking Machines Lab published a report arguing that AI systems should be distributed, customizable, and shaped by their users rather than trained in a handful of places and then frozen. The lab proposes four technical directions: training models with multimodal interaction and customizability, building tools for fine-tuning model weights, developing interfaces for better human-machine communication, and publishing research so more engineers understand how models work. The report frames this approach around the idea that much real-world knowledge is tacit and local, constantly updated through feedback, and therefore cannot be extracted into databases or centralized systems. The lab introduced Tinker, a tool that lets organizations fine-tune open-weights models and encode their own values into portable adapter weights, keeping both data and control in-house rather than renting fixed models from central providers.