When Yann LeCun, Meta’s chief AI scientist and a Turing Award winner, declares that today’s artificial intelligence systems are “not smart,” the tech world listens. His assessment cuts through the hype surrounding large language models (LLMs) like ChatGPT and Gemini, highlighting a fundamental gap between current capabilities and genuine machine intelligence. LeCun argues that while these models excel at pattern recognition and statistical prediction, they lack the core cognitive abilities that define human or animal intelligence: persistent memory, reasoning, planning, and a grounded understanding of the physical world.
Current generative AI operates largely as advanced autocomplete engines. Trained on vast swathes of internet text, they predict the next token in a sequence based on statistical probabilities. This allows them to write code, summarize documents, and mimic conversational styles with uncanny fluency. However, LeCun points out that this mimicry is brittle. These systems do not build internal models of how the world works; they merely model the distribution of language. Consequently, they struggle with multi-step logic, cannot reliably distinguish fact from fabrication (hallucination), and fail to adapt to novel situations outside their training data. They possess no “common sense” — the intuitive physics and social norms humans acquire effortlessly through interaction.
To bridge this gap, LeCun has championed a paradigm shift toward “Objective-Driven AI” centered on world models. Instead of predicting the next word, these systems would learn to predict the consequences of actions in a simulated or real environment. The architecture, often referred to as Joint Embedding Predictive Architecture (JEPA), aims to learn abstract representations of the world by observing it — much like a baby learns gravity by dropping objects. Crucially, this approach moves away from generative reconstruction (predicting every pixel or word) toward predicting high-level, abstract representations. This allows the system to ignore irrelevant noise and focus on features necessary for planning and reasoning. LeCun’s recent startup endeavors and research at Meta’s FAIR lab are actively pursuing this non-generative, energy-based model approach to create agents that can plan hierarchically and reason over long time horizons.
The distinction between “predictive text” and “world understanding” isn’t academic — it dictates what AI can safely and reliably do in daily life. Current LLMs are powerful co-pilots for drafting emails or brainstorming, but they are unreliable for high-stakes decision-making. An AI with a world model could manage complex logistics, conduct scientific research with genuine hypothesis testing, or operate robots in unpredictable home environments. For the average user, this evolution promises a shift from tools that generate content to agents that solve problems. It also addresses safety: a system that understands cause and effect is inherently more controllable and alignable than one optimizing for plausible-sounding output. While LeCun estimates human-level AI (often called AMAS — Artificial Machine Intelligence) is still years or decades away, redefining the architectural target is the necessary first step toward machines that truly understand, rather than just imitate.
Image: Photo: Kindel Media · Pexels
Based on reporting from bbc.co.uk.
Cookies & Privacy: This website uses cookies to ensure you get the best experience on our website. Read more.
