The previous pages have helped you gain insight into the underlying systems of neural networks, language models, and GPTs.
The exercises and follow-up reflection questions below allow you to assess whether you are able to draw a number of conclusions, supplemented or not with knowledge you have acquired elsewhere.

Novice learner: ChatGPT and other language models (revision)

Did you find the previous pages unclear or overly complex? If so, watch the video below by education expert Barend Last before continuing with the applications that follow:

https://www.youtube.com/watch?v=iOaJZ3oP2T4

❗️Exercise: Do you understand the meaning of the three terms (generative, pre-trained, and transformer)? Match the most appropriate term to the examples below.

❗️ Exercise: Go through the questions below and try to formulate an answer yourself first. Then check the correct answer by clicking on the arrow.

Why are you given different answers for the same input?

The answer may differ (slightly to substantially) each time for exactly the same input because language models operate with a degree of randomness. At each step, the model calculates a probability distribution for each possible next word (or token). It does not always select the most probable option, but may instead make a random choice based on that distribution.

Expert learner: Temperature
To allow language models greater flexibility in selecting one of the options in the probability distribution, a parameter called temperature is used. A higher temperature makes the model more creative and less predictable, while a lower temperature results in more consistent but less varied outputs.

You cannot adjust the temperature parameter in the standard ChatGPT web interface (chat.openai.com). There is no user setting or custom instruction that has this effect. To adjust the temperature, you must use the OpenAI API, where you can explicitly set a temperature value.

You can simulate this effect using the following transformer visualisation: https://poloclub.github.io/transformer-explainer/
More information on how this transformer works can be found here: https://www.youtube.com/watch?v=ECR4oAwocjs&t=27s

When is it better to write a prompt in English?

English is the main language of most large language models. It is the most frequently used language on the internet and therefore most of the available training data is in English. As a result, models generally interpret English prompts more precisely and produce more complete and accurate responses, particularly for technical or factual questions. Other languages, on the other hand, might be used less online and hence less frequently included in token distributions. Words in smaller languages are therefore often split into a larger number of tokens, which may lead to minor differences in accuracy or interpretation.

Some helpful guidelines:
- Use English for general, technical, or knowledge-oriented prompts where precision is essential.
- Use other languages when the content, tone, phrasing, or target audience is specific to that language, for example in text editing, communication, or context-related topics.