What is AI and where do LLMs fit in
The term “AI” gets thrown around constantly — applied to everything from a chatbot to a recommendation algorithm to a thermostat. It’s become a marketing label as much as a technical one. This post breaks down what AI actually means, where large language models (LLMs) sit in the bigger picture, and how to think about these terms without the hype.
What counts as AI
At its broadest, artificial intelligence is any system that performs tasks typically requiring human intelligence. That definition is intentionally wide, and it covers a huge range of things:
| Category | Examples |
|---|---|
| Rule-based systems | Spam filters, decision trees, expert systems |
| Machine learning | Fraud detection, recommendation engines, image classifiers |
| Deep learning | Object detection, speech recognition, language translation |
| Large language models | ChatGPT, Claude, Gemini, LLaMA |
| Robotics / autonomy | Self-driving vehicles, warehouse robots |
A spam filter that uses keyword matching is technically AI. So is a neural network with billions of parameters. The term itself doesn’t tell you much about sophistication — it’s more of an umbrella.
Note: If someone calls their product “AI-powered,” that could mean anything from a simple if-else rule to a fine-tuned transformer model. Always ask what’s under the hood.
The AI hierarchy
It helps to think of AI as layers, where each one builds on the last:
1. Traditional AI / Rule-based systems These are hand-coded rules. Think: “if the email contains these words, flag it as spam.” No learning involved — a human writes the logic. This is where AI started decades ago and it’s still used everywhere.
2. Machine learning (ML) Instead of writing rules, you feed the system data and let it find patterns. A model trained on thousands of labelled images can learn to tell a cat from a dog without anyone explicitly coding “cats have pointy ears.” The key shift: the system learns from data rather than following hard-coded instructions.
3. Deep learning A subset of ML that uses neural networks with many layers (hence “deep”). This is what powers image recognition, voice assistants, and translation services. Deep learning excels when there’s a lot of data and the patterns are complex enough that traditional ML struggles.
4. Large language models (LLMs) A specific type of deep learning model trained on massive amounts of text. LLMs are what most people mean when they say “AI” in 2025/2026. They predict the next token in a sequence, and at scale, that simple mechanism produces surprisingly coherent text, code, and reasoning.
What makes LLMs different
LLMs aren’t magic — they’re pattern machines operating at a scale that creates emergent behaviour. Here’s what sets them apart from earlier AI:
Scale matters. Models like GPT-4, Claude, and LLaMA have billions of parameters trained on enormous datasets. The sheer scale is what enables them to generalise across topics rather than being narrow specialists.
They’re generalists. A traditional ML model does one thing well — classify images, predict churn, detect fraud. An LLM can write code, summarise documents, answer questions, translate languages, and reason about problems, all with the same model.
Context windows. LLMs process input through a context window — a fixed amount of text they can “see” at once. Modern models handle anywhere from 8K to 200K+ tokens, which determines how much information they can work with in a single interaction.
They don’t “know” things. LLMs don’t have a database of facts. They have statistical patterns learned from training data. This is why they can be confidently wrong (hallucination) and why retrieval-augmented generation (RAG) exists — to ground responses in actual source material.
What is NOT AI
Not everything labelled AI deserves the title. Some things that get called AI but really aren’t:
- Hard-coded automation. A cron job that runs a script every night isn’t AI. It’s automation. There’s no learning, adaptation, or inference happening.
- Simple conditionals. If-else logic in your app doesn’t become AI just because a product manager says so.
- Search. Full-text search with ranking algorithms is clever engineering, but it’s not AI unless there’s a learning component involved.
- RPA (Robotic Process Automation). Most RPA tools replay recorded actions. They follow scripts, they don’t learn from them.
The line gets blurry when you start adding ML components to these systems. A search engine with a learned ranking model? That’s arguably AI. A cron job that triggers an LLM to process data? The LLM part is AI, the cron job isn’t.
Practical terms to know
If you’re working in or around this space, these are the terms worth understanding:
| Term | What it means |
|---|---|
| Model | The trained artefact — weights and parameters that encode learned patterns |
| Inference | Running a model to get a prediction or output (as opposed to training it) |
| Fine-tuning | Taking a pre-trained model and training it further on a specific dataset |
| RAG | Retrieval-Augmented Generation — feeding relevant documents into an LLM at query time to improve accuracy |
| Tokens | The chunks of text an LLM processes — roughly ¾ of a word on average |
| Hallucination | When a model generates plausible-sounding but factually wrong output |
| Embedding | A numerical representation of text (or images, etc.) used for semantic search and similarity |
| Agent | An LLM given tools and the ability to take actions — browse the web, execute code, call APIs |
| Context window | The maximum amount of text a model can process in a single interaction |
Where this is heading
The landscape is moving fast, but a few trends are clear:
Local inference is real. You can run capable models on consumer hardware now. A Mac Studio with 128GB of unified memory can run 70B+ parameter models locally via Ollama. The gap between cloud and local is shrinking.
Agents are the next layer. Models that can reason, plan, and use tools are where the industry is heading. Instead of asking an LLM a question and getting text back, you give it a goal and let it figure out the steps.
RAG over fine-tuning. For most use cases, it’s cheaper and more flexible to give a model access to your data at query time than to fine-tune a custom model. Fine-tuning still has its place, but RAG is the pragmatic default.
Open-source is competitive. Models like LLaMA, Qwen, Mistral, and DeepSeek are closing the gap with proprietary models. Running your own stack is increasingly viable.
The bottom line
AI is a broad term. LLMs are a specific — and currently dominant — subset. Most of what people call “AI” today is either machine learning or LLM-based, and understanding the difference helps you cut through the noise.
When someone says “we use AI,” the useful follow-up is: what kind? A rule-based system, a trained classifier, a fine-tuned model, an LLM with RAG, an agentic workflow? The answer tells you everything about what the product can and can’t do.
More posts on AI, DevOps, and developer tooling at noukeosombath.com.