technology called At the heart of ChatGPT is the potential for more than just talking. Linxi “Jim” Fan, his AI researcher at chipmaker Nvidia, has teamed up with several colleagues to develop the powerful language model GPT-4 (the “brain behind ChatGPT and a growing number of other apps and services). ”) can be freely set internally. Blocky video game “Minecraft”.
The Nvidia team, which includes the company’s director of machine learning and Caltech professor Anima Anandkumar, created a Minecraft bot called Voyager that uses GPT-4 to solve in-game problems. The language model generates goals that help the agent explore the game and code that improves the bot’s skill in the game over time.
Voyager doesn’t play games like humans do, but it can read game state directly through the API. For example, it might see that you have a fishing rod in your inventory, that there’s a river nearby, and suggest a goal to use GPT-4 to fish for experience. This goal is then used to have her GPT-4 generate the code the character needs to accomplish the goal.
Provided by NVIDIA
The most novel part of the project is the code that GPT-4 generates to add behavior to Voyager. If the originally proposed code doesn’t run perfectly, Voyager will try to improve it using error messages, feedback from the game, and explanations of the code generated by GPT-4.
Voyager builds a library of code over time to learn how to create more and more complex and explore the game further. A graph created by the researchers shows how capable it is compared to other Minecraft agents. Voyager gets more than three times as many items as she does. You can explore more than double the distance. He builds tools 15 times faster than any other of her AI agents. Fan said that adding ways to incorporate visual information from games into the system could improve this approach in the future.
Chatbots like ChatGPT have amazed the world with their eloquence and apparent knowledge, but Voyager has great potential for language models to perform useful actions on computers, even if they are often hoaxes. is shown. Using language models in this way has the potential to automate many mundane office tasks, which is potentially one of the greatest economic benefits of technology.
Provided by NVIDIA
The process by which Voyager uses GPT-4 to determine how to navigate in Minecraft could be adapted to software assistants determining how to automate tasks via the operating system of a PC or mobile phone. The startup OpenAI, which developed ChatGPT, added a “plugin” to the bot that allows it to interact with online services such as grocery delivery app Instacart. Microsoft, which owns Minecraft, is also training AI programs to play Minecraft, and the company recently announced Windows 11 Copilot, an operating system feature that uses machine learning and APIs to automate certain tasks. Did. It might be a good idea to try out this kind of technology in a game like Minecraft, where the harm from flawed code is relatively low.
Of course, video games have long been testbeds for AI algorithms. AlphaGo is a machine-learning program that mastered the extremely subtle board game Go in 2016, breaking the mold by playing a simple Atari video game. AlphaGo used a technique called reinforcement learning. It trains the game-playing algorithm by giving positive and negative feedback, such as in-game scores.
Guiding an agent in this way is more difficult in an open-ended game like Minecraft, where there is no score or set of goals, and the player’s actions may not be rewarded until much later. Whether you believe we should prepare to contain the AI existential threat now or not, Minecraft seems like a great playground for technology.
