How do you become an expert Minecraft player, or learn how to beat the game as fast as humanly possible in the shortest amount of time?
Easy, you just create a bot powered by OpenAI’s GPT-4. That’s exactly what NVIDIA AI Scientist Jim Fan and his team did, with their creation of Voyager.
Voyager, a one-of-a-kind artificial intelligence (AI) agent built using GPT-4, was set loose in the Minecraft universe. Voyager isn’t just your regular AI; it’s a lifelong learner, continuously bettering itself by writing, refining, and storing its own code.
In a groundbreaking approach, Voyager changes the way we traditionally think about AI training, explains Fan. It learns by executing code rather than going through the usual method of gradient descent. Instead of being a fixed ‘trained model’ represented by numbers in a matrix, it’s a growing storehouse of skills that it refines and uses as needed, creating a dynamic and self-improving AI. A super duper learning computer that can dig through the Minecraft world like nobody’s business.
Voyager isn’t just any Minecraft player; it’s a master at it. It collects 3.3 times more unique items, journeys 2.3 times longer distances, and advances through tech tree milestones up to 15.3 times quicker than other methods. The techniques that Voyager employs have been made open-source, meaning anyone can use them.
The structure of Voyager is made up of three parts:
This method lets Voyager tackle tasks in a way that responds to the context of the game, learning to adapt to different situations and environments. Voyager first tries to write a program to achieve a specific goal using the Mineflayer Javascript API. It then gets better and better at this program by using feedback from the game and correcting errors that occur during execution. Over time, Voyager builds a library of successful programs that it can call upon when needed (kind of scary but ingenious).
Compared to other AI agents, Voyager is by far the smartest in Minecraft. It discovers 63 unique items in just 160 attempts and travels far and wide across different terrains without needing explicit instructions to do so. This independent learning ability, together with the architecture that allows it to store and reuse learned skills, makes Voyager a powerful example of what AI can do.
According to Fan, GPT-4 is far superior even to its predecessor, when asked about other AI models tested. “We tried GPT-3.5, and the performance drops massively. So Llama/Vicuna will only be worse. GPT-4 has a quantum leap in coding and refining,” he said. Sorry kids, you’ve just lost your job.
Currently, Voyager operates only through text, but it’s ready to be upgraded with visual perception capabilities in the future. Despite this, it’s already showing impressive abilities, such as the capability to build complex 3D structures like a Nether Portal and a house.
Voyager is showing us that with the right technology and approach, AI can learn to navigate complex environments and tasks on its own. It’s a big step forward for artificial intelligence. It’s kind of creepy yet incredibly cool at the same time.
