AI and Science of Reasoning(5)

Sherlockian Way of Thinking

3. Generative AI and Art of Prompt Engineering

How AI Learns: Supervised and Unsupervised Learning

For an artificial intelligence system to perform any task effectively, it must first undergo a process of learning. Before it can act autonomously, the system must be trained—exposed to vast amounts of data to discern patterns, establish criteria, and cultivate judgment. Broadly speaking, there are two principal modes of learning: supervised learning and unsupervised learning.

In supervised learning, the AI learns from labeled data—that is, data that already carries the correct answers. Imagine showing the model countless images of cats, each time telling it, “This is a cat.” By repeating this process, the AI gradually learns to identify the defining features of a cat: the shape of the eyes, the curve of the ears, the texture of the fur. Eventually, it becomes capable of recognizing cats on its own, even without being told the answer.

There are two principal modes of learning: supervised learning and unsupervised learning.

Unsupervised learning, by contrast, operates without labeled data. In this case, the model might be shown images of cats, dogs, tigers, and foxes—without any names attached. Examining these examples, the AI begins to cluster similar data together: “These images seem related,” it infers. Perhaps cats and tigers are grouped for their similar eyes and facial structures, while dogs and foxes are clustered for their pointed ears. In this way, the AI uncovers hidden structures within data, identifying relationships and extracting features purely from internal similarity.

Conceptually, unsupervised learning resembles inductive reasoning: it observes instances and infers general patterns. Many logicians and AI theorists indeed describe induction as the human cognitive process most akin to machine learning. Yet the two are not identical. Unsupervised learning relies on statistical modeling—clustering and dimensionality reduction based on numerical similarity—while human induction generalizes through linguistic and logical abstraction.

In the early stages of AI research, most systems depended heavily on supervised learning. Clear answers meant clear direction. But as the world grew more complex, and as datasets swelled beyond human labeling capacity, supervised methods reached their limits. From that point, unsupervised learning began to rise in importance. Its ability to discover meaning in unclassified data and to create new taxonomies autonomously became essential—not only for deep learning but also for generative AI and modern recommendation systems.

How AI Learns from Experience: Reinforcement Learning and RLHF

Among the many ways machines learn, reinforcement learning (RL) most closely mirrors the human way of learning through experience. Unlike supervised or unsupervised methods that rely purely on data observation, reinforcement learning allows an AI to act, experiment, fail, and improve through feedback from its environment. It is, in essence, learning by doing.

Reinforcement learning (RL) most closely mirrors the human way of learning through experience.

The structure is simple: the AI observes a state, takes an action, and receives a reward. Positive outcomes yield higher rewards; poor choices bring penalties. Through countless iterations, the AI learns to select actions that maximize its cumulative reward. Imagine a robot navigating a maze: at first it moves randomly, but over time, by noting which directions bring it closer to the goal, it develops a strategy—a policy—to escape more efficiently. This is reasoning through trial and error.

Reinforcement learning has also become central to generative AI, particularly in language models. Here, it takes a human-centered form known as Reinforcement Learning with Human Feedback (RLHF). In RLHF, human evaluators review the model’s responses and rate them according to quality—clarity, helpfulness, tone, or factual accuracy. The AI then learns to prefer responses that receive higher human approval, refining its behavior accordingly.

This method does more than enhance technical performance. By incorporating human judgment and values, RLHF teaches AI to emulate human sensibility—to express nuance, empathy, and appropriateness. In this way, reinforcement learning combined with human feedback forms the bridge by which AI evolves from a calculating engine into a conversational presence.