AI in a Nutshell: A Practical Guide to Key Terminology

Hi there,

Understanding AI terminology like GPTs, LLMs, ML, etc. can be daunting, but today we're here to demystify some common terms.

In today's issue, you'll learn more about some popular AI terms and their connections, so you can communicate more effectively in data and AI projects.

Hopefully, this will come in handy the next time you're collaborating with colleagues from other departments or engaging stakeholders.

Let's start!

Update, July 2026: I first published this guide in June 2023, a few months after ChatGPT originally launched. A lot has happened since then, but the core taxonomy hasn’t changed. This article will still give you the right mental model, but I’ve refreshed the examples and references throughout – and added a small 2026 update to the end.

AI terminology can be quite confusion. The main reasons are:

The way some terms are being used will be different whether you're talking to a researcher, data scientist, or business executive.

I guess we just need to accept that.

If there's one thing I've learned, it's that being strict about the use of correct terminology in a business can quickly turn into fighting windmills. Remember, it's not about the terminology itself, but the concepts and ideas they represent.

So, while this article is a good launching pad, feel free to tailor it to your organization's needs to better navigate AI-related discussions and make more informed decisions.

Here's a high-level overview of the key terms in a hierarchical manner, beginning with AI and ending with ChatGPT:

Let's go through these step by step!

Artificial Intelligence (AI) has become a high-level umbrella term that generally refers to systems (machines) that behave as if they were intelligent. "Intelligence" in this sense generally refers to the ability to perform more or less cognitively complex tasks such as learning, problem solving, and decision making.

Brief history: The term AI dates back to the 1950s, with strong roots in military research. Since then, it has evolved to encompass a wide range of applications across industries.

Specifically, two main research areas have emerged: Strong AI, or Artificial General Intelligence (AGI), and Narrow AI ("weak AI").

Strong AI aims to replicate human intelligence in all aspects. The research goal here is to create systems that can solve previously unseen problems, much like a human. The ultimate goal would be superintelligence or singularity, an AI system that matches or exceeds human intelligence in all areas. However, the time when both researchers and practitioners believe this is achievable is still far in the future. It is uncertain whether it will ever be achieved.

Narrow AI, on the other hand, refers to task-specific systems, that are designed for particular jobs like language translation or image recognition. Unlike Strong AI, Narrow AI is not meant to replicate human intelligence in all aspects. Instead, Narrow AI is designed to perform narrowly defined tasks efficiently.

Narrow AI systems have a lot of practical use cases. When you encounter an AI application in business, it's narrow AI at work.

There are two ways that systems can be "intelligent". Either these systems are based on pre-programmed rules for decision-making without learning from data. ("Rule-Based Systems") Or, these systems can learn from data without having to explicitly hand-code the rules.

This is where Machine Learning (ML) comes in.

Machine learning (ML) is a technology that lets systems iteratively learn from large amounts of data, identify patterns, and make decisions with minimal human intervention.

To be clear, ML per se doesn't imply self-awareness or emotion, as is sometimes suggested in popular media.

Rather, Machine Learning is a statistical approach to automated pattern recognition in data.

Deep Learning, is a subset of machine learning, that uses a certain type of machine learning models - so-called artificial neural networks (ANNs) that typically contain many layers (hence "deep").

While the concept of ANNs was inspired by the human brain, the actual operation of these networks is vastly different from the biological processes in the brain.

Still, these models are sophisticated pattern recognition systems and should not be mistaken for conscious entities.

Essentially, deep learning is just a very complex variant of machine learning.

Generative AI is a branch of deep learning that aims to create or modify original content. It doesn't inherently have a creative mind; it uses statistical patterns in data to generate similar content without "understanding" the content the way humans do.

The first popular example of this technology was Deepfakes, where visual and audio content is altered to such an extent that it's hard to tell if the content is real or not.

Generative AI spans multiple domains and approaches, from text generation to music generation to image generation to video generation.

The Generative AI hype was started by the invention of another specific type of ANNs, namely GANs - Generative Adversarial Networks in 2014 by Ian Goodfellow and his colleagues. Since then, a lot of research has been done in the Generative AI space, which ultimately led to the development of another (and in many cases better) type of ANN architecture - Transformers, introduced in 2017 by Google.

Today, transformer-based AI models dominate text generation, while most image and video generation today uses a different approach called diffusion models. You don't need to know how either works internally – just be aware that "Generative AI" is no longer a single technology.

The Transformer architecture, introduced in 2017, was a major breakthrough for Large Language Models (LLMs). BERT, a widely recognized LLM developed by Google in 2018, leveraged this architecture and received widespread acclaim for natural language understanding tasks such as question answering and sentiment analysis.

