Meta’s Chief Scientist Yan LeCun has made a bold statement that is stirring the AI community. He claims that we should not talk about AGI, or artificial general intelligence. Instead, we should focus on what kind of intelligence current AI systems lack compared to humans and animals.
LeCun argues that current AI systems, like large language models (LLMs), lack a lot of basic skills that even pets have. They don't understand the physical world, and they lack persistent memory. This means they can't reason or plan like humans.
LeCun's team at Meta is working on something called "Joint Embedding Predictive Architectures" (JEPA). This model is trained on video data to learn about the physical world, similar to how babies learn by observing their surroundings. The goal is to create intelligent systems that can learn with fewer examples and without needing to be explicitly told what each example is.
Unlike generative models that try to recreate every detail, JEPA focuses on understanding the broader picture. This makes the training process more efficient. LeCun believes that this approach will lead to machines that can understand, plan, and reason better than current AI systems.
LeCun also points out that LLMs have limitations. They are trained on tons of text data but still don't understand basic logic or how the world works. For example, they might struggle with simple tasks like counting letters in a word unless guided step-by-step.
LeCun argues that scaling up LLMs won't lead to human-level intelligence. He believes that the future of AI lies in non-generative models that can learn from video and other non-text data. He suggests that these new models could one day achieve human-like understanding and reasoning.
Energy efficiency is another challenge in AI. Current AI systems consume a lot of power. The human brain, in contrast, is highly efficient. New technologies like photonic chips, which use light instead of electrical signals, offer a pathway to more energy-efficient computing. These chips can handle higher bandwidths and generate less heat, making them ideal for the next generation of AI.
In conclusion, LeCun's views challenge the current focus on LLMs. He promotes a new approach based on learning from videos and other non-text data. This could lead to machines with a deeper understanding of the world, better reasoning skills, and more efficient use of energy. The next few years will be crucial in seeing if his ideas will shape the future of AI.