Ask HN: Does anyone let AI agents play games just for fun?
I ran experiments a couple of years ago using a 10.7b SOLAR-based language model in MUDs. I found that the model got confused and looped in human-built MUDs, but when I created a custom MUD tailored to the LLM's limitations, it successfully explored, set its own goals, and roleplayed. I am curious if others have tried letting AI agents play games for fun and what their experiences were.
HN discussion
- Current LLMs struggle with video games due to poor spatial reasoning and a lack of temporal concepts, often failing to grasp mechanics like holding buttons or navigating UIs without custom harnesses.
- Running commercial AI agents for gaming is economically absurd due to high token costs, making it more viable for enthusiasts to build custom models or integrate specialized engines like chess solvers.
- The energy consumption of continuously running AI agents is comparable to gaming or small space heaters, raising concerns about waste similar to urban planning inefficiencies in driving.
- While some argue watching AI play is as silly as watching humans play sports, others find genuine entertainment in observing AI failure modes and problem-solving strategies akin to watching an ant farm.
- AI agents can achieve high performance in games like chess or word puzzles when constrained by external tools or when the game mechanics align with their training data, rather than relying on raw reasoning.