The authors note that while GPT-4 does not match the performance of traditional, reinforcement learning-based models, it does not require any training, relying instead on its reasoning and observational skills. They suggest that further research could improve the LLM's gaming abilities. The authors hope their work will expand the possibilities for intelligent, LLM-based agents in video games and conclude by discussing the ethical implications of their research.
Key takeaways:
- The large language model GPT-4 can run and play the 1993 first-person shooter game Doom with only a few instructions and a textual description of the game state.
- GPT-4 can perform basic game functions like manipulating doors, combating enemies, and pathing.
- More complex prompting strategies involving multiple model calls provide better results in the game.
- Despite not being as proficient as its reinforcement learning-based counterparts, GPT-4 required no training, relying on its own reasoning and observational capabilities.