Back to the Basics: What Are Large Language Models and Why Do They Matter?

This is one of those months in which so much is happening in the space and industry that I feel incapable of keeping up. Here’s a handful of things going on:

• At an international AI summit in Paris, US Vice President, JD Vance, emphasized the importance of limited regulation to foster innovation.

• OpenAI’s leadership Chief Technology Officer, Mira Murati, and co-founder Ilya Sutskever departing to pursue their own independent (billion dollar) AI ventures.

• Google has updated its AI ethics policy, dropping its previous pledge not to use AI for weapons or surveillance. The company now states it will use AI in line with international law and human rights

• Thomson Reuters secured a (potential) landmark win in a copyright infringement case against AI startup Ross Intelligence, with the US District Court of Delaware ruled in favor of Thomson Reuters, rejecting Ross's fair use defense.

This is just the tip of the proverbial iceberg, and it doesn’t even scratch at some of the innovations happening within individual companies.

That said, for today’s edition of Core Concepts, we’re going back to the basics: outlining a basic understanding of Large Language Models, covering some of the core principles and terms you may or may not already be familiar with.

Humor me, yeah?

Machine Learning: How LLMs learn

At the heart of every LLM is machine learning. This is a type of artificial intelligence that allows a computer to recognize patterns and make predictions without being explicitly programmed for every possible scenario.

Think of it like this:

• If you show a child thousands of pictures of different types of dogs, they eventually learn what a dog looks like.

• Even if they see a breed they’ve never seen before, they can still guess, based on past experience, that it’s a dog.

• Now and then they might get it wrong (for example, labeling a “fox” a dog) but they “learn” from their mistakes. Editor’s Note: Except for the fact that apparently a fox IS a dog. I just Googled it. You can’t win ‘em all!

LLMs work the same way. They are exposed to enormous amounts of text and learn patterns in language.

That’s how they know that “peanut butter and…” is likely to be followed by “jelly” and not, for example, “concrete.” (Don’t knock a PB&C until you try it!)

Neural networks: The brain of an LLM

OK, but how is machine learning powered? Enter neural networks, which are computer systems designed to mimic how the human brain processes information.

Let’s imagine a giant web of tiny connections, where each connection strengthens when it makes a correct prediction. Over time, the system gets better at recognizing patterns and generating text that sounds natural.

The specific type of neural network used in LLMs is called a transformer, which allows the model to process and understand words in relation to each other rather than one at a time. This is why LLMs can produce long, coherent responses instead of just stringing words together randomly.

Training Data: what LLMs learn from

LLMs don’t come up with new knowledge on their own—well, not that we know of. Generations come from patterns found in the data they’ve been trained on.

This includes books, articles, websites, and other forms of written text…the good, the bad, the ugly. The more diverse and high-quality the training data, the better the model’s understanding of language.

Tokens: How LLMs process words

To a human, sentences are made up of words. But to an LLM, everything is broken down into tokens.

A token can be a word, part of a word, or even a single letter in some cases.

For example, the sentence "AI is amazing!" might be broken down into three tokens: "AI," "is," and "amazing!"

Processing tokens instead of full words allows LLMs to better “understand” (I use this term loosely—there’s no human-like understanding) language structure and context, allowing them to handle things like spelling variations, slang, and different writing styles.

Context Windows: How LLMs remember what you say

Ever wonder why a chatbot sometimes forgets what you said earlier in a conversation?

That’s because of something called a context window. A kind of digital amnesia.

The context window is the amount of text the model can "remember" at any given time.

Smaller models might only remember a few hundred words, while larger models can keep track of entire conversations. Google Gemini can purportedly memorize the entire Lord of the Rings trilogy.

If a chatbot suddenly stops making sense, it’s often because the earlier parts of the conversation have fallen outside its context window, meaning it no longer has access to that information.

Fine-Tuning: Customizing LLMs for specific tasks

A general LLM like ChatGPT can handle a range of topics, but what if you need a model specifically for, say, medical advice, legal research, or customer service?

That’s where fine-tuning comes in.

Fine-tuning is when an existing LLM is trained on a smaller, specialized dataset to improve its performance in a specific area.

So…a healthcare chatbot might be fine-tuned using medical textbooks and research papers. This makes the model more accurate and reliable for specific use cases.

Fine-tuning allows businesses and researchers to create custom AI models without having to build them from scratch.

Reinforcement Learning: Teaching LLMs to improve over time

Training an LLM isn’t a one-time process. Not yet, at least.

Developers constantly refine these LLMs using reinforcement learning—essentially a technique in which the model is rewarded for good answers and corrected for bad ones.

Human reviewers guide the AI by ranking its responses. The model then adjusts itself based on this feedback, gradually improving its accuracy and usefulness.

This is why AI-generated answers today are much better than they were a few years ago—and why, theoretically, they’ll continue to improve.

Retrieval-Augmented Generation: Making AI more knowledgeable

One of the biggest challenges with LLMs? Their knowledge is limited to the training data they were given.

If they weren’t trained on recent events, they won’t know about them.

To solve this, developers use retrieval-augmented generation, AKA RAG. This technique allows AI to pull in real-time information from external sources—search engines or company databases, etc.—before generating a response.

Kind of like a student using Google to look up facts prior to answering a question. This method makes AI more reliable and up-to-date.

RAG is already being integrated into modern AI tools, which makes them far more useful for research, customer support, and other knowledge-based tasks.

Bringing it all together

At a very basic level, LLMs are just very advanced pattern-recognition systems. They process text, predict what should come next, and refine their answers based on feedback.

To summarize:

• Machine learning: the foundation of how LLMs learn

• Neural networks: the brain that processes language

• Training data: the text sources that shape the model’s knowledge

• Tokens: the way text is broken down for processing

• Context windows: how much information the model can "remember" at once

• Fine-tuning: customizing LLMs for specific tasks

• Reinforcement learning: teaching AI to improve over time

• Retrieval-augmented generation: connecting AI to real-time knowledge sources

I hope having some basic understanding these core ideas (concepts!) makes it easier to see how and why LLMs work the way they do.

Having a better understanding of what’s happening under the hood will empower you to use AI more effectively and confidently—whether for work, learning, or just curiosity (like learning that foxes or dogs).

Contact us at NorthLightAI.com to learn how we can help you build a stronger admissions and recruitment infrastructure using AI.