For decades, artificial intelligence has been striving to outperform humans at video games. In May 2023, a paper by researchers at Nvidia announced a new groundbreaking milestone: an AI that exceeded the state-of-the-art performance for playing Minecraft. However, the most astonishing part of this achievement wasn’t just the impressive performance metrics, but the method they used to do so.
This Minecraft specialized AI, named Voyager, was powered by a large language model, in other words, a chatbot. In fact, it used GPT-4, the very same model that powers the popular AI application, ChatGPT.
Imagine a chatbot learning to master the complex dynamic world of Minecraft. The complex training process of Voyager can be broken down using three main tools.
First, automatic curriculum. This is Voyager’s personal guidebook for Minecraft. It allows the AI to keep track of its progress and decide what to do next. For example, if Voyager has a stick and wooden planks in its inventory, the curriculum might suggest building a wooden pickaxe. But if Voyager already has a pickaxe, it might instead suggest building a fence.
The second tool is the skill library. This is Voyager’s memory bank. Every time it learns a new trick in Minecraft, it stores that skill in its library, ready to be used again when needed. For instance, if Voyager learns how to build a boat to cross a body of water, it stores this skill and can use it again the next time it encounters a similar situation.
Finally, the third tool is the iterative prompting mechanism. This is Voyager’s way to do trial and error. It tries different actions, sees what works and what doesn’t, and uses this knowledge to improve its gameplay. For example, if Voyager encounters a creeper, it might try different strategies like running away or fighting. It learns from these experiences and uses this knowledge to handle similar situations in the future.
Voyager’s unique way of learning has led to impressive results. It can find more unique items, travel longer distances, and achieve game goals faster than other AI programs. Even more exciting, Voyager can use the skills it learns in one game and apply them to new games. Imagine applying Voyager’s learning mechanisms to a game such as Legend of Zelda: Breath of the Wild. Voyager’s automatic curriculum could guide its exploration of the game world, suggesting new areas to explore or quests to undertake based on its current progress. Its skill library could store effective strategies for defeating enemies, allowing it to recall them in the future. The iterative prompting could help it learn from its actions and improve gameplay over time.
Voyager’s success in Minecraft has significant implications for the future of AI. Its ability to learn and adapt to different situations suggests that we might be able to build AI systems that can learn in a wide variety of real-world situations. If an AI can be used to both chat and play games, the potential applications for this technology could be vast. In essence, Voyager is a testament to the potential of AI. It’s a computer program that can play and learn in the world of Minecraft like a pro, using technology similar to ChatGPT. As we continue to develop and understand this technology, we can look forward to a future where AI can learn to navigate not just the virtual world of a game, but a wide range of real-world tasks and environments.
