Show HN: 28MB local agent solves "Gravity Othello" where GPT-5.2 fails (github.com)

A new AI project named Project A.L.I.C.E has developed a compact 28MB local agent capable of successfully tackling the "Gravity Othello" game, a modified version of Othello designed to challenge AI systems' adaptability to rule changes. This initiative reveals critical insights into an AI's ability to detect anomalies and adapt strategies in response to unexpected modifications in dynamic environments. The "Context Drift Detection Test" evaluates how AI manages cognitive flexibility through scenarios like the introduction of phantom stones and gravity physics, which force the agent to discern real game elements from illusions and adjust its tactics accordingly. The significance of this project lies in its approach to traditional AI benchmarks, which typically assess performance under static conditions. By creating a framework for testing rapid rule changes, this research aims to better prepare AI systems for the complexities of real-world applications, such as fluctuating software APIs and evolving user demands. Technical insights from the tests highlight that while some models detected anomalies promptly and adapted strategies effectively, others struggled, underscoring the diverse capabilities of current AI in handling dynamic challenges. This work contributes to the larger discourse on benchmarking AI adaptability and resilience, essential for advancements in artificial general intelligence.