Generative AI: Bisrat's insights and workshop experiences

codeveloper · General · Generative AI: Bisrat's insights and workshop experiences

Generative AI

Bisrat's insights and workshop experiences

Our colleague Bisrat has led numerous workshops on generative AI in recent years and gained valuable experience in the process. In our latest blog post, he shares insights into running successful workshops, both at the University of Rostock and at the NØRD Digital Convention. Learn how Bisrat uses practical tips and personal experiences to show how to design and deliver workshops effectively. An ideal guide for anyone looking to improve their workshop design skills!

The current wave of generative AI requires attention not only from people involved in technical fields but also from the general audience. It is with this spirit that we at DEJ Technology have gone out and given some workshops and talks to reflect with people about the direction of the current generative AI. The pace at which the field is moving is so fast that it is important to keep up with the latest developments and to also be aware of the ethical implications of the technology.

From the early days of inception, DEJ has believed in sharing the knowledge and expertise to people in different talks. So far within the last year we have given talks related to language models in different venues such as NConf, University of Rostock, NØRD Conference and Rostock business. Each one brings a different challenge and a different experience.

Workshop project week Uni-Rostock

This year we gave generative AI and LLM-related workshops at the university for the second time. The University of Rostock offers a project week for students to explore various practical applications.

Workshop on the analysis of sensor data

Two years ago, we offered a workshop on analyzing sensor data. We had our "intelligent shopping cart" with us. It was equipped with lidar, a RealSense tracking camera, a Raspberry Pi and a smartphone. We drove the trolley around and recorded our movements using the various sensors. We gave the students the task of practically processing the sensor data in order to create information from it in real time.

The students stayed with us for hours and analyzed sensor data using tools such as Python, Numpy and Matplotlib. We were amazed by their creativity and enthusiasm.

Workshops about generative AI and LLMs

Last year's workshop around generative AI and LLMs revolved around two main topics: How language models like chatGPT are built and how to use openAI's API to build specific business applications. We discussed deeper into the Transformers architecture. The audience ranged from first year students to PhD students to people working in NLP.

Practical experience and prompt engineering

What was particularly impressive last year was the fact that the students did their own prompt engineering using the small web-based tool we built. At one point, they designed their own chatbot so that it was always angry and responded with angry replies. It was a fun and engaging experience. In the few hours that the workshop lasted, we generated or requested about 100,000 tokens from the students.

Changes in the language model scene

There have been several changes in the language model scene between last year and this year. One is that our understanding of how language models are built has improved significantly. The other change is the rise of several competing generative models compared to last year, when chatGPT was way ahead of everyone else. This time Gemini and Claude were good competitors.

Open source models and new tools

Open source models, or rather models with open weights, have also become very popular in recent months. After the advent of Lama and small extensions like llama.cpp, the ability to run language models directly on the laptop is a crucial step forward. We demonstrated to the students how to run their own language model on their laptop by connecting to tools like ollama.

The NØRD Digital Convention: Experiences and highlights

Largest digital trade fair in Mecklenburg-Vorpommern

The NØRD Digital Convention is probably the largest digital trade fair in the whole of Mecklenburg-Vorpommern. High investment and many different companies occupying the scene make it a very appealing environment. Overall, however, the event was more geared towards networking and making contacts.

Presentations and challenges

Nevertheless, we are grateful that we were able to give presentations at the NØRD. Unfortunately, we encountered a few difficulties during the event. Firstly, the rooms in which the workshops took place were not well shielded from outside noise. As a result, we had to speak very loudly in order to make contact with the audience. This made communication quite difficult.

Adaptation of the presentations

Because the majority of the audience had no technical background, we dynamically adapted our presentation. We mirrored business applications instead of going deeper into neural network based architectures. There were several questions from the audience that seemed to be well thought out. We were well prepared for these questions and were able to generate an exciting discussion from them.

Exciting questions from the audience

Some of the questions were, for example:

How close are we to fully generated AI films and perfected images?
Is it better to use large context window-based language models to deal with large content, or is it better to use RAG models?
Kolmogorov-Arnold networks? Could they be an improvement on the architecture of classical neural networks?

Conclusion and outlook

The convention is a very important institution for bringing the general audience into the digital industry. The exchange of knowledge on such a large scale, interesting discussions and making valuable contacts, especially in Mecklenburg-Vorpommern, is to be encouraged. For these reasons, we were very grateful to be part of the NØRD Convention and look forward to next year.

Bisrats tips for successful workshops

#1 Know your audience

Over the years, we have learned from our workshops and presentations that it is important to adapt dynamically to the audience. In workshops and talks, it is very important to know who the audience is, what their experience is and what they are particularly interested in. Take LLMs for example: you can talk about the technical issues of transformers, how they are trained, and reinforcement learning from human feedback. But if the audience is more interested in how to use the currently popular LLMs for a specific business use case, then the workshop would be a challenge.

#2 Preparation is everything

Our experience over the years when giving workshops and talks is that it is important to dynamically adapt based on the audience. Knowing who your audience is, what their experiences are, what they are rather interested in is a very important part of giving workshops and talks. Take LLMs for instance: One can talk about the technical topics of transformers, how they are trained and reinforcement learning from human feedback. But when the audience is rather interested in the topics of how to actually use the current popular LLMs to create a certain business usecase, then the workshop would be challenging.

#3 Keep it simple

As varied as the audience could be, one ingredient that always works is explaining things as simple as possible. Everyone loves well thought out explanations supported with engaging analogies, examples and good statistics. The presentations should also always be simple, contain only few words to say and more visual. The old saying always holds in that, a picture is worth a thousand words.

#4 Engage and trigger

Unless there is a time constraint, we mostly start our workshops by asking the audience about themselves a bit. That turns out a bit more engaging and the audience feels to get a more personal touch. Additionally, it is a good idea to prepare some engaging questions. In one of our conferences, at NConf, the audience stayed for about 20 more minutes discussing among themselves about the future of generative AI, which was rather beautiful to see.

Previous slide

Next slide