✍️ Let's Journal
The first experiment / application I wanted to test out with OpenAI was journalling. In previous years I worked on Here which originally started as a mindful social network and later pivoted into a journalling app. I wanted to use GPT-3 to generate journal prompts.
Click here to check it out:
In a support article OpenAI mentioned first try zero shot, then few shot then if neither work fine-tuning – so I started with zero shot. I used the following prompt:
Generate a journal prompt to exploring the topic of gratitude:
One parameter input for the model is temperature, for simplicity I dubbed it 'Creativity' although the two aren't quite analogous.
Temperature is the randomness of the selection of tokens. The higher the temperature the more random the selection of tokens. The lower the temperature the more likely the model will select the most likely token and be more deterministic.
So for a task like keyword extraction you'd want a temperature of 0. For creativity, you seem to hit a sweet spot around 0.7. I've added a slider to control temperature so be sure to play with it for yourself and see if you can spot the language changes between 0.5 and 1.0.
On an additional read OpenAI discourage the labelling of the temperature parameter as creativity. In the future I will pull out the parameter controls for temperature, top-p and top-k and allow the user to adjust them in an 'advanced' mode.
As a first experiment I am pretty happy – the model does what it says it should, with plenty of room for experimentation. Before coming to the GPT category of models I was suprised that input wise all you need is a prompt and the importance of prompt engineering.
If this all seems a bit basic, we are only just getting started and now I've got the basics of the site, OpenAI framework and inputs setup, I'm excited to start building out more experiments.
If you have any feedback or suggestions please let me know.