We know that one of the ultimate goals of AI development is to achieve AGI (Artificial General Intelligence), which is general artificial intelligence, meaning that machines can autonomously perform all human tasks.
Currently, the applications of large language models have achieved some astonishing results. However, these models still have certain limitations, such as the phenomenon of “hallucination.” Sometimes, the model generates content that seems reasonable but is actually meaningless, just like a person talking in their sleep.
If we analyze this solely from the perspective of training large language models, there are some solutions. Additionally, analyzing it from the field of embodied intelligence may provide refreshing viewpoints.
At the World Robot Conference in 2024, Wang Xingxing, the founder of Yushu Technology, proposed an interesting viewpoint regarding the current development of embodied intelligence: only the organic combination of thought and body can achieve true AGI.
01
Dreams and Hallucinations
Wang Xingxing talked about his experiences of feeling like he was falling off a cliff in a dream or wanting to speak but being unable to. He pondered for many years why such dreams occur. He later realized that in sleep, our thoughts and body are separated, losing the feedback mechanism of the body. Thoughts in this state can become detached, leading to various unrealistic hallucinations.
In real life, when we walk down the street, our bodies feel physical feedback such as friction and gravity, which helps our brains make correct judgments. However, in dreams, we lack this feedback, causing our brains to generate strange dreams.
Similarly, when you converse with a large language model, sometimes its answers seem reasonable, but at other times, they sound like nonsensical statements, lacking any real sense of logic; this phenomenon is referred to as the “hallucination” of large models. This hallucination is akin to talking in one’s sleep, both stemming from a detached state of thought.
Because large language models lack physical perception of reality when generating content, users need to constrain their scope by adding various prompts, which is currently a challenge in using large models; not everyone can provide very accurate prompts.
02
The Perfect Combination of Thought and Body
Research shows that human intelligence is not just a product of the brain, but also the result of the collaboration between the brain and the body. The body’s perception and motor abilities provide the brain with rich information, helping it better understand and perceive the world. This collaboration not only enhances human intelligence but also enables humans to adapt to various complex environments.
Let’s hypothesize: if we transplant a human brain into a pig, can the pig really become “smart”? The fact may be no, because the pig’s bodily functions fall far short of human levels; thus, a smart brain would be a burden for it, and advanced intelligence would not emerge.
03
Embodied Intelligence and Large Language Models: A Mutual Journey
Human advanced intelligence is not just a product of the brain but the result of a perfect combination of the brain and body. Similarly, achieving AGI should not focus solely on large language models but should also engage in a mutual journey with embodied intelligence.
The importance of embodied intelligence lies in its ability to provide a physical foundation for large language models, enabling them to better understand and perceive the real world. By giving robots physical forms and sensory capabilities, they can interact with the world like humans, providing large language models with rich data and feedback. This integration can not only reduce the occurrence of “hallucinations” but also enhance the intelligence level of robots, making them closer to human advanced intelligence.
Conclusion
The emergence of AI large language models has excited the entire industry, making people feel that the future is within reach. However, upon calm examination, the rise of large models is merely a small step toward the grand goal of general artificial intelligence. As the saying goes, a tall building rises from the ground; the current development is just laying a foundation.
To promote the continuous progress of general artificial intelligence, we cannot rely solely on large language models; we must advance hand in hand with embodied intelligence, achieving an organic fusion of the two. Only then can AGI possibly become a reality.
Here are technology trends
Also, practical insights
Altruism is self-interest
Let’s grow together
