Roam
Space was never written down, so Roam is writing it down: a live, queryable record of the physical world.
A studio building the machine-readable record of physical space.
Why now
World models, sensors and interfaces all keep improving. The friction is still in turning the world itself into something software can read without flattening it into fiction.
Build the corpus. Capture the physical world continuously, resolve it into persistent structure rather than disposable frames, keep provenance attached, and expose it as something a model can query. Perception is one layer of that; the harder half is representation, persistence and retrieval.
Earlier work sits behind Roam as context.
World models, robotics, autonomy, and simulation all depend on better physical-world inputs. The bottleneck now is trustworthy data tied to the real world.
Roam sits in that gap: capture, structure, and interfaces that help machines reason about real places with less hallucination and less operational fog.
Earlier work in crypto, applied AI, and secure systems produced a bias toward systems you can audit.
That bias carries forward into Roam as a product principle: keeping provenance attached to every capture.
Then
Roam continues a long-running pattern: I pick infrastructure shifts early. Bitcoin in January 2013, a language model of my own in 2022, and now the physical world.
Now
Roam is the current expression of it: hardware, data infrastructure, and model work in the same system, aimed at agents, robots, spatial interfaces and logistics.
What stayed with me
The system has to keep history, provenance, and constraints visible.
Beneath it
The tokenizer, the pretraining corpus, the transformer, the training runs, the evals, and the serving stack were all built here.
Property, bitcoin, and what happens when software has to live with regulation instead of talking around it.