AI More Fascinating Than Ever
Let’s take a moment to explore the AI hype together and try to understand what it is really about. What exactly is artificial intelligence, and what lies behind this phenomenon? This article will only scratch the surface, as the goal is to provide an initial overview rather than dive deeply into the many complex research areas of AI. Quite apart from the fact that AI is an extremely complex subject, there are aspects of it that are sometimes impossible to fully understand—even for experts. More on that later.
We will work our way from the basic concepts of AI to large language models (LLMs). Other exciting areas of the broader AI landscape, especially generative AI such as image and video generation, will be covered in a future blog post.
Symbolic AI – Rules, Symbols, and Expert Systems
Symbolic AI is a classical approach to artificial intelligence in which knowledge is represented using symbols and explicit rules. The core idea is that every aspect of intelligence can be described precisely enough for a computer to simulate it. A pleasant side effect is that humans also find symbols easy to perceive and interpret.
These symbols represent objects, concepts, or properties such as “cat,” “fast,” or “bank account.” Systems like digital dictionaries or rule-based programs use this symbolic representation to make decisions. Expert systems, which are typically fed with knowledge from real-world experts, are also based on this principle. They are reliable and controllable, but lack flexibility and are unable to adapt to new situations. However, their strengths make them well suited for sensitive fields such as medicine or aerospace, where predictability and transparency are essential.
Decision Trees – Symbolic Learning in Practice
A simple yet effective model in machine learning is the decision tree. It works by asking yes-or-no questions about the data and making decisions based on the answers. Decision trees are particularly well suited for small, manageable datasets and have the advantage of being easy for humans to understand.
For example, deciding whether to take an umbrella can be modeled with a simple decision tree that asks questions like “Is the sky cloudy?” or “Is rain forecasted?” However, decision trees reach their limits when dealing with more complex tasks. They can become difficult to manage and are prone to errors in edge cases. For tasks such as image or speech recognition, more powerful subsymbolic methods are required.
Subsymbolic AI – The Power of Neural Networks
Neural networks are at the heart of modern AI. They consist of many small computational units, called neurons, organized into layers. By training on data, they learn to recognize complex patterns. Deep neural networks, also known as deep learning models, contain multiple hidden layers and enable a high level of abstraction.
For those who paid attention in math class and want a more technical explanation: each layer in a neural network operates using matrices, vectors, and mathematical transformations. These networks transform data into higher-dimensional spaces where patterns can be separated more effectively. However, the knowledge that emerges from this process is not directly understandable to humans. This is known as subsymbolic knowledge.
Unlike decision trees, neural networks are not transparent, but they are extremely powerful. On the one hand, a properly trained neural network usually delivers the desired results. On the other hand, as mentioned earlier, neither everyday users nor even specialized researchers can fully explain how a neural network arrives at a particular solution. Paraphrasing Socrates, one might say in the context of AI: “I know that I know nothing.” This mindset may actually be the best foundation for using AI responsibly in everyday life.
Give Me Your Data and I’ll Show You the World
Large and sufficiently diverse training datasets are the foundation of every AI model and largely determine its performance. High-quality data is accurate, consistent, and free from errors or biases. The data must also be directly related to the task the model is meant to learn.
Balanced datasets are crucial to prevent models from favoring or disadvantaging certain categories. For example, a system designed to filter job applications might systematically disadvantage women—not because the AI is biased by nature, but because it was trained on historical data from decades when women were underrepresented in leadership positions.
Since training data often comes from the internet or other broad sources, it must be carefully prepared before use. Given the sheer volume of data involved, this preparation process is extremely time-consuming and is a discipline in its own right.
Good Training Leads to Success
Neural networks can be trained in different ways, depending on the available data and the intended application. The most common approach is supervised learning, where the network is trained on labeled data—inputs paired with known correct outputs. The network adjusts its internal parameters to minimize the difference between its predictions and the actual results.
In unsupervised learning, there are no predefined outputs. The network independently searches for patterns or structures in the data, such as clusters or reduced dimensions. A third approach is reinforcement learning, where an agent learns through trial and error by receiving rewards for good actions and penalties for bad ones.
There are also hybrid approaches, such as semi-supervised learning, which combines labeled and unlabeled data, and self-supervised learning, where the network derives its own learning objectives from the data. Each method has its strengths and is chosen based on the specific problem and data situation.
Inference – What Does That Mean?
Inference refers to the process of applying a trained neural network to new, previously unseen data in order to make predictions or decisions. The input data passes through the network from the input layer, through the hidden layers, to the output layer. The learned weights—those training parameters mentioned earlier—and activation functions process the data along the way.
Unlike during training, the weights are no longer adjusted during inference. The network simply uses what it has already learned to compute an output. Inference is typically very fast and efficient, as it only requires a forward pass through the network. It represents the practical application of a neural network, such as when a language model generates a response, an image classifier identifies an object, or a recommendation system suggests a product.
LLMs – And Who Doesn’t Like Talking to a Computer?
When people talk about AI today, many immediately think of ChatGPT, Google Gemini, and other large language models. One reason for this is how surprisingly natural it feels to interact with them. In many cases, modern LLMs can pass the Turing Test, meaning the conversation feels human-like. Only careful questioning reveals that the interaction is with a machine.
But how do these large language models actually work? They begin by using neural networks to analyze vast amounts of text, learning how language is structured. A key component is the attention mechanism, particularly in the form of the transformer architecture. This mechanism allows the model to focus on relevant words in a context, regardless of how far apart they are in a sentence.
This enables the model to understand relationships and meaning across longer passages of text, which is essential for generating high-quality language. During inference, the model uses these learned relationships to respond to new inputs with contextually appropriate outputs, such as answers, text suggestions, or translations.
It is also important to understand that in large language models like GPT, each newly generated word—or token—is fed back into the model to predict the next one. The model calculates the probability of the next token based on all previous tokens in the context.
Impressive Results, but Handle with Care
The capabilities of LLMs are undeniably impressive. They can generate text, answer questions, hold conversations, and are increasingly transforming how we search for information—shifting from traditional search engines to AI-generated answers.
However, caution is always necessary. Language models can hallucinate. Hallucinations occur when a model generates false or fabricated information that sounds plausible and convincing. This happens because the model has no true understanding of the world. It merely predicts the most likely next word based on statistical patterns learned from its training data.
The model does not know whether a statement is true; it only knows what sounds probable. If a question refers to something not present in the training data, the model may produce an answer that seems reasonable but is incorrect. Therefore, while language models can be incredibly helpful in everyday life, it is essential to remain critical of their outputs and verify important information using reliable sources when in doubt.