Meta’s AI boss says current AI lacks ‘intelligent behavior’ — Report

Meta’s AI Chief Highlights Deficiencies in Current Models, Advocates for a New Approach

Yann LeCun, Meta’s chief AI scientist, has identified four crucial aspects of intelligent behavior missing from current AI models: understanding the physical world, persistent memory, reasoning, and hierarchical planning. He argues that existing Large Language Models (LLMs), which power many popular AI chatbots, fall short in these areas. LeCun believes that the current approach of “cobbling capabilities” onto existing models is insufficient and calls for a fundamental shift in training methodologies. He advocates for “world-based models” trained on real-life scenarios to achieve higher-level cognition. These models would learn by predicting the outcome of actions, mirroring the human process of understanding the physical world through abstraction. Meta is already exploring this direction through research into retrieval augmented generation (RAG) and non-generative models like V-JEPA.

The shortcomings of current LLMs are further highlighted by Meta’s recent struggles. The company is experiencing significant talent drain from its AI research team, particularly among the original Llama model creators. Only three of the fourteen original authors remain at Meta, with many joining the Paris-based startup Mistral. This exodus coincides with a lukewarm reception to Llama 4, a model that fails to compete with faster-moving rivals such as OpenAI’s GPT-4o, Google’s Gemini 2.5 Pro, and Anthropic’s Claude 4 Sonnet, all of which feature dedicated reasoning models. Furthermore, Meta’s flagship LLM, Llama 4 “Behemoth,” has reportedly been delayed.

LeCun’s critique underscores the significant challenges facing the AI industry in developing truly intelligent systems. The limitations of current LLMs, combined with the talent drain at Meta, suggest a potential turning point in the race to create advanced AI. The industry may need to move beyond incremental improvements and embrace fundamentally different approaches to achieve the level of intelligence LeCun envisions. Meta’s shift towards world-based models represents a bold, albeit potentially risky, strategy in this ongoing race.