On this page
The model is not the chat windowA simple analogy: recipe and kitchenWhat does a language model do?Models can have different strengthsDownloading is not the same as learningTry it: identify the model in your appSourcesAn AI model is a system of learned values that turns input into a prediction or result. A chat app uses a model; it is not the model itself.
3Blue1Brown: Large Language Models explained briefly
Grant Sanderson uses clear animation to explain language models. Watch for the distinction between learning patterns and writing a reply. The video is about language models, which are one kind of AI model.
Watch on YouTube or turn on captions ↗Here is what matters, why it matters and how to check it on your own setup.
The model is not the chat window
When you open a chat app, you see buttons and a box for typing. Behind that screen, a model does the prediction work. The same app can sometimes run several different models.
An AI model contains values learned from examples. For text AI, these values help the system choose likely pieces of text. They are usually saved in one or more downloadable files.
Think of a model as a learned pattern system. It is not a person hiding inside the laptop, and it is not a folder of ready-made answers.
A simple analogy: recipe and kitchen
A recipe describes how ingredients can become a meal. The kitchen provides the tools to make it. In our rough analogy, the model is like the recipe, and the engine is like the kitchen.
The analogy has a limit: a model is not a readable list of cooking instructions. It is a large set of numbers. You do not need to read those numbers to use the model.
The app sends work to an engine, which uses a model. One product may combine the app and engine.
What does a language model do?
A language model works with text. When it creates a reply, it repeatedly predicts another piece of text using the input and what it has already written. These pieces are called tokens; lesson 5 explains them.
This can produce useful writing, but it can also produce a convincing wrong answer. The model does not become a fact-checker just because the sentences sound smooth.
A model’s learned patterns are also different from live internet search. An app can add search, but that is an extra part of the system.
Models can have different strengths
One model may write clearly in your language. Another may be stronger at code. A vision model can accept images when its software supports that input. A text-only model cannot inspect a photo simply because you ask it to.
Model names often include a family, a version, and a size. Do not let the longest name impress you. Ask whether the exact model supports your task, fits your computer, and allows your intended use.
A newer or larger model can still perform poorly on your specific work. Compare answers using the same examples.
Downloading is not the same as learning
Downloading a model copies its files to your disk. Loading it makes it ready to run. Chatting usually uses the learned values already in those files.
An ordinary conversation does not normally retrain the model. An app may save chat history or notes and use them later. Saving history is different from changing what the model learned.
We will cover training later. For now, remember that installing a model does not mean your computer has to teach it from the beginning.
Try it: identify the model in your app
Find the model selector or the loaded model’s name. Write the name down separately from the app name. If nothing is loaded, do not download a large file just to complete this exercise.
Use a made-up shop notice to compare two models later. Keep the request the same and check facts, clarity, and missing instructions.
- App: where you click and type.
- Engine: software that performs the model’s calculations.
- Model: learned values used for predictions.
Changing the colour of a chat window does not change the model. Selecting a different model can change the answers.
Official facts and real user evidence
Official documentation supports product and model facts. Community discussions show real setups, failures, and questions. A community result is supporting evidence, not a promise that another computer will perform the same way.
Watch a real test. Check what fits your computer.
See practical local AI tests from DeviceTerra, then use MamiLens to build a hardware-aware shortlist for your own setup.
This guide is educational. Model software, licenses, and hardware support can change. Check official sources before an important deployment.