Today, modern LLMs typically have billions (or even trillions) of parameters and can help us solve not only tasks such as answering questions, but also writing essays, summarizing long documents, translating languages, generating code, and much more!

LLMs don't have true language understanding; their responses are based on patterns learned during training, not on inherent knowledge or comprehension of the world.

Hence, they are sometimes called "fancy autocomplete" as they essentially try to complete text inputs by predicting what comes next based on a given context and patterns in their training data.

(That's still true at the core – but see the update at the end of this article. Reasoning models have stretched this analogy quite a bit.)

Over the past few years, there has been a tremendous effort in LLM development, both from commercial organizations and open source contributions.

Generative Pre-trained Transformers (GPT) are a special type of Large Language Models that use a transformer architecture to generate original content.

GPTs are typically trained on a very large corpus of data - such as all the text on the public internet - from which they learn the patterns and structures of language to generate text.

Despite its impressive capabilities, it's crucial to note that even the most advanced GPT does not "understand" the content it generates in the way humans do.

A GPT is essentially recognizing and predicting statistical patterns in the data it was trained on – a classical machine learning approach.

Practically all leading models today are built this way – OpenAI's GPT series, Google's Gemini, Anthropic's Claude, and Meta's Llama family.

When I first wrote this article, GPT-4 was the most capable model in the world. It has long since been retired. Today, the frontier is shared between a handful of labs – OpenAI (GPT), Google (Gemini), and Anthropic (Claude) – and the leaderboard reshuffles every few months.

More useful than tracking names is knowing what today's frontier models can do that GPT-4 couldn't: they work with images, audio, and documents natively, they handle hundreds of pages of context in a single request, and they can "reason" through hard problems before answering (more on that below).

Two limitations haven't gone away, though: AI hallucinations – confident-sounding outputs that aren't grounded in actual knowledge or data – and biases inherited from training data. Better models have reduced both. They have eliminated neither.

To learn how these models are actually built, I recommend Andrej Karpathy's Deep Dive into LLMs like ChatGPT – still one of the best general-audience explanation I know.

This is where we shift from a “technology” to an “application”.

ChatGPT is a web application by OpenAI that lets users interact with their latest GPT models.

While ChatGPT is the most popular, several similar services compete in the space:

While applications like ChatGPT often seem like a thin layer around existing LLMs, it's worth noting that there's often more going on than you might expect.

For example, while ChatGPT has to deal with the same limitations as the underlying GPT models, such as AI hallucination or biased output, the application layer provides more ways to mitigate these limitations, such as setting rules and filters to keep responses within certain parameters or expected outcomes - a process often referred to as "alignment" between the model and the desired behavior.

Don't be confused: ChatGPT is the application; GPT is the model family inside it. The same naming overlap exists with Gemini and Claude. So when someone says "we use ChatGPT," you still don't know which model they're using — and for business use cases, that difference matters a lot.

The taxonomy above still holds. But (at least) three terms have entered the everyday AI vocabulary since I first wrote this guide – you'll hear all three in any AI meeting today:

Reasoning models. Modern LLMs can "think before they answer": they generate internal reasoning steps before producing the final output. This made them dramatically better at complex tasks like math, coding, and analysis. And it's why "fancy autocomplete" is now only half the story. The trade-off: reasoning costs time and money, which is why most products let you choose how much thinking you want.

RAG (Retrieval-Augmented Generation). Instead of relying on what a model memorized during training, RAG systems first look up relevant information – from your documents, wiki, or database – and let the model answer based on that. It's still the standard architecture for making AI useful on your company's knowledge. Full deep dive here: Understanding RAG for AI Applications.

AI agents. This is the biggest shift. An agent doesn't just answer, it acts: it uses tools, browses, writes files, and works through multi-step tasks toward a goal. Under the hood, it’s still an LLM that generates text commands, but the blast radius has obviously increased. AI agents have moved the key question from "is the answer correct?" to "what is this system allowed to do?" I've written about how to think about them in AI Agents are Dogs and Agents Are Ready (Your Org Probably Isn't).

If you want to stay current without following the daily noise, subscribe to my free weekly newsletter:

I hope this article gave you a good overview of some key AI concepts and how they are connected.

As you continue to interact with AI, feel free to come back to these explanations and keep in mind the potential applications and limitations of these technologies.

Good communication involves tailoring the message to the audience and using appropriate language to convey the intended meaning.

Feel free to build an AI vocabulary that helps you along your journey - and sometimes that might involve not talking about "AI" at all.

And if you're past the terminology stage and want to know how to make AI actually pay off in your business, start here: Profitable AI is rare, fragile, and misunderstood.

See you next Saturday,

Tobias