Planetary Prediction Engine
Google Research's Earth AI system automates geospatial data discovery, curation, feature engineering, model search, evaluation, and reporting from a natural-language prediction request.
How reliably does the system select causal rather than merely predictive signals when data shifts across geography, time, and institutional context?
The Planetary Prediction Engine (PPE) is an experimental system for turning a geospatial question into a trained and evaluated predictive model. It uses language models to orchestrate data discovery, dataset construction, feature selection, model search, overfitting checks, and final reporting.
Google Research reports improvements over baselines across public health, food security, environmental risk, and socioeconomic tasks. In one outbreak nowcasting example, the system combined epidemiological signals with geospatial embeddings to identify new disease hotspots. These are research demonstrations, not evidence that the system can be deployed without domain review.
Mechanism
The design keeps major stages modular and passes data artifacts through opaque handles instead of serializing everything into the model context. That reduces context pressure and makes the pipeline easier to inspect stage by stage.
ASI relevance
PPE is a concrete example of research automation beyond manuscript generation. It closes a bounded loop from question to data to model to evaluation. Its strength comes from the existence of measurable prediction targets; the open problem is whether its chosen signals remain valid under distribution shift and real decision pressure.