At the AI Action Summit in Paris, Meta’s chief AI scientist Yann LeCun shared concerns about today’s artificial intelligence systems. He said current AI models, like popular chatbots, still fall short of true human-level intelligence.
LeCun explained that these systems cannot remember past interactions, reason logically, understand how the physical world works, or plan tasks step-by-step. Instead, they rely on recognizing patterns in large amounts of data but lack real understanding.
To address this, Meta is developing new types of AI. One project involves a world-based model that uses retrieval-augmented generation (RAG) to fetch accurate information from outside sources. Another system, called V-JEPA, learns by predicting missing parts in video clips, helping the AI grasp real-world scenarios more effectively.
LeCun believes these new models could be the future of AI. They are designed to learn from real life and make abstract predictions like humans, potentially allowing AI to handle complex problems better.
However, Meta faces internal challenges. Many key researchers who helped build its 2023 AI model, Llama, have left the company. Only three of the original 14 researchers remain. Some have joined Mistral, a fast-growing AI startup in Paris that is attracting attention with its innovations.
Meanwhile, Meta’s latest model, Llama 4, has not gained much support from developers. Rivals like OpenAI’s GPT-4o, Google’s Gemini 2.5 Pro, and Anthropic’s Claude 4 Sonnet are currently leading the way in AI reasoning capabilities.
The release of Llama 4 Behemoth, Meta’s most ambitious AI model yet, has been delayed. This has raised concerns about whether Meta can keep up with its competitors. As the company works on more advanced and realistic AI models, it must also focus on keeping its talent and staying ahead in the fast-changing AI race.
