By Andreas Refsgaard & AIxDesign Community

https://andreasrefsgaard.dk/

https://www.aixdesign.co/

Setup

Make a user with OpenAI and make sure you can log in: https://openai.com/api/

All assignments below should be done in the Playground (https://beta.openai.com/playground) with default settings.

Assignments

<aside> 👉 Assignment 1 - “Order a text”

Go to https://beta.openai.com/playground

Describe in a single sentence, what kind of text you would like and let GPT-3 write it for you. Make sure there are no line breaks or concluding spaces in your prompt, it often confuses GPT-3. Try with different genres of text (article, story, recipe, poem etc)

Examples: Write a poem about #FF00FF.

Write list of things to do when you are sad.

Create an outline for an essay about the role of computers in poetry.

**Write a fairytale about a jar of honey and a Nintendo Wii controller. ** Write tagline suggestions for a shop that sells smoked salmon and leather boots.

Save your best text here: https://docs.google.com/document/d/1zkkl7AZb8kC7rw-rsl1GT-ZB1p3QMZcduX5ne3wihM8/edit


</aside>

<aside> 👉 **Assignment 2 - Let’s make some better poems about things!

1)** Find 2-3 different images on your computer or online

2) Go to https://huggingface.co/spaces/Salesforce/BLIP, upload your images and get a descriptive image caption for each image.

3) Go to https://beta.openai.com/playground and use the image captions from step 2 to generate a poem based on the image.

Your prompt instruction could be as simple as Write a poem about [insert image caption].

Or perhaps try to “steer” the poem in a certain direction, by exporing different adjectives that describe the poem, or by adding in the style of Sylvia Plath or similar to your instructions.

Save your best poem + image here: https://docs.google.com/document/d/1zkkl7AZb8kC7rw-rsl1GT-ZB1p3QMZcduX5ne3wihM8/edit

</aside>

<aside> 👉 Assignment 3 - providing examples

Sometimes just asking for a text does not really give you the results you want. It often helps if you add examples of the kind of inputs (prompts) and outputs (generated texts) you want.

Example: Let’s say we want to generate a chatbot that has a certain tone of voice, and for instance want it to be really sarcastic. Instead of only writing an instruction, we can provide a few examples of questions and answers showing the API how to reply. It will not be perfect, but generally a few examples improves the outputs quite a lot.

Try it out here: https://beta.openai.com/playground/p/default-marv-sarcastic-chat

⏳If we have time: Try to change both the instructions and the examples for the chatbot, to give it a different tone of voice. Lets say you want to make it very sweet😍, secretive🥸 or perhaps very melancolic🥺, how would you design the instruction and the examples?

If you have a lot of data with input-output pairs, you can also fine-tune GPT-3, which improves quality, lowers latency and is cheaper once done. See this guide https://beta.openai.com/docs/guides/fine-tuning

</aside>

<aside> 👉 Independent hands-on exploration time

Now you have a bit of time to explore on your own. Spend some more time doing some of the assignments, check out the GPT-3 documentation for text completion, the offical examples or learn more about some of my works (scroll down on this page👇).

****Let’s meet back in the video call at 17:50 (Copenhagen time) for a quick show and tell from some of you, and perhaps a discussion of ideas or approaches.


</aside>