We’d rather build the compiler
Tom Jeong
August 2026
Waymo built the world. Hand-authored road networks, scripted pedestrians, edge cases specified one at a time. An enormous engineering effort to approximate something that already existed. Tesla instrumented a fleet, shipped cameras on every car, and mined what actually happened.
Both produced working systems. But they scale differently: one with engineer headcount, the other with cars on the road.
RL environments are being built the Waymo way. Engineers clone LinkedIn, clone Salesforce, clone realistic coding problems. They spend thousands of hours rebuilding what already exists.
We want to be Tesla. But Tesla had to sell cars to get its data: factories, dealers, regulators, recalls. Everyone pays a version of that tax. In coding, it’s running a product business. Cursor accumulated the best coding data in the industry and trained its own models on it. Design Arena went the other way and built the product so there’s free model comparison, and every vote is a label.
So the question is whether there’s work that produces instrumented trajectories, is worth doing on its own terms, and requires selling nothing to anyone for the data to exist.
ML research is the best fit we’ve found. The training run happens whether or not a customer exists, and it logs itself: the loss curve is the reward signal, the ablation is a labeled comparison, the failed run is a negative example. The output has value independently: a result, an expert-reviewed paper, a capability. No fleet to build. No platform or product to sell to get the data.
This extends to autoresearch, where an agent runs the experiments and produces trajectories in the format the environment needs, at machine volume.
Every environment company is manufacturing what a world model can compile. We’d rather build the compiler: the metric from the researcher, the task distribution from the model.