Prompt-engineering for GPT-4 and other language models

Large Language Models (LLM) are models that predict the next logical word. These have become so good at it that you can also ask questions of the language model. And in most cases, it answers correctly. As a result, you can talk to a language model like you talk to a chatbot. The best-known example of this, of course, is Chat-GPT.

For Highlight Delft's chatbot, Kris uses GPT-4; the technology behind Chat-GPT. GPT-4 is a language model created by OpenAI. GPT-4 is easy to "program" by giving instructions to the chatbot. These instructions are called prompts. For example, for Highlight Delft, Kris tells GPT-4 that the chatbot is a virtual tour guide, must answer questions from visitors, and when it doesn't know the answer it must admit it honestly.

So you program the chatbot by giving commands. This information sends the language model to OpenAI's server and they run the language model. This model then predicts the outcome word for word and from this informative text follows. You integrate the chatbot with an API on your own website. At the bottom of this article, Kris shares his tips for good prompts.

Kris' idea is to use this technique on Highlight Delft's website. Highlight Delft stands for combining new technology with art. A language model is a very accessible way to let people see and experience what is now technically possible. Kris creates a chatbot for each artwork on Highlight Delft. These chatbots can be accessed via a QR code. And you can scan those QR codes at the various artworks at the festival.

Art and technology are sometimes very abstract and profound and raise many different questions from people. That is why it is interesting that visitors themselves ask questions to the chatbot, instead of reading a standard information sign. For example, ask the chatbot, "Why is there a blue light flashing at this artwork?". Kris and Highlight Delft feed the language model with information about the artworks and, in addition, it uses information on the Internet. And with that, the artificial intelligent tour guide answers all questions about the artworks on Highlight Delft.

Question: who is Sophia Bulgakova? And what kind of artwork does she have?

Answer:

Sophia Bulgakova is an interdisciplinary artist born in 1997 in Odesa, Ukraine, who currently lives and works in The Hague, Netherlands. She has a background in ArtScience and her work is at the intersection between art, technology and society. Bulgakova focuses on exploring the relationship between light, perception and imagination. She uses different sensory inputs in her installations and performances to influence spectators' perceptions and explore new possibilities.

One of her notable artworks is 2019's "YOU ARE SOURCE PROJECTION AND REFLECTION," a projection-based installation that explores notions of physical space and nonlinear legibility in motion. This installation creates an infinity of possible compositions centered on the viewer's imagination. The audience is invited to explore space individually by immersing themselves in an environment of merging clouds of words, using transparent screens to navigate. The project was developed in collaboration with the Summer Sessions Network for Talent Development, Metamedia Association and V2_ Lab for Unstable Media.

Would you like to know more about her artistic approach, techniques or specific works?

Another interesting use of a language model is to keep track of what questions visitors ask. All those questions the language model stores in a logbook. This provides insight into the interests of visitors and gives the festival a lot of feedback.

The technology behind these language models has only been around for a year. There are already several applications, but as a personal tour guide, Sping has not seen this before. So it is a real innovation and we see a lot of potential for other festivals, museums and cultural events. Kris is now building this for Highlight Delft, contributing to the festival's mission of letting people experience technology. And at the same time it is a test for Sping for a possible wider application of this technology.

Kris also involves his colleagues in this new technology. For example, we are testing internally a new way of building applications using language models. This project for Highlight Delft is thus also an experiment for us; both in terms of the new technique, and the new application of AI.

With these tips, you will probably go a long way in setting up your own chatbot based on GPT-4. Do you still have questions or would you rather leave it to Kris and his colleagues? Feel free to contact us.