Recently, the Verses team developed the Genius intelligent agent, which achieved remarkable results in the classic game Pong. Using only 10% of the data and 2 hours of training time, it surpassed top human players and other AI models. This groundbreaking advancement marks a new milestone in AI technology and indicates the future direction of intelligent agent development.
The success of the Genius agent is attributed to its unique design philosophy. Compared to traditional large models, Genius is only 4% the size of the SOTA model IRIS and can run on a standard M1 chip MacBook. Researchers were inspired by an experiment from four years ago, where scientists discovered that a "brain in a dish" could learn to play Pong in just 5 minutes, prompting them to consider mimicking the way the human brain works.
Image Source Note: Image generated by AI, licensed through Midjourney
The Verses team believes that traditional large model-based AI agents have significant shortcomings in logical reasoning. Existing models tend to rely heavily on memorizing reasoning steps from training data, lacking true initiative and curiosity. The Genius agent employs the concept of a cognitive engine, which not only possesses cognitive, reasoning, and decision-making abilities but also empowers the agent with active learning capabilities.
In comparison tests with IRIS and other AI models, Genius demonstrated strong learning abilities. Researchers trained Genius with 10,000 steps of game data in 2 hours, and the results showed that its performance surpassed that of IRIS, which had been trained for two days. The success of Genius lies not only in its rapid learning ability but also in its proactive behavior during gameplay. For instance, in a Pong match, Genius was able to come from behind to win, a phenomenon that did not occur during IRIS's training.
However, researchers also caution that, despite Genius's impressive performance, there is currently a lack of a unified standard to comprehensively measure AGI performance. Diverse testing is needed to validate its adaptability and reliability across different fields.
This research not only advances the development of AI agents but also provides new ideas and methods for future explorations in machine intelligence.
Paper link: https://arxiv.org/pdf/2410.05229
