As traditional AI benchmarking techniques struggle to keep pace with advancements in technology, developers are increasingly exploring more creative methods to assess the capabilities of generative AI models. One such innovative approach comes from the realm of gaming, specifically the popular sandbox-building game, Minecraft. The newly launched website, Minecraft Benchmark (or MC-Bench), serves as a platform where AI models compete against one another in head-to-head challenges, generating unique creations based on user prompts.
MC-Bench invites users to vote on which AI model has executed a prompt more effectively, but the twist is that voters only discover which AI was responsible for each creation after they cast their votes. This element of surprise adds an engaging layer to the process. According to Adi Singh, a 12th grader and the founder of MC-Bench, the appeal of Minecraft lies not only in its gameplay but also in its widespread recognition. As the best-selling video game of all time, it provides a familiar framework for users to evaluate AI creations, even if they haven't played the game extensively.
“Minecraft allows people to see the progress of AI development much more easily,” Singh remarked in an interview with TechCrunch. “People are used to Minecraft, used to the look and the vibe.” This familiarity contributes to the platform's effectiveness in showcasing the advancements in AI technology.
Currently, MC-Bench boasts eight volunteer contributors who help maintain and develop the platform. Notably, major players in the AI industry, including Anthropic, Google, OpenAI, and Alibaba, have provided support for the project by allowing the use of their products to run benchmark prompts. However, it’s essential to note that these companies are not officially affiliated with MC-Bench.
“At the moment, we are focusing on simple builds to reflect on our progress since the GPT-3 era,” Singh explained. “We envision expanding to longer-form plans and goal-oriented tasks in the future.” He further suggested that using games like Minecraft for benchmarking could provide a safer and more controllable environment for testing AI reasoning skills.
In addition to Minecraft, other games such as Pokémon Red, Street Fighter, and Pictionary have been utilized as experimental benchmarks for AI. This trend arises from the inherent challenges in benchmarking AI models. Traditional standardized evaluations often grant AI a home-field advantage, allowing them to excel in narrow problem-solving areas, particularly those involving rote memorization or basic extrapolation.
For instance, while OpenAI’s GPT-4 may score in the 88th percentile on the LSAT, it struggles with even the simplest tasks, such as counting letters in a word. Similarly, Anthropic’s Claude 3.7 Sonnet demonstrated 62.3% accuracy on a standardized software engineering benchmark but performs poorly in games like Pokémon compared to most five-year-olds.
MC-Bench, while fundamentally a programming benchmark, asks AI models to write code to create prompted builds like “Frosty the Snowman” or “a charming tropical beach hut on a pristine sandy shore.” This approach makes it easier for users to visually assess which creation is superior, broadening the project’s appeal and enhancing its data collection capabilities. The ability to evaluate visual outputs rather than delving into code enables a wider audience to participate, thereby enriching the benchmarking process.
However, the significance of these scores in terms of AI usefulness remains a topic of debate. Singh believes that the scores serve as a valuable indicator of AI performance. “The current leaderboard closely reflects my own experiences with these models, which differs significantly from traditional text benchmarks,” he stated. “MC-Bench could potentially provide useful insights for companies looking to gauge their progress in AI development.”