Yann LeCun, Chief AI Scientist at Meta, has predicted that within the next 12 to 18 months, we will have a generalized method to train hierarchical world models.
Key Developments
According to Yann LeCun, these models will not only learn from text like current LLMs, but will learn directly from video and real-world data. The core difference lies in the ability to understand physical laws and causal relationships. Once a world model is established, AI can predict the outcomes of actions and thus plan to achieve specific goals.
This technology is expected to be widely applied in Robotics, where machines need to interact with the physical environment, as well as in healthcare and other fields requiring decision-making based on complex contexts.
Why It Matters
LeCun's prediction marks an important shift from purely language-based AI to AI with real-world knowledge (World Models). For the Vietnamese tech community, this is a signal to focus on research combining AI and robotics (Embodied AI). If successful, this will address the issues of logical reasoning and planning capabilities that current GPT models are struggling with.