Yann LeCun, one of AI's most influential researchers, has made a dramatic exit from Meta to pursue a fundamentally different vision for artificial intelligence. Rather than betting on large language models (LLMs), which predict text one word at a time, LeCun is building world models that learn how the physical world actually works. His new company, Advanced Machine Intelligence (AMI), represents a direct challenge to the industry consensus that language is the path to human-level AI.

What Exactly Are World Models, and Why Does LeCun Think They Matter?

World models are AI systems designed to predict the consequences of actions in real environments. Unlike LLMs, which excel at language manipulation, world models simulate how the world responds to decisions. LeCun explained that intelligent behavior requires three core capabilities: the ability to predict action consequences, planning through reasoning and optimization, and the capacity to handle new tasks without extensive retraining.

LeCun argues that LLMs fundamentally lack these abilities. "LLMs don't have the ability to predict the consequences of their own actions, nor do they have real planning ability, because reasoning in LLMs is just predicting the next token, not searching," he stated. This distinction matters enormously for applications beyond chatbots, from robotics to drug discovery to autonomous vehicles.

Why Did LeCun Leave Meta, and What Changed?

By early 2024, LeCun found that Meta's research division, known as FAIR (Facebook AI Research), no longer aligned with his vision. The company's pivot toward competing in the LLM race meant that exploratory research into world models was deprioritized. While his own work on JEPA (Joint Embedding Predictive Architecture) and world models wasn't directly affected, the broader organizational shift signaled that Meta was no longer the right home for his ambitions.

LeCun acknowledged that Meta's leadership, including CEO Mark Zuckerberg, was supportive of him personally. However, the company's acquisition of Scale AI and its focus on catching up with competitors like OpenAI meant that breakthrough research took a backseat to product development and scaling. "By the end of last year, it was obvious that Meta was no longer the right place to advance this project, so I left and founded AMI," LeCun explained.

How Does LeCun's Approach Compare to Other World Model Competitors?

LeCun is not alone in pursuing world models. Runway, a video-generation startup valued at $5.3 billion, launched its first world model in December and plans another release this year. Google's Genie model, Fei-Fei Li's 3D space models, and startups like Luma and World Labs are all chasing similar goals. However, their approaches differ significantly.

Runway co-founder Anastasis Germanidis believes that training models directly on observational data from the world, rather than on text from the internet, is the next frontier. "Language models are trained on the entire internet, on message boards and social media, on textbooks, distilling the existing human knowledge. But to get beyond that, we need to leverage less biased data," Germanidis noted. Runway's path emphasizes video generation as a stepping stone to world models, while LeCun's JEPA approach focuses on learning predictive representations without generating pixel-perfect outputs.

Anastasis Germanidis

Steps to Understanding the World Models Revolution

• Language Limitations: LLMs are trained on text from the internet and excel at language tasks, but they cannot predict how physical systems respond to actions or plan sequences of decisions in real environments.
• Observational Learning: World models learn from video, sensor data, and other observational inputs, allowing them to understand physics, causality, and real-world dynamics without explicit programming.
• Practical Applications: These models have near-term use cases in robotics training, interactive entertainment, gaming, drug discovery, climate modeling, and autonomous systems that require real-world reasoning.

What Does This Mean for the Future of AI Research?

LeCun's departure signals a philosophical divide in AI development. While OpenAI, Anthropic, and others continue scaling LLMs, LeCun and his peers are betting that the next breakthrough requires a fundamentally different architecture. He has been vocal about his skepticism toward LLM-focused research, even advising PhD students to avoid the field. "If you're doing a PhD, don't work on LLMs. It's meaningless. You can't make a contribution," he remarked.

This contrasts sharply with the industry's current trajectory. LeCun also criticized Geoffrey Hinton and Yoshua Bengio, fellow Turing Award winners who have recently become vocal about AI risks. He suggested they embraced LLMs only after GPT-4's release and now promote AI regulation through fear. "I feel like he just wants to slack off: 'Okay, this is what we need. I can declare victory.' Yeah, I can retire," LeCun said of Hinton.

The stakes are enormous. If world models prove to be the path to artificial general intelligence (AGI), the companies that master them first will reshape AI's trajectory. Runway has raised $860 million to date, including a $315 million round in February from AMD Ventures and Nvidia. Competitors Luma AI and World Labs have raised $900 million and $1.29 billion respectively, according to available funding data. Yet all face competition from tech giants like Google, whose parent company Alphabet is worth $4.86 trillion.

LeCun's optimism about world models is tempered by realism about the challenge ahead. He believes that achieving breakthrough research requires hiring the best people, providing them resources, and then "getting out of the way." Whether AMI can execute on this vision while competing against well-funded incumbents remains an open question. What is clear is that the AI industry's consensus around LLMs is fracturing, and the next chapter of AI development may be written by those willing to bet against the crowd.