{"id":15426,"date":"2025-04-28T18:23:12","date_gmt":"2025-04-28T23:23:12","guid":{"rendered":"https:\/\/www.inthacity.com\/blog\/uncategorized\/metas-ai-boss-done-with-llms-shift-to-advanced-ai-models\/"},"modified":"2025-05-16T20:54:57","modified_gmt":"2025-05-17T01:54:57","slug":"metas-ai-boss-done-with-llms-shift-to-advanced-ai-models","status":"publish","type":"post","link":"https:\/\/www.inthacity.com\/blog\/tech\/ai\/metas-ai-boss-done-with-llms-shift-to-advanced-ai-models\/","title":{"rendered":"Meta’s AI Boss Declares He’s Done With LLMs, Shifting Focus to Advanced AI Models"},"content":{"rendered":"

The Case Against LLMs: Why Yann LeCun Isn\u2019t Impressed<\/h2>\n

Yann LeCun, a pioneer in the AI space and Chief AI Scientist at Meta, isn\u2019t shy about sharing his opinions. At the Nvidia GTC 2025 conference, he dropped a bombshell: \u201cI\u2019m not so interested in LLMs anymore.\u201d<\/em> For those who\u2019ve been following the hype around ChatGPT and similar models, this might sound shocking. But LeCun\u2019s skepticism isn\u2019t unfounded. He argues that LLMs, while impressive, are limited in their ability to understand the physical world, reason, and plan\u2014key components of true artificial general intelligence (AGI).<\/p>\n

LeCun points out that LLMs excel at generating text by predicting the next word in a sequence. But text, he says, is a poor model for understanding the complexities of the real world. \u201cText is a very lossy compression of the world,\u201d he explains. \u201cIt\u2019s like trying to understand a movie by reading the script.\u201d In other words, LLMs can mimic human language, but they lack the depth and richness of human understanding.<\/p>\n

World Models: The Missing Piece of the AI Puzzle<\/h2>\n

So, if LLMs aren\u2019t the answer, what is? LeCun believes the key lies in \u201cworld models\u201d\u2014systems that can understand the physical world as humans do. A world model is essentially a mental representation of how the world works. For example, we know that if we push a bottle from the top, it\u2019ll tip over, but if we push it from the bottom, it\u2019ll slide. These intuitive understandings of physics are something humans develop in the first few months of life. But AI systems, despite their computational power, struggle with this.<\/p>\n

LeCun argues that the architectures we use for AI today\u2014like transformers, which power LLMs\u2014are ill-suited for building world models. Transformers are great at predicting the next token in a sequence, but they fall short when it comes to reasoning about the physical world. \u201cTokens are discrete,\u201d LeCun explains. \u201cBut the world is continuous and high-dimensional.\u201d This mismatch, he says, is why AI systems still struggle with tasks that come naturally to humans and even animals.<\/p>\n

Meet JEPA: The Future of AI Architecture<\/h3>\n

Enter JEPA, or Joint Embedding Predictive Architecture, LeCun\u2019s proposed solution to the limitations of LLMs. JEPA is designed to learn abstract representations of the world, allowing AI systems to reason and plan more like humans. Unlike generative models, which try to predict every detail of an image or video, JEPA focuses on understanding the underlying structure. This makes it more efficient and better suited for real-world tasks.<\/p>\n

LeCun and his team have been working on JEPA for years, and they\u2019re now gearing up to release version 2. Early results are promising, with JEPA demonstrating the ability to predict physical outcomes in videos\u2014like determining whether a sequence of events is physically possible. This is a significant leap forward, as it moves AI closer to understanding the world in a more human-like way.<\/p>\n\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t
<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t
<\/div>\n\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\n\t\t\t\t\t
\nPrev<\/a>\n1<\/span> of<\/span> 1<\/span>\nNext<\/a>\n<\/div>\n
\n\t