“Yann LeCun, Meta's chief AI scientist, is leaving the company to start his own venture that he said will partner with the social media giant.
LeCun said in a LinkedIn post Wednesday that he was creating a startup "to bring about the next big revolution in AI."
The company plans to focus on "systems that understand the physical world, have persistent memory, can reason, and can plan complex action sequences," he wrote.
The world models LeCun described learn by taking in visual information, much like a baby animal or young child does.
LeCun created an AI research organization inside Meta in 2013 and led it for four years before becoming Meta's chief AI scientist. His views on how to advance the state of AI have diverged from Meta's in recent years.
Meta, like many of its industry peers, has raced to build large language models and continually scale them, in the hopes of reaching superintelligence, or intelligence that is smarter than humans. LeCun has said repeatedly that he doesn't believe that LLMs will lead to that level of intelligence, instead putting his weight behind emerging world models.
Over the summer, Meta named another chief scientist within its lab focused on developing superintelligence, while LeCun held the role within the Fundamental AI Research group.
The Wall Street Journal reported last week that LeCun had recently talked to associates about leaving Meta -- recruiting colleagues for his new venture and speaking to investors. He plans to leave Meta at the end of the year.
A Meta spokesperson confirmed the news but declined to comment further.” [2]
1. It is not accurate to say world model is definitively better than LLM, as they serve different, though related, purposes, and world models are seen by many as a necessary evolution beyond LLMs, especially for physical tasks.
LLMs excel at processing and generating text by identifying patterns, while world models are internal, predictive systems that simulate reality, allowing AI to "imagine" consequences before acting, which is crucial for robotics, planning, and understanding the physical world.
Large Language Models (LLMs)
Function: LLMs are trained on vast amounts of text and other data to perform tasks like language translation, summarization, and content generation.
Strength: They are powerful at handling language and finding patterns within their training data, but their understanding of the real world can be limited and brittle.
Limitation: They primarily rely on pattern matching and do not possess a robust internal simulator of how the world works, which can lead to unexpected failures.
World models
Function: These are internal simulations of the environment that allow an AI to predict and reason about the consequences of actions before they happen, much like a human "imagines" the outcome of touching a hot stove.
Strength: They are essential for tasks requiring a deep understanding of physics, causality, and the real-world dynamics needed for planning and control in physical environments, such as robotics and embodied AI.
Advantage: They provide a crucial ability to learn, reason, and plan in the physical world that goes beyond the text-based pattern matching of LLMs.
The relationship between LLMs and world models
World models are not a replacement for LLMs but are seen as the next step for creating smarter, safer AI systems.
The two can work together: a world model can be used to enhance an LLM by providing it with a better understanding of cause and effect, and LLMs can be used within the architecture of a world model to help with tasks like planning and social understanding.
2. Meta's Chief AI Scientist To Depart for New Venture. Bobrowsky, Meghan. Wall Street Journal, Eastern edition; New York, N.Y.. 20 Nov 2025: B4.
Komentarų nėra:
Rašyti komentarą