[Tech Talk] What Happens When AI Starts Dreaming?

English Startup's 'World Model' Draws Attention
AI Enhances Learning Ability Through 'Jawakeumong' (Lucid Dreaming)
Human and Machine Learning Growing Increasingly Similar

Dreams we have while sleeping are considered the exclusive domain of higher organisms with consciousness. Although some mammals like dogs and cats, as well as certain highly intelligent creatures like octopuses, do dream, humans are known to be the only species that experience long and vivid dreams.

So, could highly advanced artificial intelligence (AI) also dream? And if AI were able to dream, what impact would that have on technology?

World Model That Can Create First-Person Perspective Videos Like Dreaming

The operation of Gaia-1, a world model developed by the UK autonomous driving technology company 'Wave'. This is a video generated by artificial intelligence, not an actual driving scene. [Image source=Wave]

On the 21st of last month (local time), the UK-based autonomous driving technology startup Wayve introduced a completely new type of AI called the "World Model." Wayve's World Model is named "GAIA," and it has the function of showing video footage based on commands input by users.

At first glance, GAIA does not seem much different from the recently popular generative AI. To some extent, that is true. GAIA itself is an AI created by integrating various image and video generation AIs.

However, GAIA is distinct in that it is a model specialized for autonomous vehicles. It also differentiates itself by implementing commands into very specific videos. It is as if the driver is experiencing a vivid dream from a first-person perspective. Moreover, depending on the commands, it can change the time of day in the video between day and night, look around the surrounding scenery, or even change the viewpoint within the same video.

Dreaming AI Advantageous for Handling Real-World Variables

Through Gaia-1, autonomous driving AI can create and learn numerous driving scenario situations. [Image source=Wave]

People who witnessed demonstrations of the GAIA model described it as "AI dreaming." Strictly speaking, it is closer to a "lucid dream" where the user can control the dream environment themselves.

Alex Kendall, founder of Wayve and a renowned authority in computer vision AI, also predicted that AI performance would advance once again through the GAIA model.

So why is a "dreaming AI" important? Kendall emphasizes that through GAIA, AI can continuously learn, simulate, and understand its own actions across various situations. This capability is crucial for controlling scenarios involving diverse "real-world variables," such as autonomous driving.

The biggest obstacle in autonomous driving technology is safety. Roads where humans and vehicles coexist are spaces filled with countless variables, and current AI struggles to actively respond to unpredictable situations. Therefore, various driving scenarios must be converted into data for learning, but conventional machine learning methods have limitations.

However, if AI can virtually encounter multiple situations through the GAIA model and learn various scenario patterns, the cost and time required for training would be greatly reduced.

Humans Also Relieve Brain Load Through REM Sleep

During the REM (Rapid Eye Movement) sleep stage, when humans dream, the brain is highly active and undergoes a process of organizing the stimuli received throughout the day.

In this regard, the GAIA model functions similarly to the dreams humans actually experience. We still do not know why humans and other higher vertebrates dream. However, we do understand what effect dreams have on the organism's body.

During the REM (Rapid Eye Movement) sleep stage, when people begin to dream, almost all bodily functions stop, but the brain becomes much more active. Because of this, some in academia argue that dreaming is a way for a resting human to unconsciously regulate brain functions.

In fact, in 2021, researchers at Tufts University in the U.S. proposed a hypothesis that REM sleep may be a process that restores the brain from an "overfitting" state. Humans receive countless stimuli while awake, and these stimuli remain in the brain as "data."

However, if the amount of certain data becomes too large or too small, resulting in an "overfitting" state, brain function abnormalities occur. Dreams resolve overfitting by randomly recombining stimuli received in reality and showing them to the brain.

Surprisingly, overfitting is a phenomenon frequently observed in AI as well. When AI receives too much specific data, it becomes overconfident in certain patterns, and when it receives too little, it becomes confused about situations. Wayve's GAIA model adjusts the behavior of autonomous driving AI by showing a virtual mix of various data the AI has previously learned.

Human and Machine Learning Growing More Similar

The autonomous driving development division of the American electric vehicle company Tesla also built a similar generative AI-based video measurement model. [Image source=Twitter]

Wayve is not the only company competing in the development of world models. Tesla, which is developing the autonomous driving service "FSD," has also started developing a model similar to GAIA.

Furthermore, Kendall predicts that world models will cause a major impact not only in autonomous driving but also in the robotics field. By synthesizing computer vision data, which has been a chronic problem in robot development, into world models, faster and cheaper AI training will become possible.

Looking at these developments, AI machine learning and human intelligence are becoming increasingly similar. Just as humans organize the stimulus data their brains receive through dreams, computers have developed methods to optimize their own performance through various generative models.

This content was produced with the assistance of AI translation services.