By Gopal Gupta, Professor of Computer Science and Director of the Center for Applied AI and Machine Learning, University of Texas at Dallas
With recent advances in generative artificial intelligence (GenAI), the question that keeps arising is whether artificial general intelligence (AGI)—AI that exhibits human-like intelligence—is within sight.
Broadly speaking, intelligence has two components: learning and reasoning. Learning involves the use of our senses and language capabilities to recognize various patterns over time. The learned patterns are stored in our mind as associations between the pattern and its corresponding name (e.g., imagery of a flower). Once the patterns are learned, we use pattern matching for recognizing them later. We also observe complex patterns and generalize them into rules (e.g., if I release an object that I am holding, it will fall to the ground). These rules are also stored in our mind. These associations and rules constitute our knowledge. Of course, we may also be taught this knowledge by others who have already learned it.
As our brain receives sensory inputs, we apply pattern matching (“this object is a chair”) to acquire knowledge about our surroundings. Next, we use reasoning (i.e., apply the learned rules) to check the consistency and completeness of this knowledge. We also use reasoning to draw new conclusions, and take actions based on those conclusions. To illustrate, consider that as a child we may hear a sharp, high-pitched sound that our parents may tell us to be the sound of a siren. This association is recorded in our minds. We may also record in our mind—as knowledge—the learned rule of thumb that, in the plains of North Texas, a siren warning during inclement weather means a tornado is close by and that we should immediately seek shelter. The next time we hear a siren sound during inclement weather, we will immediately recognize it—via the pattern-matching process—and subsequently use our reasoning ability to conclude that we should immediately seek shelter. We might further employ more complex reasoning steps to find the safest place in our house for shelter. While this is a simplified perspective, our day-to-day lives nevertheless depend critically on both learning/pattern-matching (to store associations and upon seeing a pattern, retrieve its name) as well as reasoning (to process learned rules).
We humans are not the only ones who use pattern matching to make sense of the world around us. All other animals do the same. In fact, nature has equipped animals with significantly more advanced pattern-matching capabilities than humans. Eagles can see significantly better than us, a dog’s sense of smell and sound is significantly more advanced than a human’s, a camel can sense water tens of miles away, and so on. Thus, nature’s evolutionary processes have produced extremely sophisticated pattern-matching capabilities.
Nature has likewise equipped all animals with instincts. An instinct represents an automatic response to some stimulus, where the stimulus is produced via sensing and pattern matching. It stands to reason then that had pattern matching and instinctual reaction alone been sufficient to produce sophisticated human-like intelligence, it would have emerged in other species also through evolution. However, this is decidedly not the case, even though evolution has led to non-human animals possessing very sophisticated pattern-matching abilities. Rather, evolution took a different path to produce higher intelligence that we humans possess—it incorporated reasoning in the human brain. Other (non-human) animals can reason as well but in a very limited form. As we go up the evolutionary chain, we see that reasoning capabilities improve, culminating in us humans. As an example, an albatross is unable to reason that if it does not find its newborn chick in its nest, then it should search in the vicinity. This is precisely what reasoning capability can achieve that instinctual, pattern-matching-based response may not.
Reasoning, one could therefore conclude, is indispensable for AGI. However, popular AI systems of today are primarily based on machine learning, i.e., automated association-based pattern-matching. They lack (automated) reasoning capabilities, i.e., the ability to represent rules and apply them. Automation of reasoning has been vigorously pursued by AI researchers over the last 60+ years. However, the task has turned out to be quite hard due to well-researched problems of undecidability, non-monotonicity, and incompleteness that have stood in the way of this automation.
As we know from our own experiences, pattern matching is easy. We can identify objects, sounds, smells, etc., effortlessly. However, reasoning is much harder and requires significantly more effort, and we tend to shy away from it. In his book “Thinking, Fast and Slow,” Nobel Laureate Daniel Kahneman has referred to learning and pattern matching as fast (or reflexive, or System 1) thinking, and reasoning as slow (or deliberative or System 2) thinking. Reasoning or System 2 thinking in humans is slow for the same reason that automation of reasoning has proven to be difficult. So even if we manage to automate reasoning and combine it with machine learning in the best possible way, the resulting AGI will face the same challenges that we humans face with respect to reasoning. It follows then that it may not be possible to build intelligent systems that can outperform humans. At best, we can build an AI system that outperforms humans simply because humans often make mistakes due to being tired, inattentive, emotional, etc., that an AI system will not. Perhaps the best we could do is to build an AI system that can match our collective human wisdom and knowledge.
