AI is moving towards AGI – is this what the world is waiting for?

Dr Keith Darlington

Introduction
AI (Artificial Intelligence) has long been touted as a saviour for humanity’s problems, and it’s not difficult to see why. AI excels at identifying patterns in massive, complex datasets that far exceed human capacity. For example, AI is being used in thousands of diverse applications, from smartphone notifications to complex hospital operations support for patients, and is also being utilised in highly complex tasks such as simulating protein folding, climate modelling, and drug discovery. In doing so, AI is infinitely better than humans at many tasks.

Much of its success in recent years is due to the quantum leap impact of large language models (LLMs) like ChatGPT and Gemini. They provide human-level conversational knowledge on every conceivable subject. This has led to a flurry of activity, prompting many to believe that AI is now rapidly moving towards AGI (Artificial General Intelligence). This stage can match a human at any intellectual task. AI has achieved many milestones and feats in the last decade, but this would be the “Holly Grail”. In this article, I discuss AGI and show that, although LLMs may not be sufficient on their own, the level of investment combined with other AI techniques is likely to make this happen within the next few decades.

The limitations of AI

AI, despite its remarkable achievements to date, remains predominantly single-purpose. This means that it can be used for single tasks, such as diagnosing a patient’s disease, conversing with someone via a chatbot, or driving an autonomous vehicle. They are all important uses of AI, but they cannot simultaneously understand and apply knowledge across various domains, enabling them to solve more general problems as humans do. Humans can handle diverse intellectual tasks and transition seamlessly from one task to another. For example, our driving knowledge enables us to drive many models and sizes of cars that we have never encountered before, because we have reasoning capabilities that will allow us to understand similarities and differences, as well as abstract features. Furthermore, we could combine this knowledge with our financial and domestic knowledge to decide what car to buy if we wanted to make a change. This is what is required for a transition to AGI: the ability to use and combine intellectual understanding from a variety of tasks.

The Transition from AI to AGI

According to many, we are getting closer to AGI than ever before. Thanks in large part to LLMs that have extraordinary learning capabilities in spoken languages. LLMs are trained on massive amounts of text, images, and video, primarily sourced from the Internet, enabling them to learn patterns and relationships in language. This means they possess in-depth knowledge of almost everything. LLMs have led to a significant leap in AI applications since ChatGPT’s debut at the end of 2022.

Ferocious competition in Silicon Valley R&D has paved the way for massive scale-up improvements that continue unabated. However, it is unlikely to be enough, because true AGI will require solving further challenges. As Gary Marcus, a leading academic critic of AI, wrote: “Anybody who thinks LLMs are a direct route to the sort of AGI is kidding themselves.” A move to AGI will likely necessitate entirely new architectures that incorporate breakthroughs in reasoning and common sense, as well as deep integration with the physical world, known as embodiment.

Reasoning and Common Sense
Rapid progress in LLMs indicates that many of them can reason in ways that approach human standards. However, they still struggle with human-level common sense despite much research effort. Human common sense refers to our understanding and knowledge about the world acquired through everyday experiences and interactions. We take common sense for granted, yet we acquire a great deal of it during our lives. This includes facts, such as men shave, dogs don’t use the Internet, and when two people are talking to each other, they are talking in the same language.

The difficulty with incorporating human-level common sense is that machines will always lack the same real-world understanding that we do. Social norms change regularly, and machine common sense will always struggle with this because it lacks the real-world understanding of such events. LLMs empower AI with an enormous capability to understand the relationships between words and language, but is that enough to understand the world as we do, and does it matter in the pursuit of achieving AGI? To acquire common-sense knowledge and a deeper understanding of the real world, it may be necessary to have some form of physical embodiment in the real world.

Embodiment

Embodiment refers to a (robotic-like) entity having a physical or virtual body that interacts with an environment through sensors to perceive and act on events in the same way as we humans do. This could help the entity learn common sense from its interaction with others as well as relate to human emotions in a more meaningful way. Furthermore, embodiment enables the entity to gain context from its interactions, providing it with a deeper understanding of each situation it encounters.

Many in the AI community believe that embodiment is a necessary requirement for the move to AGI, as they consider intelligence to be more than just cognitive processes in the brain, but also influenced by the physical interactions an entity has with its environment. Both embodiment and common sense have been active areas of research for many years; however, their focus will now shift to their incorporation into the structures of LLMs.

Will AGI Happen?

There is a broad consensus that supports the belief that AGI will happen in the next 5 to 20 years. For all the challenges, the huge investments and activity at all levels of business and government mean it’s likely to happen. It may take persistence and many fresh ideas and minds, but so did the invention of the automobile engine. The development of the combustion engine, by Nikolaus Otto and later by Daimler and Benz, involved a significant amount of trial and error by many researchers over several years before it became commercially viable. The same may happen with AGI. Like many great inventions, will require persistence and may involve various approaches, including trial and error, but it will ultimately succeed.

Dr Keith Darlington is a retired AI university lecturer and author of five books on AI and computing topics, as well as over 70 magazine articles on AI and related subjects