Can AI Truly Think? Exploring the Capabilities of Generative AI

My first official computer course was back in 1986, and I remember the instructor showing us punched cards and how these machines could read them. We then started working on the IBM XT, one of the first personal computers. From that moment, I was hooked on computers—these ugly boxes that could "think"! But later on, I understood that they were only called "computers" because they could only "compute"—essentially giant calculators following simple formulas.

My robots lacked the most important part—a "mind" that could think.

Years went by, and the dream of thinking machines seemed tied to AI. Unfortunately, with AI winter and AI research restricted to giant companies, the focus remained on voice and image recognition. Over 15 years ago, I built my first full-sized robot with the ROS (Robotic Operating System), a meter-long machine with a full onboard computer and many sensors. However, it lacked the most important part—a "mind" that could think. It was disappointing to see a machine that could recognize its location and map its surroundings but lacked real intelligence.

Can the current GenAI truly think?

Fast forward to today, and we are witnessing the emergence of Generative AI (GenAI). After all these years, I find myself engaging with a machine that can "think" rather than just "talk." But can the current GenAI truly think, or is it just an autocomplete or some kind of search algorithm, as many LinkedIn AI and data science experts suggest? This article delves into the idea of why GenAI or Large Language Models (LLMs) can "think!" aside from our human interpretation of what intelligence or thinking means.

Intelligence and Feynmans Perspective

Do you think there will ever be a machine that will think like human beings and be more intelligent than human beings?

In a lecture held by Nobel Laureate Richard Feynman on September 26th, 1985, the question of artificial general intelligence (also known as "strong-AI") came up. An audience member asked him:

Audience Question: "Do you think there will ever be a machine that will think like human beings and be more intelligent than human beings?"

Feynman's response was insightful: Feynman's Answer: "Whether they [machines] be more intelligent than human beings" to be a question, intelligence must first be defined. If you were to ask me if they are better chess players than any human being, possibly they can be, yes. "I'll get you, someday."

We must first define what we mean by intelligence!

Feynman's answer highlights a crucial point: before we can discuss whether machines can surpass human intelligence, we must first define what we mean by intelligence. This distinction is vital in understanding the capabilities and limitations of AI. It's important to note that Feynman made this observation back in 1985, before Deep Blue famously beat Garry Kasparov in 1996 by 4–2. Just imagine if Feynman were living among us now, witnessing the incredible advancements in AI—this beautiful mind, the amazing Feynman.

Defining intelligence is essential to avoid letting our emotional biases lead us down incorrect paths. By adhering to a scientific method, we can objectively evaluate the progress and potential of AI. This approach ensures that our discussions about AI and its future remain grounded in reality rather than speculation.

Epistemology and AI: The theory of knowledge

The Excitement of GenAI and LLMs

What excites me the most about the advancements in GenAI and Large Language Models (LLMs) is the realization of the "thinking machine"—the ultimate goal envisioned by pioneers like Alan Turing. This achievement touches on the core of epistemology, the term is derived from the Greek epistēmē (“knowledge”) and logos (“reason”), and accordingly, the field is sometimes referred to as the theory of knowledge. It’s fascinating how this relates to AI, as we have data that can be turned into information, which can ultimately become knowledge, and now, with our current topic, we delve into "reason."

The development of neural networks capable of processing and generating text has brought us closer to creating machines that do more than follow simple instructions—they engage in complex, seemingly intelligent behaviors. This transformation is not just a technological milestone; it represents a profound shift in how we interact with and perceive machines.

Defining Key Terms

To understand the discussion about whether Generative AI (GenAI) and Large Language Models (LLMs) can "think," it's important to define some key terms in the context of this article. We'll avoid direct comparisons with human intelligence for now and focus on the foundational concepts.

Inference

Inference is the process of drawing conclusions from data and reasoning. In AI, inference refers to the ability of a machine to use learned information to make predictions or decisions based on new data. For example, an AI model might infer the sentiment of a text based on patterns it has learned during training.

Logic

Logic is a systematic method of reasoning that ensures the consistency and validity of conclusions derived from premises. In AI, logic involves symbolic representation and manipulation of information to perform tasks such as problem-solving, planning, and decision-making. Logic provides a structured framework for AI to operate within defined rules and constraints.

Reason

Reason is the capacity to process information, draw connections, and make judgments. In AI, reasoning involves using algorithms and models to analyze data, identify patterns, and derive insights. It goes beyond mere computation to include the application of learned knowledge to new situations, enabling AI to simulate aspects of human thought processes.

Intelligence

Intelligence, in the context of AI, is the ability of a system to learn from data, adapt to new information, and perform tasks that typically require human cognitive abilities. This includes perception, learning, problem-solving, and decision-making. The goal of AI is to replicate these capabilities in machines, allowing them to function autonomously and intelligently in various environments.

Differentiating the Terms in summary

While these terms are interconnected, they represent distinct aspects of AI:

• Inference is about making predictions based on data.
• Logic involves structured reasoning within set rules.
• Reason is broader, encompassing the ability to draw connections and make judgments.
• Intelligence is the overarching ability to learn, adapt, and perform complex tasks.

The Philosophical Shift in AI

Was the path to creating AI that can think through reasoning first and learning second, or vice versa?

