Achieving human-level AI may be at least a decade away, according to Meta’s AI scientist, Yann LeCun. Current AI systems, like large language models, fall short of true reasoning, memory, and planning, even though companies like OpenAI market their technologies with terms like ‘memory’ and ‘thinking’. LeCun cautions against the hype, saying these systems lack the deeper understanding required for complex human tasks.
LeCun argues that the limitations stem from how these AI models function. LLMs predict words, while image and video models predict pixels, making them capable of only single or two-dimensional predictions. In contrast, humans operate in a three-dimensional world, able to plan and adapt intuitively. Even the most advanced AI struggles with everyday actions, such as cleaning a room or driving a car, tasks children and teenagers can learn with ease.
The key to more advanced AI, according to LeCun, lies in ‘world models’ – systems capable of perceiving and predicting outcomes within a three-dimensional environment. These models would allow AI to form action plans without trial and error, similar to how humans quickly solve problems by envisioning the results of their actions. However, building these systems requires massive computational power, driving cloud providers to partner with AI companies.
FAIR, Meta’s research arm, has shifted its focus towards developing world models and objective-driven AI. Other labs are also pursuing this approach, with researchers such as Fei-Fei Li raising significant funding to explore the potential of world models. Despite growing interest, LeCun emphasises that significant technical challenges remain, and achieving human-level AI will likely take many years, if not a full decade.
