Tech leaders eye world models as link to smarter AI

Your house cat might be smarter than ChatGPT. While artificial intelligence can now write poetry and pass law school exams, it still can’t match the basic reasoning abilities that let your tabby navigate window ledges and pounce on prey.

That gap—between AI’s command of human knowledge and its inability to match an animal’s understanding of the physical world—has emerged as a crucial challenge in the quest for more intelligent machines. Meta’s chief AI scientist, Yann LeCun, thrust the issue into the spotlight last week at a Paris AI safety summit, arguing that “world models“—AI systems that form internal representations of structure, dynamics and causal relationships—could be key to advancing artificial intelligence.

However, some leading researchers question whether this approach is as groundbreaking as it might seem. 

“AI has been using world models since the 1950s, and several subfields of AI depend entirely on world models,” Stuart J. Russell, Professor of Computer Science at the University of California Berkeley, told IBM Think. “It’s about as novel and ingenious as suggesting that mathematics could be useful for physics.”

The debate highlights an intensifying chase to achieve artificial general intelligence (AGI), or AI systems that can match or exceed human-level learning and reasoning across virtually any domain. Unlike today’s specialized AI systems, which excel at narrow tasks like chess or language processing, AGI would need to understand the world more fundamentally, combining reasoning abilities, physical understanding and adaptability.

Major tech companies have taken different approaches to this challenge. OpenAI, through its GPT series, has focused on scaling up language models to achieve increasingly sophisticated reasoning. Google DeepMind, with systems like AlphaFold and AlphaGo, has targeted specific domains while working toward more general capabilities. Meanwhile, Meta has emphasized the importance of learning from real-world interaction rather than text alone.

“Never mind trying to reproduce human intelligence,” LeCun said at the summit. “We can’t even reproduce cat intelligence or rat intelligence. Any house cat can plan very highly complex actions.”

In recent years, AI systems have rapidly increased their scores on benchmarks in specific domains. But LeCun points to what researchers call Moravec’s paradox: the observation that the skills humans find easy often prove hardest for machines to master.

“Things that we take for granted because humans and animals can do it, we think it’s not complicated—it’s actually very complicated,” LeCun said. “And the stuff that we think is uniquely human, like manipulating and generating language, playing chess, playing go ... turns out to be easy, relatively.”

This paradox—that AI excels at tasks we consider intellectually demanding while struggling with seemingly basic physical and perceptual abilities–draws skepticism from some experts.

“The process of going from a sequence of retinal images to, say, Maxwell’s equations for electromagnetism took the human race tens of thousands of years and required a cumulative process of concept formation and mathematization,” Russell says. “I’ve not seen anything resembling this in current deep learning systems.”

To illustrate current limitations, LeCun pointed to a striking comparison: modern LLMs train on trillions of tokens—an amount of text that would take a human half a million years to read. Yet a four-year-old child, awake for only about 16,000 hours, processes a comparable amount of data through visual perception alone.

“A four-year-old child has seen as much data as the biggest LLM in the form of visual perception, and for blind children, it’s touch,” LeCun said. “That tells you a number of things. We’re never going to get to human-level intelligence by just training on text.”

This observation has pushed researchers to explore new approaches. At Meta, LeCun’s team has even moved away from the term “artificial general intelligence” altogether, preferring “advanced machine intelligence” (AMI). “The reason being that human intelligence is actually quite specialized, and so calling it AGI is kind of a misnomer,” LeCun said.

LeCun thinks we need to completely rethink how AI will evolve. Rather than trying to copy the human brain, he wants to change how AI learns about and makes sense of the world. He argues that AI needs to build its own mental picture of reality by taking in information through its “senses,” like learning how physical objects behave just by watching videos. These AI systems would need to remember things consistently and know how to plan their actions step by step to get things done.

Computer scientists differ on how AI systems might develop their own understanding of the world. University of Washington Professor Emeritus Pedro Domingos sees progress as possible but not immediate. “It’s entirely realistic to build AIs that develop their own world models,” he told IBM Think, “but we don’t yet know how, and more research is needed.”

Current AI capabilities fall short of human-level reasoning, particularly in handling complex tasks. As Russell puts it: “Since the 1960s, we have understood that using simple, uniform world models leads to completely intractable reasoning and planning problems. (Imagine trying to plan a vacation by figuring out in advance the exact sequence of 800 million muscle activations that will be needed.)” Humans, he notes, process information differently: “The principal tool that humans use to overcome this is hierarchy—we operate with many models at many levels of abstraction, all the way from big, high-level actions ... down to tiny, low-level actions, such as moving a finger to type the next letter ‘i’ in this email.”