Meta’s AI Chief: AGI is Still Far Away

Artificial general intelligence (AGI) is the ultimate goal of many AI researchers, who envision a machine that can perform any intellectual task that a human can. However, achieving AGI is not an easy task, and it may take much longer than expected.

In a recent interview with the Wall Street Journal, Meta’s chief AI scientist Yann LeCun said that he does not expect AGI to be developed in the next five years. He believes that AI systems still have a long way to go before they become sentient or have common sense.

LeCun, who is widely regarded as one of the pioneers of deep learning, said that he is more interested in creating artificial intelligence that can augment human capabilities and solve real-world problems. He said that he thinks of AGI as “a very ambitious goal” that requires a lot of research and experimentation.

LeCun also said that he does not think that current AI models, such as ChatGPT-4, are close to achieving AGI. He said that these models are good at generating realistic text and images based on specific prompts, but they do not have any understanding or reasoning behind their outputs. He said that these models are still limited by their data and algorithms.

LeCun’s views are in contrast with some other prominent AI researchers, such as Demis Hassabis, the CEO of Google’s DeepMind. Hassabis has claimed that AGI might be developed within a decade, and that he expects very capable and general systems in the next few years. Hassabis has also argued that AGI is not far away from being achieved.

However, some experts have expressed concerns about the potential dangers of unleashing powerful AI systems into the world without proper regulation and oversight. They have called for more ethical and social considerations when developing advanced AI technologies. They have also warned about the possibility of creating artificial superintelligence (ASI), which could surpass human intelligence and control in all domains.

Therefore, it seems that there is no consensus on when or how AGI will be achieved. It remains to be seen whether LeCun’s prediction will come true or not. What is clear is that AGI is one of the most fascinating and challenging topics in AI research today.

One of the main challenges of creating AGI is to endow machines with common sense, which is the ability to understand and reason about the world in a human-like way. Common sense is essential for AI systems to handle novel situations and tasks that they have not encountered before. However, common sense is not easy to define or measure, and it is not clear how to teach it to machines.

Artificial general intelligence (AGI) researchers have proposed different approaches to develop common sense in AI systems, such as using large-scale knowledge graphs, natural language understanding, computer vision, and reinforcement learning. However, none of these methods have been able to achieve human-level common sense so far. Moreover, some experts have argued that common sense is not a single skill, but a collection of many different abilities that depend on the context and the task.

Therefore, creating AGI that can exhibit common sense is still an open and difficult problem. It may require new breakthroughs in AI research and engineering, as well as interdisciplinary collaboration with other fields, such as psychology, neuroscience, and philosophy. It may also require a better understanding of how humans acquire and use common sense in their everyday lives.

The AI Act faced several challenges and criticisms, especially after the emergence of ChatGPT, which showed the potential and the limitations of AI, as well as the difficulties and dilemmas of regulating it. Some of the challenges and criticisms were: The AI Act was too vague and broad, and did not provide clear and consistent definitions and criteria for the risk categories and the requirements.

For example, it was unclear how to determine the level of risk of a foundation model like ChatGPT, which can serve both the benign and the malignant, depending on the user’s input and output. It was also unclear how to ensure the transparency, safety, and human oversight of such a system, which can generate unpredictable and diverse texts on any topic.