Truly Intelligent AI Could Play by the Rules, No Matter How Strange
To build safe yet powerful AI models, researchers are exploring a new approach – testing their ability to play games on the fly. Games, with their diverse rules and strategies, have always been a significant part of human culture. Navigating and adhering to these rules is a fundamental aspect of human intelligence.
The ultimate goal of artificial intelligence research is to achieve artificial general intelligence, which closely mimics human cognitive abilities. One crucial aspect of this is developing AI systems that can interpret, adapt to, and rigorously follow the rules set for them.
Inspired by the work in general game playing (GGP), a new test called the Gardner test is being proposed. In this test, AI models are presented with the rules of a game at the start and are expected to play by those rules without any human intervention. This challenges AI systems to adapt to new rule sets on the fly.
The Gardner test builds upon the advancements in large language models (LLMs) like ChatGPT, enabling AIs to understand and follow game rules expressed in natural language. This test aims to evaluate AI systems on their ability to play a variety of strategy games without prior knowledge of the rules.
Unlike current game-playing AI models such as AlphaZero, which rely on pre-defined rules for training, AI systems that pass the Gardner test would be capable of adapting to new rules without prior data. They would exhibit precision in following any novel rule set, showcasing a higher level of flexibility and adaptability.
By developing AI systems that can rigorously follow rules, we can ensure their safe and reliable operation in various real-world scenarios. From national security to finance, precise rule adherence is crucial to prevent catastrophic errors that could arise from misinterpretation or misapplication of rules.
Ultimately, testing AI systems on their ability to play games with novel rule sets could lead to the creation of more human-like machine intelligences. Just as humans can generalize their skills across different games, AI systems capable of playing any game and following any rule set with precision could pave the way for more versatile and adaptable artificial intelligences.
In conclusion, the path to building powerful yet safe AI lies in testing their ability to navigate and adhere to rules in dynamic and challenging game-playing scenarios. This approach could unlock new possibilities for creating AI systems that are not only intelligent but also capable of following rules with utmost precision and reliability.
