Part 1 (20 minutes): Origin of the idea and illustrative history of development
The idea that machines could exhibit intelligent behavior has roots in several strands: philosophical questions about the nature of thinking, the development of formal logic, and the emergence of programmable computing machines in the 20th century. Alan Turing formulated in 1950 the fundamental question of whether machines could think and proposed pragmatic tests for assessment; his essay contains both technical and conceptual impulses for further research [1]. Immediately relevant was the notion of formally describing the building blocks of intelligence; in the now-famous proposal for the Dartmouth workshop in 1956 the hypothesis was formulated that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it" — this formulation shaped the early research program of AI research and explains the initial optimism of many researchers [2].
An early practical outcome of this phase was the development of programs that addressed logical problems and drew simple inferences; pioneers such as Newell and Simon demonstrated that certain forms of problem solving can be automated with explicit rules. In parallel, initial approaches to learning machines emerged, such as the perceptron algorithm, which took biological models as inspiration and described a simple model "node" for pattern recognition [3][6].
For a clear analogy: early AI research can be compared to a workshop testing two basic approaches. The first approach encodes knowledge explicitly as rules and symbols — comparable to a collection of instruction manuals; this path led to expert systems that could perform practical tasks with coded rules. The second approach tries to learn structures automatically from many examples — comparable to training a craftsperson through practice; this field corresponds to what is today called connectionism or neural networks [8][12].
In the 1960s and 1970s there were both visible successes and sobering limitations: well-known demonstration systems like ELIZA showed that linguistic interactions can be simulated without genuine understanding, while robotics projects like SHAKEY required complex technical coordination and made clear how much peripheral functionality (sensing, control, world models) is necessary for behavior to become robust [5][4]. The result was a twofold gain in insight: concepts worked in narrow, controlled areas, but transfer to open, real-world environments remained difficult.
The initial euphoria led to high expectations among some research groups and funding institutions; historical accounts show that in the early decades there was a widespread expectation that fundamental problems of intelligence could be solved within a few decades. These expectations proved overoptimistic because they underestimated the complexity of real environments and practical constraints [2][3].
Part 2 (20 minutes): Deepening and introduction of central technical terms
To place the historical stages more precisely, some central terms are helpful. "Symbolic AI" denotes the approach of representing knowledge explicitly as symbols and rules and building reasoning systems. From the 1960s to the 1980s this was the dominant paradigm; programs worked with explicit heuristics and knowledge bases, as later expert systems showed [3][8].
"Connectionism" or "neural networks" denote models made of many simple nodes whose connections are weighted and adjusted through experience. The perceptron is considered an early, very simple neural model; its limited capabilities were outlined in an influential analysis in 1969, which temporarily dampened research on simple networks [6][7]. The fundamental learning rule that returned to the center of attention due to later research is the gradient method "backpropagation", which was revived and formalized in the 1980s (Rumelhart, Hinton, Williams) and made training multilayer networks practical [9].
"Expert systems" are applications of symbolic AI that store domain-specific knowledge in the form of rules and use this knowledge for decision support. In the mid-1970s and 1980s expert systems became commercially interesting because they delivered high performance in narrow application areas; at the same time their limitations (maintaining large rule sets, lack of flexibility) became apparent [8].
"Deep Learning" is a modern umbrella term for methods with many processing stages (layers) in neural networks. The current success of deep learning methods is based on a combination of factors: algorithmic advances (e.g., specific architectural and optimization techniques), greater computational power, and significantly available datasets. Work in the 2000s and 2010s showed how deeper networks can be trained in practice (e.g., Deep Belief Nets and later convolutional networks), which became evident in 2012 with strong results in image classification competitions [10][11][12].
Among newer architectural ideas is the Transformer model, which shifted the focus from recurrent structures to attention mechanisms. Transformer architectures have proven particularly effective in natural language processing and are the technical basis for many current language models [13]. Scientific papers provide methodical definitions and show how the respective architectures exploit training and computational scale.
Finally, it is important to distinguish the terms "training" and "generalization": "training" denotes the process of adjusting model parameters to data; "generalization" denotes the ability to respond sensibly to new, not exactly seen situations. Historically, many advances depended on methods that provided better generalization under realistic conditions; many setbacks occurred precisely where generalization failed [12].
Part 3 (10 minutes): Applications, limits, and short thought exercises
Concrete applications illustrate the historical stages: chess programs and other game programs served early on as testbeds for planning and search; expert systems such as medical decision support systems (classic examples are mentioned in the historical literature) demonstrated practical usefulness in limited domains; image classification with deep nets (notably since work around 2012) changed applications in image recognition and many industrial areas [11][8].
At the same time, historical and current reports reveal systematic limits. These include the need for large, representative datasets, difficulties in explaining decisions (interpretability), problems with robustness and generalization in changed environments, and the risk that systems reproduce biases present in the data. Technical reviews make clear that these limits are technical in nature and cannot be solved by more compute power alone; algorithmic and methodological research therefore remains central [12][9].
Three short thought exercises to consolidate learning:
1) Consider why an expert system that works very well in one clinic does not automatically perform the same in another clinic. Which practical differences (data, workflows, notation of knowledge) play a role? Hints can be found in accounts on expert systems and domain specificity [8].
2) Take the perceptron as a simple model: what kinds of problems can a single perceptron solve, and where does it reach its limits? The historical critique of perceptrons highlights exactly such limitations and explains why multilayer models became necessary [6][7].
3) Briefly compare two modern approaches to language processing: recurrent networks versus Transformer models. Which architectural differences allow the Transformer to handle large amounts of text more efficiently? The introduction of attention mechanisms is described here as central [13].
For these exercises: answers are not gained by observation alone but by critical examination of assumptions and the data landscape. The literature points out that many historical surprises in AI resulted from a misjudgment of these assumptions [2][3][12].