Thinking Machines Lab published a report arguing that most AI today is trained in a handful of places, then frozen, which excludes the people a model serves, and instead proposes AI that is distributed, customizable, and shaped by its users. The lab names four technical directions: training strong models with multimodal interaction and customizability; building tools that let people fine-tune and train model weights themselves; developing interfaces that widen the human-to-machine communication channel; and publishing research so more engineers understand how models are made. The report cites Michael Polanyi and Friedrich Hayek to support a claim that much know-how is tacit, local, and updated constantly, arguing AI must be distributed to use distributed knowledge, with chess and math as stated exceptions. It reframes two technical bottlenecks: the communication channel, which the lab’s interaction models address by taking in audio, video, and text continuously using roughly 200ms micro-turns, and evaluation, arguing benchmarks like METR’s measure how long a model works alone, missing what people and machines accomplish together. The report argues values should be encoded in model weights, not prompts, and introduces Tinker, which fine-tunes open-weights models such as Llama and Qwen using LoRA, exposing low-level primitives and allowing export of portable adapter weights.