There is so much AI (Artificial Intelligence) terminology being thrown around these days that it’s becoming difficult to know what people are actually talking about. Every week, there seems to be another acronym. LLMs, RAG, Agents, and Inference. If you don’t keep up, it all sounds far more complicated than I actually think it is.

The way I think about it is actually fairly simple.

  • AI is the overall umbrella.
  • Machine Learning detects patterns in data to make predictions.
  • Generative AI creates new content.
  • LLMs and SLMs are language models of different sizes.
  • Training teaches the model; inference is using it.
  • RAG lets the model look up information before answering.
  • Agents combine all of these capabilities to accomplish real tasks.

All of these are really just different capabilities that build on each other.

AI itself isn’t actually new. The term has been around since John McCarthy coined the phrase “Artificial Intelligence” and organized the 1956 Dartmouth Conference, which formally established AI as an academic discipline. What’s changed is that we now have enough computing power, enough data, and better algorithms to make it genuinely useful in everyday business.

The two types of AI you actually need to understand

When people talk about AI today, they’re typically referring to one of two things: Machine Learning (ML) and Generative AI.

Machine Learning is really the AI everyone was talking about before ChatGPT became mainstream. Think of ML as finding statistical relationships in data, often numerical data, to identify patterns and make predictions. Feed it enough data, and it is very good at identifying patterns, making predictions, spotting anomalies, or recommending the next best action. That’s why it’s been used for years for things like fraud detection, predictive maintenance, forecasting, recommendation engines, etc.

Generative AI is different. Think of GenAI as learning the statistical relationships within language rather than numbers. Instead of trying to predict an outcome, it’s trying to predict the most likely sequence of words based on everything it has learned. That’s also why people sometimes say that AI “hallucinates”. The model isn’t trying to tell you the truth. It’s trying to produce the statistically most likely response based on the information available to it. Most of the time, those two things are the same, but not always.

At the center of all of this is the model itself

Whether people talk about a Large Language Model (LLM) or a Small Language Model (SLM), they’re really just talking about different sizes of language models, but the underlying idea is the same. Larger models generally have broader capabilities and can perform a wider range of general-purpose tasks, but they also require more computing power. Smaller models can often be faster, cheaper, and sometimes they’re actually the better option if they’re solving a very specific problem.

Teaching AI versus using AI

Another couple of terms that often get mixed together are training and inference.

Training is simply the process of teaching the model. That’s where it learns the statistical relationships from very large amounts of data.

Inference is when you actually use the model. Every time you type something into ChatGPT, ask Claude a question, or use an AI assistant at work, you’re performing inference.

Where your own data changes the game

One limitation of language models, however, is that they only know what they were trained on. That’s where Retrieval-Augmented Generation, or RAG, comes in. Rather than relying purely on what the model already knows, a RAG allows it to retrieve additional information within your own guardrails before providing an answer. In practice, that usually means connecting the model to your own documents, knowledge base, policies, manuals, or other business information.

There’s quite a bit of technology behind this; however, the important part is that with a RAG, the AI is no longer relying purely on information captured during training. It’s answering using your information as well.

From answering questions to getting work done

Finally, there are AI Agents.

This is where many of these capabilities start coming together. Instead of simply asking AI a question and getting an answer back, you now give it a goal. The agent can retrieve information, reason about what it’s found, leverage other software, make decisions within pre-defined guardrails, and complete a series of tasks. The way to think about it is that a chatbot answers questions, while an AI Agent can complete a series of tasks towards a defined goal.

Once you start thinking about AI this way, the terminology becomes much easier to understand. Instead of just being dozens of disconnected buzzwords, they are different building blocks that all fit together. And I think that’s the important part. You don’t necessarily need to understand every technical detail behind AI. But understanding how these building blocks fit together makes it much easier to see how the technology can genuinely create business value.