Human level is still a distant vision | Meta AI image generator | WhatsApp Business | Meta AI app | Turtles AI
Human-level AI research is still a long way off. Meta’s Yann LeCun advances the idea of “world models” as a key to progress. Current AI systems, while promising, do not reach human cognitive capabilities.
Key points:
- Yann LeCun states that human-level AI is still decades away.
- Current AI models lack true understanding of the world.
- “World models” could revolutionize the approach to AI.
- The research is complex and requires significant time and resources.
In the current debate on AI, statements by experts such as Yann LeCun, head of AI at Meta, lead to deep reflection on the real capabilities of current models. According to LeCun, although progress has been made, we are still far from achieving an AI that can truly think, plan and reason like a human being. During a recent talk at the Hudson Forum, LeCun questioned the optimism of some pioneers in the field, such as Elon Musk and Shane Legg, who claim that human-level AI is imminent. His view is clear: true AI requires machines capable of understanding the world, equipped with memory, intuition and common sense, capabilities that current systems do not possess.
Current language models, such as ChatGPT’s, work primarily through prediction of the next token, a process that, although sophisticated, remains limited. LeCun explains that while humans learn simple tasks such as clearing the table or driving a car in a few days, more advanced AI models struggle to operate effectively in physical situations, despite training on huge datasets. This limits their practical applicability and ability to solve everyday problems.
To improve, LeCun proposes the development of “world models,” an approach that aims to create three-dimensional representations of reality that can be used to plan and predict actions in the real world. These models could absorb information more efficiently than current language models and, although they require significant computational power, promise to offer a more complete and useful view of the surrounding environment. Models of the world, as explained by LeCun, are based on building a mental image of how objects and situations interact, allowing effective action sequences to be planned.
Many research laboratories are currently investing in the concept of world models. One example is World Labs, a startup founded by renowned researchers including Fei-Fei Li, which recently secured $230 million in funding to delve into this new AI frontier. Similarly, OpenAI has already indicated its intention to explore similar approaches in its Sora video content generator, although details are still vague.
However, LeCun warns that the road to realizing these models is still long and winding. Despite the promise and potential of this technology, there are complex challenges ahead, making it difficult to make predictions about when such systems might become a reality. “It will take years, if not a decade,” LeCun concluded, stressing the need for patience and perseverance in an ever-evolving field.
As discussions about human-level AI intensify, the idea of world models emerges as a possible turning point, but it remains a challenge to be approached with caution and foresight.
