Current artificial intelligence systems such as ChatGPT, Claude and Gemini are powerful tools for writing, coding and solving maths problems, but leading AI researchers say they still fall far short of understanding the real world.
One of them is Yann LeCun, a pioneer in artificial intelligence and former chief AI scientist at Meta, who argues that today's AI cannot match even the basic understanding of a rat when it comes to interacting with the physical world.
"We don't have robots that are nearly as good at understanding the physical world as a rat," LeCun said during the VivaTech technology conference in Paris.
After leaving Meta in 2025, LeCun founded Advanced Machine Intelligence Labs (AMI Labs) with the aim of developing a new generation of AI capable of handling real-world situations, such as household chores and robotics.
According to LeCun, large language models (LLMs), the technology behind ChatGPT and similar chatbots, are excellent at tasks such as writing text, coding and solving mathematical problems because these tasks are well-defined. However, he believes they are not designed to understand how the real world works.
"They basically accumulate knowledge and reproduce it, but they are not particularly smart because they don't have a true understanding," he said.
LeCun argues that current AI models rely on statistical patterns rather than reasoning about physical reality. As an example, he said if a pen is balanced upright and released, even a young child knows it will fall over. However, no one can predict exactly which direction it will fall because the outcome depends on many unpredictable factors.
An LLM may attempt to predict one specific outcome based on its training data, but such a prediction would likely be wrong because it lacks a real understanding of the physical world.
To address this limitation, AMI Labs is developing a new AI system called Joint Embedding Predictive Architecture (JEPA).
Instead of trying to predict every possible detail, JEPA creates simplified representations of the real world, allowing AI to focus only on the information that matters. In the pen example, the system would recognise that predicting the exact direction of the fall is unnecessary.
The approach has attracted strong investor interest. Earlier this year, AMI Labs raised more than $1 billion in seed funding, one of Europe's largest early-stage investment rounds. Investors include US chipmaker Nvidia and the investment fund managing Amazon founder Jeff Bezos' private wealth.
Researchers say building AI that understands the physical world has become increasingly important as companies invest billions of dollars in humanoid robots.
Although robots have become more capable, teaching them to safely perform everyday household tasks such as ironing clothes or loading dishwashers remains difficult and expensive.
LeCun believes current language models are unlikely to solve that problem.
"LLMs are largely hopeless for robotics," he said, rejecting claims that simply making today's AI models larger will eventually produce superhuman intelligence.
His views are shared by many AI researchers.
Ingmar Posner, Professor of Applied Artificial Intelligence at the University of Oxford and director of its Applied AI Lab, believes future AI systems must be able to explain their decisions and understand cause-and-effect relationships.
"You need models that can answer questions like: What matters? What causes what? What would happen if I took a different action?" Posner said.
Posner and his research team have spent the past four years developing an alternative approach known as World Models.
The concept has existed for decades but gained renewed attention after a 2018 research paper by David Ha and Jurgen Schmidhuber suggested AI could learn by building internal simulations of the world.
Since then, companies including Google DeepMind have expanded research in this area. One version of Google's Dreamer World Model learned to collect diamonds in the video game Minecraft by imagining possible future scenarios before making decisions.
Posner's team is developing what he calls a "mechanistic world model," designed to organise knowledge so AI can efficiently recall, combine and update information when needed.
However, he cautioned that predicting when these new systems will become practical is difficult.
He noted that only a few years before ChatGPT was launched in late 2022, many researchers believed similar technology was still decades away.
Several major AI companies are now investing in world-model research. Google DeepMind is developing its Genie model, London-based Wayve has built a system called Gaia, while AI pioneer Fei-Fei Li founded World Labs in San Francisco in 2023 to develop another new AI architecture.
LeCun said AMI Labs plans to continue refining its AI system throughout this year and hopes to deploy it first in industrial applications next year.
If successful, he believes the technology could eventually evolve into general-purpose AI systems capable of handling a wide variety of real-world tasks with minimal additional training.
Despite concerns about increasingly capable robots, LeCun believes humans will continue to play the central role.
"We're still going to need humans to figure out what questions to ask, what to build and what to create," he said.
He expects future AI systems, even those that may surpass humans in some abilities, to function as assistants rather than replacements.
"Our interaction with future AI systems, even if they are smarter than us, will be like the relationship between a business leader or political leader and a team of highly capable assistants," he said.
Source: BBC