GPT-4 faces laundry drying dilemma, humans assist in solving it, when will AI gain common sense?

Discussing the key role of embodiment and emotional factors in the development of artificial intelligence. Analyzing the importance of these two elements for achieving true intelligence, and their potential impact on AI research. The embodiment and emotional factors play crucial roles in the development of artificial intelligence. These elements are essential for achieving genuine intelligence and have significant potential impacts on AI research. Embodiment refers to the idea that intelligence requires a physical form or body to interact with the environment. This concept suggests that cognition and learning are deeply rooted in physical experiences and sensory-motor interactions. For AI, embodiment implies that true intelligence may require more than just processing power and algorithms; it may need a physical presence to fully understand and interact with the world. Emotional factors, on the other hand, relate to the capacity for AI to recognize, process, and express emotions. Emotions are fundamental to human intelligence and decision-making, influencing our perceptions, motivations, and social interactions. Incorporating emotional intelligence into AI systems could lead to more nuanced and human-like interactions, as well as improved decision-making capabilities. The importance of these elements for achieving true intelligence: 1. Enhanced learning and adaptation: Embodied AI can learn from physical interactions, leading to more robust and adaptable intelligence. 2. Improved context understanding: Physical presence and emotional awareness can help AI better understand complex situations and human behaviors. 3. More natural human-AI interaction: Embodied and emotionally intelligent AI can communicate and interact with humans more effectively. 4. Ethical decision-making: Emotional factors can contribute to more empathetic and ethically-aware AI systems. Potential impacts on AI research: 1. Shift in research focus: More emphasis on developing physical AI systems and emotional intelligence algorithms. 2. Interdisciplinary collaboration: Increased cooperation between AI researchers, roboticists, psychologists, and neuroscientists. 3. New evaluation metrics: Development of new ways to measure AI performance that include physical and emotional intelligence. 4. Ethical considerations: Greater focus on the ethical implications of creating emotionally intelligent and embodied AI systems. 5. Advancements in robotics: Accelerated development of more sophisticated and human-like robots. By incorporating embodiment and emotional factors into AI development, researchers may be able to create more advanced, adaptable, and human-like artificial intelligence systems, potentially leading to significant breakthroughs in the field.

In this QuantaMagazine podcast, the host interviews Yejin Choi, a computer science professor at the University of Washington. They explore interesting topics such as whether AI must acquire embodiment and emotions to develop human-like common sense.

Currently, GPT-4 has shown some impressive "human-like awareness". In this podcast, Yejin Choi and host Steven Strogatz discuss the capabilities and limitations of chatbots and the large language models (LLMs) that build them. They explore whether AI can truly understand the world and the questions they answer.

As early as the 1960s, computer scientists dreamed of brain-inspired computers exhibiting human-like intelligence. With the rise of the internet, the emergence of large text datasets, and significant advances in computing power, we seem to have reached a crucial moment. Today's LLMs appear to possess something close to human intelligence.

Theories proposed by Sanjeev Arora of Princeton University and Anirudh Goyal, a research scientist at Google DeepMind, suggest that today's largest LLMs are not random parrots. As models grow larger and train on more data, their language abilities improve and combine skills in ways that suggest understanding, developing new capabilities.

Choi explains that LLMs simply read vast amounts of text and learn to predict the next word, but on a massive scale. They don't necessarily "reflect word-for-word" on the training data but can generalize to some extent. If text is repeated frequently enough in internet data, it will indeed memorize it verbatim.

The process of training LLMs can be boiled down to building a very large neural network with layers upon layers of neurons stacked together, then feeding internet data sequentially. The goal of the learning process is to predict the next word based on the sequence of previous words.

Although simple, this training method can produce powerful results, allowing LLMs to answer various questions in text. However, the process of training LLMs is vastly different from how humans understand the world. Humans learn through courses and curiosity, making assumptions about the world.

Choi believes that LLMs are a kind of "thought-emotion soup". They mimic human emotions and intentions because people do indeed invest emotions and intentions in their writing. But ultimately, they don't truly possess the genuine emotions that humans have.

Link to original article