Product leaders today operate in a field shaped by AI. This is the first in a series of posts that maps that field, starting with the foundational concepts: what AI is, how machine learning works, and what distinguishes the main approaches in use today.
AI vs. Machine Learning
Artificial intelligence refers to systems that can perform tasks typically requiring human reasoning: recognizing patterns, understanding language, making decisions, interpreting images. In 1950, mathematician Alan Turing posed the question of whether machines could think. Five years later, John McCarthy coined the term “artificial intelligence” in a proposal for a summer research workshop that would convene at Dartmouth in 1956.
What has made AI practically useful at scale is not the original idea but the methods developed to implement it. The dominant approach is machine learning, and most of what organizations are deploying today falls under it.
How Machine Learning Works
Traditional software follows rules a developer writes explicitly. If a transaction exceeds a certain amount and originates from an unusual location, flag it. Machine learning works differently. Instead of a developer specifying every rule for every situation, you train a model on examples, and the model finds patterns in those examples on its own. This is not a recent idea. Arthur Samuel coined the term “machine learning” in 1959, demonstrating that a program could improve its own performance through experience rather than explicit instruction. What has changed is the scale at which it can now be applied.
Machine learning encompasses several distinct approaches, each suited to different types of problems.
Supervised learning trains a model on labeled data, where each input comes paired with a known correct output. A fraud detection model, for example, is trained on thousands of past transactions already labeled as fraudulent or legitimate. It learns which patterns are associated with fraud and uses those to flag new transactions in real time. The constraint is the labeled data itself: producing it takes time, domain expertise, and careful quality control, and the model is only as good as what it was trained on.
Unsupervised learning works with data that has no labels. Rather than predicting a known outcome, the model finds structure on its own by identifying which data points are similar, which are unusual, and which variables tend to move together. A retailer feeding customer purchase history into an unsupervised model might get back segments it never defined: frequent buyers of one category, seasonal shoppers, price-sensitive customers. Nobody told it what to look for. The model surfaces structure; someone still needs to decide what it means.
Reinforcement learning takes a different approach entirely. The model learns by taking actions and receiving feedback, a reward when it does well, a penalty when it does not. Over many iterations, it learns which actions lead to better outcomes. Teaching a model to play chess is the standard example: it plays game after game and gradually learns which moves lead to wins. The same principle applies in robotics, where a system learns to navigate through repeated trial and error rather than explicit instruction.
Deep Learning and Large Language Models
Deep learning is a form of machine learning that handles problems too complex for simpler methods. It is based on neural networks, systems designed to loosely mimic how the brain processes information. Face recognition illustrates how this works: the system starts by detecting basic shapes and edges, then identifies features like eyes and a nose, and combines those into a match. This layered approach is what makes deep learning effective for tasks involving speech, medical imaging, and language, where the patterns are far too complex to specify manually.
Large language models (LLMs) are deep learning models trained on large volumes of text. They learn by predicting what comes next in a sequence, and at sufficient scale, develop the ability to generate, summarize, translate, and reason about language. The technical foundation most modern LLMs are built on is the Transformer architecture, introduced in a 2017 paper by researchers at Google. Transformers process all words in a sequence simultaneously and learn which words are most relevant to each other, giving the model a richer understanding of context than earlier approaches allowed. ChatGPT, Gemini, and Claude are all built on this foundation.
What Comes Next
The posts that follow will go deeper on how these systems are built and evaluated, where they perform reliably and where they fail, and what that means for product and business decisions.
