Two Ways to Simulate a Planet: EarthOS and NationalOS
Simulation at two scales
VIDRAFT runs two large-scale simulation efforts that sit at opposite ends of the map. EarthOS models the physical planet from observation data. NationalOS models a human population as a swarm of AI agents. Both are bets that a small team can build simulation infrastructure usually associated with much larger labs -- method over scale.
EarthOS: an Earth-observation foundation model
EarthOS is an Earth-observation foundation model that absorbs the OlmoEarth approach: pretrain on satellite and remote-sensing data to produce embeddings that transfer to downstream geospatial tasks such as land cover, change detection, crop and vegetation analysis, and disaster response.
The design goal is a shared representation. Instead of training a bespoke model per sensor or per task, you learn one backbone over multi-modal Earth observation and adapt it with lightweight heads. That is the same transfer-learning logic that made language and vision foundation models useful, now applied to pixels that arrive from orbit.
License matters here. EarthOS ships under terms that exclude surveillance and policing use. Earth-observation models are dual-use by nature, and we would rather constrain the license up front than pretend the risk is not there.
NationalOS: a society twin
NationalOS is a different kind of simulation -- a society twin. Instead of physical pixels it models a population as millions of AI-agent NPCs, each an individual with its own attributes and behavior. The Korea build targets roughly 52 million agents, on the order of one per person in the country. A USA variant applies the same approach to a different population.
The point is emergent, population-level behavior. When you instantiate millions of interacting agents you can pose what-if questions about policy, information flow, and collective response that aggregate statistical models handle poorly. It is agent-based modeling pushed to national scale.
| System | Models | Scale | Primary use |
|---|---|---|---|
| EarthOS | The physical planet from EO data | Foundation-model embeddings | Geospatial downstream tasks |
| NationalOS | A human population | ~52M agents (Korea build) | Policy and society simulation |
Why build both
The two share a worldview: complex systems are easier to reason about when you have a learned, queryable model of them rather than a spreadsheet. EarthOS gives you a substrate for the environment. NationalOS gives you a substrate for the people who live in it. In principle they compose -- environmental signals feeding a population model -- though that integration is future work, not a shipped feature.
Running both on a ~24-GPU footprint forces discipline. We cannot brute-force these problems, so the engineering leans on transfer learning, efficient agent representations, and reuse of pretrained backbones rather than training everything from scratch.
Honest scope
Both systems are research-stage. EarthOS is a foundation-model effort following a published approach, not a certified operational forecasting service; downstream accuracy depends entirely on the task and the fine-tuning data you bring. NationalOS is a simulation of behavior, not a prediction oracle -- agent-based models illustrate dynamics and stress-test assumptions, but their outputs are only as good as the behavioral priors encoded in the agents, and they should never be read as literal forecasts of what specific real people will do. Agent counts describe modeling scale, not validated demographic fidelity. We are sharing these as active builds and will report concrete evaluations as they mature.
More: https://vidraft.net
Top comments (0)