To reinforce the points made above, let’s look at another AI technology that has received significant attention like generative AI, but has yet to fulfill its promise: autonomous driving. We were promised fully autonomous cars by 2020, but that has not yet materialized. Apple, for example, killed its autonomous car project two years ago after sinking hundreds of millions of dollars on it. Robotaxis do employ some forms of reasoning, however, they are geo-fenced and still have run into problems. The reason, again, is that autonomous driving relies principally on machine learning, which by itself cannot emulate human-like intelligent behavior. Human drivers, in contrast, use both pattern matching and reasoning to drive a car. We recognize objects in our current surroundings (via pattern matching), construct a model of the observed world, and then apply reasoning over these objects to make driving decisions in real time using knowledge acquired over a lifetime. For example, we will apply common rules of thumb for reasoning such as, “I see a policeman at the intersection, I must follow his commands if he gives any”. Additionally, we can draw conclusions even when the information we have is incomplete. To fully automate driving, we must make use of reasoning also, rather than rely mainly on machine learning. Our knowledge and reasoning capabilities allow us humans to navigate an imperfect, ever-changing physical world, even when our knowledge is incomplete.
Likewise, large language models (LLMs) such as ChatGPT and other generative AI systems are prone to hallucinations because they also rely on pattern matching alone. No explicit representation (knowledge) is created by LLMs, and no explicit reasoning is involved in the LLM technology: at the most basic level, it is all statistical pattern matching and not much more. On the ARC AGI 3 benchmark that requires one to reason, humans score 100% while LLMs score below 1%. True reasoning is an additional, crucial tool that humans use for decision-making and navigating life. Adding true reasoning into LLMs, unfortunately, is far more difficult to achieve than basing LLMs on machine learning, pattern matching, and statistics alone, as is done today.
What does all this mean for the finance and accounting industry? Accounting is fundamentally an endeavor requiring high accuracy. It relies on regulatory rules, mathematical certainty, and precise logical relationships. Finance similarly depends on optimization and quantitative modeling and is subject to legal constraints. Correctness and precision are more important than intuition fluency for these areas so one must be cautious in using generative AI technology. Applications that principally involve pattern matching, such as document understanding, knowledge retrieval (“explain deferred tax liabilities”), report summarization, drafting documents, coding assistance, etc., will work well. For applications such as tax computations, financial statement preparation, accounting rule enforcement, portfolio optimization, autonomous accounting, etc., LLMs may not be the most appropriate technology, as accuracy of the results they produce cannot be guaranteed. If one still insists on using LLMs for these purposes, then one must check the output produced by the LLM extremely carefully.
In conclusion, machine learning technology alone cannot lead to AGI. Large language models, while amazing in their capabilities, will not be able to reach human level intelligence given the current technology that they are based on. Automated reasoning is indispensable for achieving AGI. However, automated reasoning comes with its own set of limitations—limitations that humans also face—which prevents us from building AI systems that can exceed human capabilities in any meaningful way. So, at best, we can hope to press forward and develop AI technology that combines machine learning, LLMs, and automated reasoning—what is referred to as neuro-symbolic AI—to build AGI systems that are as good as the unerring, nonstop-working human or that can match collective human wisdom. Building systems that exceed human intelligence—superintelligence—is likely an unachievable dream.
Gopal Gupta is a professor of computer science and the Director of the Center for Applied AI and Machine Learning at the University of Texas at Dallas where he conducts research in automating commonsense reasoning, explainable machine learning, and their use in developing trustworthy AI applications.
Learn more about CAIA Association and how to become part of a professional network that is shaping the future of investing, by visiting https://caia.org/
