A lit server rack casts a hazy image of a woman's face on a wall.

Where this language comes from

The story begins long before today’s chatbots. In 1943, Warren McCulloch and Walter Pitts described a mathematical model of neural networks. They reduced nerve cells to highly simplified switching elements that could be described using logic. This was not a digital replica of a brain. It was a computational idea inspired by biology. [1]

In the 1950s, Frank Rosenblatt developed this connection between biology and computing further with the perceptron. His 1958 paper described it as a model for the storage and organisation of information in the brain and explicitly discussed learning curves. Terms such as perception, recognition, and learning therefore became part of the technical description of a computational model. They did not claim that the machine had human experiences or an inner life of its own. [2]

Outside this specialist context, the same words quickly acquired an additional meaning. At the perceptron’s public presentation in July 1958, Rosenblatt spoke of a machine with an “idea of its own”. The New York Times wrote that it learnt by doing and became wiser in the process. A technical description had become a story about human-like abilities. The problem was not the terminology itself, but the unexamined transfer of its everyday meaning to the machine. [3]

Joseph Weizenbaum’s ELIZA program showed as early as 1966 how quickly human interpretation can eclipse a term’s technical meaning in conversation. ELIZA searched input for keywords and used them to form suitable follow-up questions. Even so, people attributed understanding and empathy to the program. Weizenbaum explained that the people using it supplied the missing knowledge and supposed insight themselves. [8]

As technical language, comparisons with biology are useful. They give a new technical idea an understandable name and indicate where it came from. The comparison does not imply complete equivalence. An artificial neuron is not a nerve cell, and machine learning is not the same process as human learning. The terms remain useful as long as we distinguish their technical meanings from their familiar everyday ones. [1] [2]

Modern language models are not simply enlarged perceptrons. They still belong to the technical family of artificial neural networks. They too process information in connected computational units whose weights are adjusted during training. The methods and scale have changed fundamentally, but terms such as neuron, network, weight, and learning have remained part of the field. [1] [2] [4]

Other terms were added later. The transformer architecture behind modern language models uses “attention” as the name of a precisely defined weighting calculation. Software around the model stores context and is therefore described as having “memory”. When such a system carries out several steps and software calls without individual human input, we call it an “agent”. Today’s technical language therefore combines older biological comparisons with newer terms for specific technical functions. [5] [6] [7]

What these words mean today

Neuron

An artificial neuron is a computational unit within a network. It processes numbers received from other units, weights them, and passes on a calculated result. The name points to its biological inspiration, but it does not turn the unit into a tiny brain cell. [1]

Learning

When specialists say that a model learns, they mean a technical process: its parameters are changed during training. These include the weights that determine the model’s behaviour. Through these adjustments, the model represents patterns found in its training data. The term therefore describes a measurable change in the model. It does not claim that the software gathers experiences or understands its mistakes as a person would. [4]

Memory

In computing, “memory” does not describe one single ability either. It may refer to the weights stored in the model, the current conversation context, or information held in an external database. Chat software can save earlier messages and send them to the model again with the next request. This feels like remembering, but technically it is a new model call supplied with the information once more. The term remains useful as long as it is clear which kind of storage is meant. [5]

Attention

“Attention” also refers to a specific calculation. A transformer uses it to calculate which parts of the input should receive greater weight for the current step. The word therefore describes how information is selected and connected computationally. It does not claim that the model consciously concentrates or considers something important out of personal interest. [6]

Reasoning / thinking

“Reasoning” can refer to generated intermediate steps, the decomposition of a task, or special methods for multi-step problems. Visible “thoughts” or “thinking logs” are themselves generated text. They can make a solution easier to follow, but they do not provide a complete account of the calculations performed inside the model. The term therefore describes a technical method or a form of output, not conscious thought. [10]

Hallucination

A “hallucination” is an output that sounds plausible but is false or unsupported. The word describes an error in generated content. It does not mean that the model perceives something that does not exist. The US National Institute of Standards and Technology therefore notes that the term can anthropomorphise the software and uses “confabulation” instead. [11]

Decision

When a system “decides”, it selects an output or action from its inputs, calculated values, and predefined rules. That selection can have real consequences. The technical term does not claim that the system weighs moral considerations or bears responsibility for the outcome. [7] [12]

Tool

A tool is a function, program, or service that an agent system can call. The model can request such a call in its output. The controlling software executes it and provides the result for the next model call. [7] [9]

Harness

A harness is the controlling software around a model. It assembles requests, stores state, starts model calls, processes the outputs, and executes requested tools. The model generates outputs; the harness turns them into a continuing technical process. [9]

Agent

“Agent” is also, first of all, a technical term describing a function. In current LLM systems, it refers to the combination of a model, controlling software, stored context, connected software, and granted access rights. Within these constraints, such a system can carry out several steps independently. Here, “independently” means that a person does not need to trigger every individual step. It does not mean that the agent has goals or a will of its own. [7]

Autonomy

“Autonomous” or “independent” describes how many steps a system can take without individual human confirmation. This scope for action comes from the harness, the available tools, and the permissions granted to the system. Greater autonomy therefore means greater technical scope, not goals of its own or independence from the infrastructure provided. [7] [12]

Why these meanings are so easily confused today

These words developed over decades as technical language. Today, however, we often encounter them outside their specialist context. Chatbots speak in the first person, have names, and answer in human-sounding voices. Product descriptions tell us about systems that think, learn, and act independently. As a result, we hear not only the technical meaning, but automatically bring the familiar human meaning of each word with us. [13] [14]

We will continue to use these technical terms in the articles that follow. But we will always ask which specific technical process a term describes. This allows us to discuss what a system actually calculates, stores, or executes without attributing feelings, intentions, or responsibility to it. Its real capabilities and risks do not become smaller. They are simply described more precisely.

In the next chapter

Next, we turn to the calculations themselves: a GPU can perform the many similar calculations required by an LLM faster than a CPU, even though it is not faster at every task.

Next article →

Deeper into the Rabbit Hole

If you would like to explore the subject further, these texts and videos continue the most important lines of enquiry:

Texts

Videos

Sources

  1. Warren S. McCulloch and Walter Pitts: A Logical Calculus of the Ideas Immanent in Nervous Activity, 1943.
  2. Frank Rosenblatt: The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain, 1958.
  3. Melanie Lefkowitz, Cornell Chronicle: Professor’s Perceptron Paved the Way for AI – 60 Years Too Soon, 2019.
  4. Google for Developers: Machine Learning Glossary — parameter.
  5. Anthropic: Context windows.
  6. Ashish Vaswani et al.: Attention Is All You Need, 2017.
  7. Anna Gutowska, IBM: What Are AI Agents?.
  8. Joseph Weizenbaum: ELIZA—A Computer Program for the Study of Natural Language Communication Between Man and Machine, 1966.
  9. OpenAI: Hugging Face Incident Technical Report, 2026.
  10. Fabien Roger et al.: Let’s Think Dot by Dot: Hidden Computation in Transformer Language Models, 2024.
  11. NIST: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, 2024.
  12. High-Level Expert Group on AI: Requirements of Trustworthy AI, 2019.
  13. Gavin Abercrombie et al.: Mirages. On Anthropomorphism in Dialogue Systems, 2023.
  14. Michelle Cohn et al.: Believing Anthropomorphism: Examining the Role of Anthropomorphic Cues on Trust in Large Language Models, 2024.