Basics of AI models

​

The life cycle of an AI model

A Large Language Model (LLM) undergoes two main phases:

1. The training phase
   • The model is trained on large data sets
2. The usage phase.
   • The model can be used to generate an answer
   • The model can not learn anymore

​

Training an LLM

To build an AI model, it needs to be trained on large data sets. This happens with large sets of data. Based on this data, the model learns which words are most likely semantically connected.

For example, it will be tested to fill out gaps in a text. With that, it learns probabilities of words in different situations (more about that below).

After the model is fully trained, it can not learn anymore. The date when the model is stopped to train is called “knowledge cutoff date”, because the model only learned facts up until this date and does not know anything happening afterwards.

​

Using an LLM

After that, the model is deployed and can be used to generate an answer. With the data from the training phase, it learned probabilities of different words occuring together. Obviously, the probability that Artificial Intelligence and Large Language Model appear together, than Artifical Intelligence and Banana Cake.

When a user sends a prompt to the model, the model will choose the next word or word-piece (token) based on these probabilities. For example, the word Hifrom the user lets the model probably answer with a greeting. It answers with Hello. Then, it generates the next most likely word based on Hiand Hello. This process is repeated until the model decides the request was sufficiently answered.

​

Influencing the output of a response

As mentioned above, the models can not learn after being deployed and after the training has been finished. So how do the models now know about the user and refer to previous messages in a chat?

For that, you can take a look at how Langdock works. Simplified speaking, the user sends a prompt in the interface to the model, the model generates an answer and sends the answer back to the user.

The model is always set back to zero, as it can not learn over time. The way it works is by not only sending the prompt from the user to the model, but also the previous chat, attached documents, assistant instructions etc. All this is put into the context window of the LLM (=the maximum amount of text a model can process in one message it receives).