I am starting a new series on this blog. Every Friday, we will look at a technical term or concept and explain it using common language, debunking myths and mysteries along the way.
This is a very simplified explanation—not scientifically perfect, but close enough to help you have meaningful conversations with engineers and parse the fast-paced industry news.
What is an LLM? First, we need to understand what an LLM (Large Language Model) actually is. Whether it’s GPT-4, Claude, or DeepSeek, at their core, they are just very large files full of numbers.
Imagine a massive mind map. The circles are all the possible words in every language. Each word is connected to thousands of others. The “magic” lies in the lines connecting them. Each line has a number attached to it (engineers call these weights) that tells the computer how likely these words are to appear together.
(In reality the weights don’t actually connect words, but let’s not worry too much about it)
A “model” is essentially a giant digital look-up table of these weights. Materially, there is nothing else. When you hear “this model has 70 billion parameters,” just imagine a spreadsheet with 70 billion rows of word-association rules.
Can I run this on my laptop? There are massive engineering hurdles to make these files work quickly, mainly because they are often larger than a computer’s memory. However, you can download smaller versions (say, 5GB to 10GB) onto a standard business laptop, and the concept is exactly the same.
How to try it: There is a website called Hugging Face (the “GitHub of AI”) which offers many many models for download. While the site itself is for pros, you can download a simple app like LM Studio or Ollama. These apps let you browse Hugging Face, click “download” on a model, and chat with it entirely offline.
The takeaway: When you ‘use’ or ‘download’ a specific model, you are merely downloading a mathematical description of a mind map. You aren’t giving an “evil superintelligence” access to your hard drive or selling your soul. It’s just a file. It only “thinks” when you give it power and run it through a program.
Let me know in the comments if you heard a different explanation that made sense to you, or if you have other terms that you’d like to see explained in the future. Unless someone has a better idea we will talk about ‘multi modal models’ next week 😉
Leave a Reply