Skip to content
Tekoälli

Layer 3

Models

Training is paid for once. Running it is paid for every single time.

On this page
  1. What happens in this layer
  2. A common misconception
  3. Key terms
  4. Articles

Training a model means adjusting the numbers inside it. Picture billions of tuneable dials that training searches for settings for. Inference means using the finished model, and of the two it is inference that is billed again on every question anyone asks.

What happens in this layer

  1. 01

    Pre-training produces a base model

    The model picks up language, the relationships between things, regularities about the world and a certain amount of problem-solving.

  2. 02

    Post-training sets the behaviour

    Examples of good answers, comparisons between answers, safety training and practice at using tools. Pre-training grants the capability; post-training points it somewhere.

  3. 03

    An answer is assembled one piece at a time

    Text is split into tokens. The model computes the probabilities for the next piece and picks one, then repeats until the answer is finished. No completed answer is looked up anywhere.

  4. 04

    A base model is not a database

    What it learned is spread across its numbers, so it cannot reliably say where any single fact came from.

A common misconception

Key terms

Whole glossary →
Inference
Using a finished model. Every question and answer consumes computation, and therefore money.
Transformer
A network design in which every token gets to weigh which other tokens matter to it.
Backpropagation
The method for working out how much each number in the model contributed to the error.
Learning rate
How large an adjustment is made to the weights on the strength of one error.
Base model
The general-purpose model that comes out of pre-training. Also called a foundation model.
Frontier model
The newest and most capable generation of models at any given moment.
Parameter (weight)
A number inside the model that was learned during training.
Token
Text chopped into small pieces. The model reads and writes tokens, and they are what you are billed for.
Context window
How much material the model can take into account in a single run.
Quantisation
Storing the model's numbers at lower precision.
Distillation
Transferring the abilities of a large model into a smaller one.
Open weights
The model's weights can be downloaded and run on your own hardware.

Articles on this topic

Sovereign AI and base models

Two phrases that turn up in the same sentence but answer different questions. One says what kind of model it is, the other says on whose terms it is used.

7 min