When people hear the word model in AI, they often imagine something very complex and mysterious. And of course, the training process is complex. But what a model actually is can be explained in a much simpler way. This episode is not about how models are trained in detail, not about data sets or pipelines or all the other technical steps. It's about the simplest and most useful mental model: what a model really is and how it works. Let's start with large language models because they are the easiest example.
A language model works with tokens. Tokens are small pieces of text, parts of words, whole words, sometimes punctuation. When you write a prompt, your text is first broken down into these tokens. Those tokens are then passed through a neural network. And this neural network does one main thing based on everything it learned before. It predicts what the next token should be. So the model is not a database of text. It doesn't store sentences or facts. It learns how tokens usually follow each other. No rules written by humans, no explicit knowledge, just learned patterns.
Now, here is the part that many people find surprising. All of this learning, all of the text, all of the data, all of the training doesn't stay alive inside the model. It collapses and it crystallizes into one thing, a file. After training is finished, a model is just a large file full of numbers. These numbers are called weights. They represent how strong different connections inside the neural network are. Training is simply the process of finding better and better weights. The better the training is, the better these numbers reflect the structure of the data.
This becomes very obvious if you have ever tried open-source models. If you have ever used tools like Ollama or similar software, you had to download the model, usually a very large file from websites like Hugging Face. Once that file is on your machine, you can disconnect your computer from the internet and the model will still be working. Nothing is missing, because all of the learned behavior is already crystallized inside these numbers within the model. Of course, in reality, things are a bit more complex. Models can be split into multiple files.
They can be encrypted. They can be protected in different ways. And large labs invest a lot of money in the security of these files. But no protection is 100% guaranteed. In the end, once the file exists, copying it is always much easier than training it. And that explains why model security is taken very seriously by labs like OpenAI, Google, and others. So when we ask what a model really is, a good answer is the following one. A model is a learned structure that predicts what comes next.
And all that learning eventually becomes just a file full of numbers. That's a note on AI for today.
