Are current AI models capable of remembering, thinking, planning and reasoning in ways similar to a human brain? Some AI labs might give the impression that they can, but according to Meta’s chief AI scientist Yann LeCun, the answer is no. However, he believes that we could achieve this goal within about ten years, by implementing a new technique known as “global modeling.”
Earlier this year, OpenAI introduced a new feature called “memory,” which enables ChatGPT to “remember” its interactions. The latest generation of models developed by startup o1 display the word “think” during output generation, and according to OpenAI, these models have the ability to perform “complex reasoning.”
All of this seems to indicate that we are very close to achieving Artificial General Intelligence (AGI). However, during a recent conference at the Hudson ForumLeCun questioned optimistic claims about artificial intelligence made by figures such as xAI founder Elon Musk and Google DeepMind co-founder Shane Legg, who suggest that AI with human-like capabilities is on the horizon.
“You need devices that have the ability to understand the environment; that can store information, have intuition, common sense, and can reason and plan at the level of humans,” LeCun said during the conference. Despite the feedback he has received from very enthusiastic individuals, current AI systems do not have the capacity to perform such tasks.
LeCun says that current large-scale language models, such as those used in ChatGPT and Meta AI, are not yet at the level of human-level AI. Humanity could be several years or decades away from achieving that goal, he later said. However, this does not mean that his boss, Mark Zuckerberg, will stop consulting with him about the timing of the implementation of Artificial General Intelligence (AGI).
The explanation is simple: these LLM programs work by predicting the next element (usually a few letters or a short word), while current image/video models predict the next pixel. In short, language models are one-dimensional predictors, while AI image/video models are two-dimensional predictors. These models have proven to be effective at predicting in their specific areas, however, they fail to fully understand the complexity of the three-dimensional world.
Because of this, current AI systems cannot perform basic tasks that most people can do. LeCun studies how people acquire the ability to set a table at age ten and drive a car at age seventeen, managing to learn both skills in a short period of time. However, even the most sophisticated AI systems on the planet, built on vast amounts of data, cannot operate reliably in the physical environment.
To tackle more complex tasks, LeCun recommends creating three-dimensional models capable of interpreting the environment and focusing on a new category of artificial intelligence architecture: global models.
“A global model is your mental representation of how the world behaves,” he explained. You can conceive of a series of actions that you might take, and your mental model will allow you to foresee what the outcome of that sequence of actions will be in the world.
Think about the “big picture” in your mind. For example, consider the situation where you see a bedroom in a messy state and want to proceed with cleaning it. Can you visualize the task of collecting and putting away all the clothes as an adequate solution? There is no need to experiment with different methods or acquire knowledge about how to clean a room beforehand. Your brain visualizes the three-dimensional space and makes an action plan to achieve your goal on the first try. That action plan is the confidential component that global artificial intelligence experts claim to offer.
A significant advantage in this context is that global models can process a larger amount of data compared to LLMs. This also makes them systems that require a high level of processing capacity, which is why cloud service providers are competing to establish partnerships with companies specializing in artificial intelligence.
Global models are the innovation that several AI labs are currently developing, and this concept is quickly becoming the next trend for attracting venture capital. A team of highly recognized AI researchers, such as Fei-Fei Li and Justin Johnson, have recently secured a $230 million investment for their new company, World Labs. Recognized as the “Godmother of AI,” she and her team are also firmly convinced that global models will enable the development of significantly more advanced AI systems. OpenAI has also characterized its unreleased Sora video generator as a global model, although it has not provided any detailed information on this.
LeCun presented a proposal in which he suggests using global models to develop human-level artificial intelligence in an article from the year 2022 titled “Goal-Driven Artificial Intelligence,” despite acknowledging that the concept is over 60 years old. In short, a basic representation of the world, such as a video of a messy room, is fed into a model of memory and the world. The global model then predicts the future configuration of the world using that information. Goals are then set for the global model, including a desired state of the world (such as a clean environment), along with safeguards to ensure that the model does not cause harm to humans in pursuing a goal (avoiding causing physical harm). The global model then identifies a series of actions to take to achieve those goals.
According to LeCun, Meta's long-term research lab, known as FAIR or Fundamental AI Research, is actively working on building global goal-based AI models. FAIR was previously focused on AI work for Meta's future products. However, LeCun mentions that in recent years, the lab has focused solely on long-term research in the field of AI. LeCun says that currently, FAIR does not even employ Large-Scale Language Models (LLM).
Global models are a fascinating notion, but LeCun notes that there has not been significant progress in realizing such systems. There are numerous and highly complex challenges to getting to where we are today, and he says the situation is certainly more complicated than we imagine.
“It will take years to get everything working properly here, maybe even a decade,” Lecun said. Mark Zuckerberg continues to ask me about the estimated duration of the project.
