Most of what passes for AI vocabulary in public discourse is marketing language grafted onto engineering concepts. The gap between how these terms are used in press releases and what they mean in practice is wide enough to cause real confusion, so this episode with Marius Högger sorts through the core terminology.
A classical algorithm is a fixed sequence of instructions: given input X, perform steps A, B, C, return output Y. In image recognition, a traditional algorithm would search for predefined features edges, shapes, contrast ratios to identify an object like a house. The rules are hand-coded by the developer.
An artificial neural network works differently. Instead of explicit rules, it uses layers of artificial neurons connected by weighted edges. Training the network on thousands of labeled images adjusts those weights until the system learns to extract the relevant features on its own. The developer never specifies what a house looks like; the network discovers that from the data. This distinction hand-coded rules versus learned representations is the dividing line between classical programming and machine learning.
Data science is the discipline of extracting knowledge from data: collecting it, cleaning it, analyzing it, and drawing conclusions. Machine learning is one tool in that toolkit, not the whole field. A data scientist predicting real estate prices might start with a simple regression on historical trends. When the model needs to account for location, floor area, amenities, and macroeconomic indicators simultaneously, machine learning becomes the more effective approach because it handles high-dimensional feature spaces that overwhelm classical statistical methods.
The confusion between the two terms matters because it leads organizations to deploy ML where a straightforward statistical analysis would suffice, and to underinvest in data quality the part of the pipeline that determines whether any model, classical or learned, will actually work.
The boundary between ML and AI is blurry, but a useful distinction exists. Machine learning systems typically excel at a single, narrowly defined task: classifying images, predicting churn, detecting fraud. AI systems like ChatGPT operate across many tasks analyzing text, writing code, drawing inferences, composing poetry using a single underlying model. That versatility is what earns the label "intelligence," though the term remains contested.
Artificial General Intelligence, the hypothetical system that matches or exceeds human cognition across all domains, remains out of reach. Current AI systems are powerful within their modality text, image, audio but they cannot transfer learning across modalities the way humans do. A language model that writes excellent prose cannot learn to navigate a physical environment without an entirely separate training pipeline.
Terminology shapes perception, and the labels attached to AI products often distort more than they clarify. The word "chatbot" carries the baggage of early rule-based systems that matched keywords to canned responses. Calling ChatGPT a chatbot obscures the fact that it runs a 175-billion-parameter transformer. The suffix "-bot" implies something mechanical and scripted, which is the opposite of what these systems do.
Labels like "assistant," "agent," and "copilot" are more accurate. "Copilot" in particular positions the system as a collaborator rather than an autonomous actor, which aligns well with how most productive deployments actually work: a human setting direction, the model handling execution.
Under the hood, ChatGPT is a mathematical function. It takes a sequence of tokens as input and outputs a probability distribution over the next token. The token with the highest probability is not always selected; a parameter called temperature controls how much randomness is injected into the sampling process. Higher temperature produces more varied, less predictable outputs; lower temperature makes the model more deterministic.
This mechanism explains both the system's strengths and its failure modes. It generates fluent, contextually appropriate text because it has learned statistical regularities across a massive training corpus. It hallucinates because the same statistical machinery can assign high probability to plausible-sounding but factually wrong continuations. There is no internal fact-checking step, no world model, no understanding just a very large, very well-trained next-token predictor.
A classical algorithm follows a predefined sequence of steps, like a recipe. AI uses artificial neurons that learn from example data during training. Developers do not need to program explicit rules; the system discovers relevant patterns on its own.
Data science is the broader process of generating knowledge from data, encompassing collection, preparation, and analysis. Machine learning is a specific tool within data science, typically deployed when the volume or complexity of the data exceeds what classical statistical methods can handle.
The term 'chatbot' is historically associated with simple, rule-based systems with scripted responses. Applying it to systems like ChatGPT understates their complexity. Terms like 'assistant' or 'agent' more accurately describe the adaptive, multi-task nature of modern language models.
Current AI systems are confined to specific modalities such as text, image, or audio and operate only within those boundaries. AGI would describe an intelligence that matches or exceeds human capability across all domains, with the ability to learn and reason across modalities.
ChatGPT is a mathematical function with billions of parameters. It generates the next token based on the preceding context. The apparent creativity in its responses comes from controlled randomness governed by a parameter called temperature.
No. AI systems have neither understanding nor consciousness nor emotions. They are sophisticated text prediction models that produce convincing outputs through extensive training, but they remain mathematical functions at their core.
.
Copyright 2026 - Joel P. Barmettler