These pages show what happens to a small AI language model's output when simulated cosmic radiation flips bits in its memory during inference — the kind of thing an AI chip on a satellite has to survive. The radiation environment is crudely calibrated to the median altitude (~1250 km) of a 500–2000 km low-Earth-orbit constellation.
One model, one solar-storm run: the side-by-side output diff, the degradation-over-time timeline, and a plain-English verdict.
The same model under different space environments (LEO, polar, the South Atlantic Anomaly, a solar storm) compared side by side.
Source & docs: github.com/davedx/cosmicgpt.
Reports are generated by the cosmicgpt CLI and are fully self-contained.
The demo model is tiny and untrained, so the baseline text is gibberish — what matters is
how much the radiation changed it.