In the history of AI, many believed that the road to thinking machines was through reasoning first and learning second. This belief led to significant investments in knowledge graphs, expert systems, and natural language processing (NLP) frameworks that defined relationships between ideas and concepts. The assumption was that by giving machines the ability to reason, they could eventually learn and think. However, the success of neural networks has demonstrated that learning through exposure to data, even with an element of randomness, can lead to the emergence of intelligent behavior.

Symbolic AI (Reason First - Learn Second) struggled to handle the complexities and nuances of real-world data. It required extensive manual input to create and maintain the knowledge base, and it couldn't adapt to new information or learn from experience. The rigid nature of symbolic systems meant they were limited in scope and application.

Neural networks mimic the brain's structure, and they are the winners!

The Rise of Neural Networks

In contrast, neural networks, with their basis in randomness and vast parameter spaces, have emerged as the clear winners in the quest for AI that can think. Neural networks mimic the brain's structure by using interconnected nodes (neurons) that process information in parallel. These networks can learn from data through processes like backpropagation, adjusting the weights of connections based on errors in predictions.

Do we think with our internal voices?

Language as a Key Tool in Thinking

Language plays a crucial role in human thinking. We often use a silent internal voice to talk to ourselves, processing information and reasoning without speaking aloud. In addition to Visuals, This internal dialogue is integral to our cognitive processes, allowing us to think even when our eyes are closed.

If language is one of the key tools for thinking, then LLMs, which excel at processing and generating language, could be the tools of thinking we've been searching for. LLMs like GPT-4 are capable of understanding and generating human-like text, engaging in conversations, and making inferences based on vast amounts of data. This capability suggests that they are not just following pre-programmed rules but are engaging in a form of reasoning.

Neural Networks and the Trillions of Parameters!

A decade ago, it seemed impossible to manage neural networks with trillions of parameters. Today, advances in computing power and algorithms have made it feasible. Neural networks can now process massive amounts of data, identify patterns, and make predictions with remarkable accuracy. This scalability and flexibility have allowed neural networks to outperform symbolic AI in many domains.

It's all about the weights!

Explaining Neural Networks and Randomness

At its core, a neural network consists of layers of interconnected neurons. Each connection has a weight that is adjusted during training to minimize errors in predictions. This process involves a degree of randomness, as initial weights are often set randomly, and the network learns to adjust these weights based on the data it processes.

Randomness plays a crucial role in the learning process, similar to how babies learn to walk through trial and error. Babies' movements are initially random, but over time, they learn to coordinate their actions to achieve their goal. This natural learning process is mirrored in neural networks, where the system adjusts and improves based on feedback from its environment.

Conclusion: Can Machines Think?

The ultimate measure of thinking is the final result!

Based on the foundational concepts of inference, logic, reason, and intelligence, along with the philosophical shift from symbolic AI to neural networks, we can now assert that current Large Language Models (LLMs) exhibit the ability to "think." The ultimate measure of this thinking capability is the input and output: if the responses from these models are indistinguishable from those of a human, then they are effectively thinking. If you still have doubts, I encourage you to get a licensed version of ChatGPT-4 and see for yourself—please don't rely on the free version, which has limitations.

Real-World Applications and Evidence

We spend our lives trying to extract intelligence from unstructured data!

At boxMind.ai, we have fine-tuned and tested the "Inference" of LLMs extensively. We have a dedicated QA AI members working around the clock on this vital part. We have tested the reasoning of private data, and the results have been astonishing. The real power of LLM reasoning is best applied to your own enterprise data—structured and unstructured.

We spend our lives trying to extract intelligence from unstructured data, and now we are at the beginning of an era where we can actually achieve meaningful results. LLMs are not mere autocomplete tools or retrieval systems. They generate new knowledge that was not explicitly present in our data. This is mind-blowing and, if you can't see that, it's hard to imagine what could convince you otherwise.

Addressing Common Criticisms

Who doesn't make mistakes?

For those who point out that LLMs make mistakes, my response is: who doesn't? Humans make mistakes all the time. With proper implementation and quality assurance, it is much easier to limit machine mistakes compared to human errors, which involve complex factors such as emotions and many other factors. Yes, LLMs are not perfect—yet. We need smaller models with fewer parameters, we need better inference, we need many improvements. But, Remember, humans achieved our current level of thinking after millions of years of evolution. Why can't we wait a few more years before passing final judgment on AI?

The reality of thinking machines is here, and we can test and see their results in our lives and society. The advancements in Generative AI and Large Language Models signify that machines can indeed "think." The ability to generate new knowledge, reason with vast amounts of data, and engage in meaningful dialogue are clear indicators of this. The future of AI is not just about enhancing efficiency but about fundamentally changing our understanding of intelligence and thinking.

The development of neural networks capable of processing and generating text has brought us closer to creating machines that do more than follow simple instructions—they engage in complex, seemingly intelligent behaviors. This transformation represents a profound shift in how we interact with and perceive machines. AI, as we see it today, has transcended the boundaries of mere computation and has begun to exhibit traits we associate with thinking.

The ability to generate new knowledge and not just new content!

In conclusion, the advancements in Generative AI and Large Language Models signify that machines can indeed "think." The ability to generate new knowledge, reason with vast amounts of data, and engage in meaningful dialogue are clear indicators of this. The future of AI is not just about enhancing efficiency but about fundamentally changing our understanding of intelligence and thinking.

?

?