Artificial Intelligence: Is it on the way to surpassing human intelligence?

The emergence of artificial intelligence (AI) has transformed our world in record time, and the question of whether it is poised to surpass human intelligence is fueling much debate. Since Alan Turing’s earliest theoretical conceptions, in which he envisioned machines capable of holding conversations indistinguishable from those of a human, technological progress has been phenomenal. Language models like ChatGPT now appear to pass the famous Turing Test, although this test is now contested regarding its relevance as a criterion of intelligence. With the ability to manipulate large corpora of text, learn new tasks, and provide answers extrapolated from training data, current systems continue to amaze us. Nevertheless, many challenges remain, particularly regarding the ability of AI to perform mathematical reasoning or master more intuitive skills typical of human intelligence. While some see hybrid systems as the key to achieving human-like intelligence, the path to true**artificial general intelligence** is still fraught with challenges.For decades, artificial intelligence (AI) has been at the heart of technological and philosophical discussions. With the emergence of large language models **(LLMs), current examples of this technology have surpassed many of the tests that previous generations failed, but remain limited in some areas. This article explores the current challenges of AI, the question of its intelligence, and the future prospects offered by hybrid systems.**LLMs: Remarkable but Limited Progress 75 years ago, computer science pioneer Alan Turing introduced the idea that a machine could be considered intelligent if it were able to pass itself off as human during a written exchange. Although current LLMs, such as ChatGPT, pass this Turing Test, it remains debatable whether they can be considered truly intelligent. This criterion relies excessively on the ability to produce language, a characteristic that can easily be misleading. A conversational agent that mimics human behavior and generates plausible text can create the illusion of intelligence without possessing real cognitive skills. The Challenges of Artificial General Intelligence

Currently, artificial general intelligence is defined by equalizing or surpassing human cognitive performance in various domains. This includes not only natural language interaction, but also solving novel problems, learning novel tasks, and transferring skills from one domain to another without reprogramming. Although LLMs have demonstrated prowess in some standard assessments, their weaknesses manifest themselves in basic tasks that require non-purely linguistic abilities, such as accurate counting. The Cognitive Limitations of LLMs The ability of LLMs to generate correct answers relies primarily on a huge training corpus encompassing virtually the entire internet. This generates correct answers when a question echoes the data included in the model. However, as soon as one moves away from the corpus, the answers can become uncertain and unreliable. For example, LLMs lack reliable mechanisms for performing precise calculations, which can lead to glaring errors in tasks requiring mathematical reasoning.

Problems Still Difficult for AI

While many challenges have been addressed by current AIs, some human cognitive abilities remain elusive. French engineer François Chollet designed the ARC test battery to specifically assess AI’s ability to tackle truly novel problems. ARC results still show little significant progress, yet these tests do not assess the full spectrum of human intelligence. The difficulties encountered by machine vision systems, such as those used in self-driving cars for reliable object recognition, illustrate these persistent limitations. Towards a Future of Hybrid Systems Despite the impressive advances in LLMs, they still struggle to achieve general artificial intelligence. One promising approach lies in hybrid systems that combine specialized programs with the linguistic power of LLMs. These systems could offer more reliable answers and precise solutions by drawing on the strengths of different modules. This development could potentially pave the way for intelligence that could one day rival that of humans in several aspects.

Notez cet article