Nvidia’s GPT-4 powered bot in Minecraft outperforms all other AI agents

The pairing will have implications far outside the block-based game.
Ameya Paleja
GPT-4 powered agent can automate tasks inside Minecraft
GPT-4 powered agent can automate tasks inside Minecraft


A team of researchers at chip maker Nvidia plugged the powerful language model GPT-4 inside the Minecraft bot and found that it could now build tools 15 times faster than other artificial intelligence (AI) agents used in the game.

An agent in Minecraft is a bot that can be programmed to automate a set of tasks. OpenAI, the creators of ChatGPT, have also dabbled in creating a bot that can play Minecraft independently. To do so, the company spent more than 70,000 hours in training the bot.

This is the most conventional of approaches used when it comes to training AI. AlphaGo, the program that made global news by defeating a world champion in the board game, began its training by watching others play the game.

However, a research team at Nvidia, the company that made the chips to herald the age of artificial intelligence (AI), has found a nifty way to let agents overcome hurdles in the game.

Plugging GPT-4 into Minecraft

The research team, headed by Anima Anandkumar, the director of machine learning at the company and a professor at Caltech, created a bot called Voyager which can tap into the powers of GPT-4 to play the game. The language model can generate objectives for the bot to explore the game and provide it with the necessary code to help it improve its skill.

The bot does not play the game like a regular human would. Instead, it seeks the events in the game through an API and then uses the information available to plan its next steps.

For instance, the bot can use the fishing rod in its inventory at a nearby river to gain fishing experience and enhance its skill, Wired said in its report. Since the process can be error-prone, the researchers have also provided a feedback mechanism that the bot can use to improve its actions.

Over time, the bot can build a library of code that helps it learn to make complex things. A comparison with the progress rate of other agents shows that the GPT-4 powered bot can gather three times as many items, explore twice as far and build things 15 times faster. The researchers even shared a chart to demonstrate the pace of progress.

Nvidia’s GPT-4 powered bot in Minecraft outperforms all other AI agents
A comparison of exploration performance between different AI agents

Going beyond games into the real world

The process to help the bot navigate Minecraft can also be used to train software to automate tasks on a computing device. Like Nvidia, Microsoft owns Minecraft and has been training programs to play the game. However, it has also developed its Windows 11 Operations System features, where APIs and machine learning can work together to automate tasks.

Interesting Engineering has previously reported that AI is expected to impact 80 percent of jobs in the U.S. and EU markets. As companies prepare to lay off thousands of people and adopt AI instead, the technology must be tested heavily for its ability to make the right decisions.

A game like Minecraft can be a great testing bed for such technology since even grave errors will not impact the real world.

Add Interesting Engineering to your Google News feed.
Add Interesting Engineering to your Google News feed.
message circleSHOW COMMENT (1)chevron
Job Board