Although Google DeepMind may seem like an unlikely location for the next huge leap in unnatural intelligence, it has just released an AI software that can learn how to complete tasks in a number of activities, including Goat Simulator 3.
Most amazingly, the program you consistently perform tasks when it is first introduced to a game by adapting what it has learned from previous games. The program, known as SIMA, for Scalable Instructable Multiworld Agent, builds on current AI advancements that have enabled big language models to create extremely able chabots like ChatGPT.
According to Frederic Besse, a study expert at Google DeepMind who was involved in the project, “SIMA is greater than the sum of its pieces.” It can use the common ideas in the game to improve its abilities and improve its ability to follow directions.
Expanding the types of data that algorithms may study from offers a way to more prominent capabilities as Google, OpenAI, and others compete to gain an edge in creating on the new conceptual AI boom.
In their latest video game project, DeepMind suggests that AI systems like Google’s Gemini and OpenAI’s ChatGPT could rapidly take control of computers and execute complex commands, giving users more opportunities to talk and create images or videos. That’s a vision being chased by both separate AI fans and big firms including Google DeepMind, whose CEO, Demis Hassabis, recently told WIRED is “investing strongly in that direction”.
” The paper is an exciting advance for embodied officials across multiple models,” says Linxi” Jim” Fan, a senior research scientist at Nvidia who works on AI game and was a part of a 2017 OpenAI job called World of Parts. Fan says the Google DeepMind work reminds him of this project as well as a 2022 effort called , VPT , that involved agents learning tool use in Minecraft.
He claims that SIMA expands its scope and shows stronger generalization of brand-new games. ” Similar to SIMA, the number of environments is still very small,” he said.
A New Way to Play
DeepMind’s new approach to AI technology, which the company has pioneered in the past, is illustrated in SIMA.
Before DeepMind was acquired by Google in 2013, the London-based startup demonstrated how a method known as reinforcement learning, which involves training an algorithm with positive and negative feedback on its performance, could aid computers in playing classic Atari video games. DeepMind created AlphaGo, a program in 2016 that used the same strategy to defeat a world champion of Go, an ancient board game that calls for subtle and instinctive skill.
For the SIMA project, the Google DeepMind team collaborated with several game studios to collect keyboard and mouse data from humans playing 10 different games with 3D environments, including No Man’s Sky, Teardown, Hydroneer, and Satisfactory. Later, DeepMind added descriptive labels to that information to reshape the actions users made, such as whether they were a goat searching for its jetpack or a human character looking for gold.
The human players ‘ data was then fed into a language model similar to that used in modern chatbots, which had developed the ability to process language by digesting a sizable corpus of text. The SIMA could then execute actions in response to commands that were typed. And finally, humans evaluated SIMA’s efforts inside different games, generating data that was used to fine- tune its performance.
After all that training, SIMA is able to carry out actions in response to hundreds of commands given by a human player, like” Turn left “or” Go to the spaceship “or” Go through the gate “or” Chop down a tree. More than 600 actions, ranging from tool use to exploration, combat, and the program can be performed. In accordance with Google’s ethical guidelines on AI, the researchers avoided games that feature violent actions.
” It’s still very much a research project,” says Tim Harley, another member of the Google DeepMind team”. However, one could envision agents like SIMA playing games with you and your friends one day.
Video games offer an incredibly safe environment in which AI agents can perform tasks. Agents must become more trustworthy in order to perform useful office or administrative tasks every day. According to Blake and Besse at DeepMind, they are developing methods to improve the agents ‘ dependability.
Updated 3/13/2024, 10: 20 am ET: Added comment from Linxi” Jim” Fan.
