The ship’s computer in Elite: Dangerous is officially called the Cockpit Voice Assistant, or COVAS for short. While its voice output is available in several languages and voices, it’s very limited. COVAS announces the most important information, but in a purely static way. “Refueling complete,” “Frame Shift Drive active,” or “Landing gear deployed”—that’s all you get. On top of that, it’s all very one-sided, as COVAS is purely a voice output and cannot accept voice commands. I would have expected more in the game’s year of 3311.
But then there are the players. COVAS: NEXT is supposed to be the better version of the Cockpit Voice Assistant—more dynamic and interactive. To achieve this, the program reads game events from the log files, has them interpreted by ChatGPT, and then outputs ChatGPT’s response via text-to-speech. Additionally, you can use a microphone to send commands to OpenAI’s Large Language Model. It doesn’t just give you a response; it can also control actions in the game—at least, actions that have a keybinding. This sounds fantastic and is one of the more sensible applications for these supposed artificial intelligences.
I say “supposed” because even as an artificial copilot, ChatGPT doesn’t cut it. Although you pre-define the basic parameters for the assistant, writing its backstory and personality, the result is hardly predictable. Even after several iterations of the input parameters, I couldn’t train some annoying habits out of the artificial intelligence. Despite the instruction “You do not ask me questions or give me instructions!”, almost every message from the AI ended with a question. “You have arrived in the XXX system. What should we do now?” or “You have landed. Be careful while exploring the planet.” If there were at least some variety, it wouldn’t be so bad. But the AI has a habit of always saying the same sentences. New system, same message, same question. Since it’s not brief and ends in a question, it’s significantly more annoying than the game’s own dumb COVAS, whose messages are at least short and informative.
Is it my parameters? Is it because the same input from the logs always leads to the same results? I honestly expected more from the big Large Language Models here. Sure, with longer texts, you can often still recognize them by the style and structure of the text. But I didn’t expect it to be so obvious with such short texts. Since I still have some credit left with ChatGPT, I will try to further optimize COVAS: NEXT to meet my standards. If it works, you’ll read about it here.