How are you feeling? What drives you? What connects us? Why? As unspectacular as such questions may seem at first, they can be transformative. Especially when they are asked in the right context and at the right moment. What else leads to the new, shared answers that we need everywhere?
Since 2017, we at denkwerk have been using tscheck.in to help people find questions and inspire them to ask themselves questions together more often. Since the pandemic at the latest, we have seen that this is actually happening: Hundreds of thousands of people from really every (!) country in the world have searched for questions via the site millions of times to check in together. To work together better, learn from each other, help each other, inspire each other ...
As good as the site was, it had reached its limits and was getting on in years. Ultimately, the breakthrough of generative artificial intelligence was the perfect opportunity for us to rethink the tool. And to make it new.
From the very beginning, tscheck relied on a few very simple design principles that were crucial to its long-term success: As soon as the user visited the page, a (constantly new) question was displayed, and each time in a surprising color - in many cases, users were already made happy this way.
We have remained true to these principles in a fascinating way: Despite the extended functionality, the questions remain the focus – with fresh colors and moving shapes.
What's more, while it was previously left to chance and the user to find the perfect question, it is now possible to search specifically for a suitable one. And what doesn't yet exist is simply generated: What is the quickest way to break the ice in space - when changing the crew of the ISS? How do you humorously moderate a tricky situation with the customer? How could Harry Potter fans start the day more inspired - or the curiosity of over 80-year-olds be aroused?
Inseparably linked to the topic of generative AI is the way we interact with it. It was therefore exciting to explore the topic of prompting - and find new solutions. ChatGPT - the interaction is in the name - has helped the technology achieve a breakthrough with its text-based, dialogical approach. We decided to make the relevant section of the prompt visible to the user. A slider can be used to add further criteria and thus refine it - the highest level is free text input and direct interaction with ChatGPT.
In purely mathematical terms, we have defined 4,480 contexts - in purely practical terms, it would be too inefficient to start a query every time. For the environment, the users and your wallet. Accordingly, we automatically generated suitable answers during development. The result: around 25,000 matching questions and a database structure that could not have been mapped manually.
We had the first functional "proof of concept" after a few hours. The actual implementation involved a few hurdles. The backend is implemented in Java, but there are no official OpenAI libraries available. There are community projects that you can fall back on. Otherwise, the client for the API had to be developed from Scratch. Thousands of requests were made to the OpenAI API when pre-generating the questions. We stayed below the specified limits - but the API was still often unstable and responded with errors after a certain time. So far, GPT-3.5 Turbo has delivered better results than GPT 4. This remains exciting - and we are keeping an eye on it.
Where is AI an integral part of your everyday (working) life? If you still lack access and an overview of the topic or you have an exciting use case or idea that you would like to bring to the streets - let's talk! We offer formats that are as inspiring as they are helpful in dealing with the topic. We look forward to it! By the way, feedback on tscheck.in is also welcome!