By Dr. Gopalakrishnan E. A., Principal, School of Artificial Intelligence & School of Computing, Amrita Vishwa Vidhyapeetham, Bengaluru Campus
Most of us have heard the story of the five blind men and the elephant. Our understanding of Artificial Intelligence (AI) is often just as limited and fragmented as the perception each blind man had of the elephant. As American futurist Eliezer Yudkowsky rightly points out, the greatest danger of AI lies in the conclusions people draw about it without truly understanding it.
So, what exactly is AI? According to John McCarthy—considered the father of AI and the person who coined the term Artificial Intelligence—“AI is the science and engineering of making intelligent machines.” Today, AI has permeated nearly every aspect of human life, from the arts to economics, healthcare to outer space exploration.
A pressing question that many are curious about is: What is the next major paradigm shift in AI? The answer appears to be Artificial General Intelligence (AGI). AGI refers to AI systems that are as capable as humans (or even superior) in performing a wide range of tasks, and are far more versatile than current “narrow” AI systems. The rise of Large Language Models (LLMs) has reignited interest in developing AGI, with some experts even claiming that present-day LLMs exhibit “sparks” of AGI. However, the key difference between the current AI systems and the AGI is the ability of the AGI to perform complex tasks involving creativity and logical reasoning indistinguishable from their human counter parts.
It’s worth delving deeper into what we mean by AGI—that is, human-like intelligence. Importantly, the focus should be on the capabilities of such a system, rather than the processes it uses. If we define AGI based on processes, we risk entering philosophical territory with questions like, “Will such systems be conscious?”—questions that are often unanswerable and not empirically verifiable. One of the hallmark traits of AGI will be its ability to generalise across diverse tasks. Another debate concerns the nature of these tasks. When we think of AGI, we should focus on cognitive and meta-cognitive tasks rather than physical ones, which would require robotic components. It is interesting to note that it is much easier to develop an AI that will beat you in chess than an AI that will help you fold your clothes.
Let us try to understand AGI in more detail. Although we lack a standard definition for AGI, the following features are essential for an AI system to be qualified as AGI. The first and foremost requirement is that AGI should surpass (or at least match) human performance across a broad class of cognitive tasks, including but not limited to perception, reading comprehension, reasoning, etc., in different contexts and environments. This essentially means that AGI will have the capacity to generalise or transfer knowledge from one environment (perhaps the environment for which it is trained) to various other contexts or environments, unlike the narrow AI systems that we have currently. Another interesting feature that AGI will have is its capacity to handle challenges that could be quite different from those anticipated by its creators. Again, this is so natural to us humans and not seen in current narrow AI systems. Some of the key features that an AGI should possess are visual perception, audio perception, creativity, problem-solving, and social/emotional engagement.
The most important aspect of AGIs will be their ability to perform counterfactual reasoning like human beings. Counterfactual reasoning involves thinking through “what if” scenarios to understand alternative courses of events to those that actually occurred. Suppose we develop an AI system to predict the occurrence of cardiac disease given a patient’s vitals. The system would be able to forewarn the patient about an incipient cardiac disease if the vitals cross a threshold. However, an AI system equipped with counterfactual reasoning would be able to provide information about alternate scenarios—such as how the occurrence of cardiac disease could have been delayed if the patient had maintained lower blood pressure and exercised regularly. The capability for counterfactual reasoning is inherent to humans, and it is essential for an AI system to explain, adapt, and optimise its decisions (not just predict like current narrow AI). Hence, counterfactual reasoning will be a key capability of an AGI.
Do we already have AGIs among us? The answer is both yes and no. Current LLMs represent an early phase of AGI, as they already outperform unskilled or semi-skilled humans in certain task categories. We are not too far from AGIs that may surpass even skilled humans in many domains.
However, we face two major challenges in developing AGI:
- Designing algorithms that emulate human-like intelligence.
- The enormous computing power required to run such systems, which may render them unsustainable.
Researchers around the world are actively working to overcome these obstacles. The first challenge is being tackled through the integration of causal reasoning into AI systems, along with a growing emphasis on making these systems more explainable and interpretable. The second challenge may be addressed through next-generation computing—such as the development of quantum algorithms and smart materials—to create energy-efficient computing infrastructures.
Let us all hope and strive to witness the advent of AGI in our lifetimes—a development that has the potential to make life on this planet more exciting, sustainable, and meaningful.
