AI that Learns Like a Baby

Yann LeCun, a Turing Award recipient and AI pioneer, holds several contrarian views regarding the current trajectory of artificial intelligence. His ideas focus on moving beyond language-centric models toward systems that understand the physical world through observation and open collaboration.

The Limitations of Large Language Models (LLMs)

LeCun believes the industry’s current obsession with LLMs is "wrong-headed" and will ultimately fail to solve many pressing problems. While he acknowledges they are useful for tasks like writing code or text, he argues that the belief that scaling them up will lead to human-level intelligence is a "delusion". According to LeCun, LLMs are limited because:

• They are confined to the discrete world of text and lack a model of the physical world.

• They cannot truly reason or plan because they cannot predict the consequences of their actions.

• They suffer from the Moravec Paradox, where tasks easy for humans (like perception and navigation) are incredibly difficult for computers.

World Models and the JEPA Architecture

Instead of LLMs, LeCun advocates for "world models"—AI designed to accurately reflect the dynamics of the real world. His specific solution is the Joint Embedding Predictive Architecture (JEPA). Key features of this approach include:

• Non-Generative Learning: Unlike generative AI, JEPA learns to represent the world in an abstract space, ignoring unpredictable details to focus on underlying rules.

• Learning from Observation: Similar to a baby learning about gravity, JEPA learns by observing video, audio, and sensor data rather than just text.

• Foundation for Common Sense: This architecture is intended to provide AI with the common sense necessary to reason and plan in the real world.

The Importance of Open-Source and Sovereignty

LeCun is a staunch advocate for open-source AI and criticizes the "closed" approach of labs like OpenAI and Anthropic. He views AI as a platform that should be open to ensure a high diversity of AI assistants with different linguistic abilities and value systems. Furthermore, through his new company, AMI Labs, he aims to provide a "sovereign" alternative to the current US-China binary, offering a credible frontier AI option for European and global interests.

Future Applications: Robotics and Industry

LeCun argues that a world model is the "key to unlocking" advanced technology, such as Level 5 autonomous driving and truly useful domestic robots. He critiques current humanoid robots as being "planned in advance" and lacking the intelligence to generalize when an environment changes. He believes "agentic systems" will only become reliable when they possess a world model to predict the consequences of their actions.

The Role of Academia

LeCun advises universities to stop working on LLMs, as they cannot compete with the massive computing resources of industry. Instead, he urges academia to focus on long-term objectives and conceptual breakthroughs, noting that the most exciting work on world models is currently coming from academic settings rather than big industrial labs.