Meta's chief AI scientist, Yann LeCun, has shared insights into the future of artificial intelligence, suggesting that current AI systems are still far from achieving human-level intelligence. While some advancements, such as OpenAI’s “memory” feature and “complex reasoning” in ChatGPT, may seem like steps towards Artificial General Intelligence (AGI), LeCun believes we are still a decade away from reaching that milestone. He emphasizes the need for a new approach: building AI systems based on "world models."
The concept of a world model is about enabling AI to not just predict or react to individual inputs, like language models or image models do, but to understand and plan in a three-dimensional world, much like humans do. Today’s large language models (LLMs) are limited because they only predict the next word or token, and image or video models predict the next pixel. While effective in their specific tasks, these models do not truly comprehend the real world, which is crucial for achieving AGI.
LeCun points out that humans, by the age of 10, can learn to perform physical tasks such as clearing a table or driving a car with minimal training. In contrast, even the most advanced AI systems, despite being trained on massive amounts of data, still struggle with basic real-world tasks. To bridge this gap, LeCun suggests AI needs to develop a deeper understanding of the physical world, which is where world models come into play.
A world model is essentially an AI's internal understanding of how the world operates. It allows the AI to predict the outcomes of actions before they are taken. For instance, if you look at a messy room, you can plan how to clean it without trial and error. This is the type of reasoning that world models aim to replicate in AI systems.
World models have the potential to handle more data and perform more complex tasks, but they are also computationally demanding. Many AI labs are now exploring this approach, including a startup called World Labs, which raised $230 million to advance this technology. OpenAI has also referred to its upcoming video generator, Sora, as a world model, although details remain scarce.
LeCun himself has discussed world models in his 2022 paper on "objective-driven AI," a concept that dates back over 60 years. The idea is to feed a model with real-world data, give it an objective (like cleaning a room), and let it predict the sequence of actions needed to achieve that objective while following safety guidelines.
LeCun’s lab at Meta, known as FAIR (Fundamental AI Research), is focusing on long-term AI projects like world models. Although progress has been slow, LeCun believes these models are essential for creating AI systems that can think, plan, and reason at the same level as humans. However, achieving this goal will take years, if not a decade, to fully realize.
