If you thought language models were built only for writing emails, pondering the meaning of life, and drafting polite rejections, hold on tight. ChatGPT unexpectedly took second place in the international Kerbal Space Program Differential Game Challenge for spacecraft control. That is especially impressive given that the model had never learned to fly and had never aspired to be a “space pilot.”
In the competition, the AI had to handle “space” tasks in a simulator, from intercepting a satellite to escaping pursuit. The researchers took an unconventional approach: they passed spacecraft state parameters in text form, and ChatGPT analyzed them and suggested actions. As a result, it lost only to a specialized algorithm tuned for motion and orbital physics. What is especially amusing is that the experiments were carried out on an older version of the model.
The number of satellites and spacecraft in orbit is growing, and controlling them from Earth in real time is becoming more difficult. Autonomous systems are becoming not just desirable, but necessary. And now it turns out that the AI that helped you write a résumé yesterday can confidently command a virtual spacecraft today. The main thing is that it does not glitch in space and start insisting, with total confidence, that there are Martians hiding on board.
