When Kai Yu Ma joined the exquAIro Biomedical AI bootcamp as part of Class 2 in the fall of 2024, he already had substantial knowledge of artificial intelligence. As a researcher turned educator at the UMCG, Kai had developed much of his knowledge of AI through self-directed learning, including reading papers, watching educational videos and exploring other learning materials. Still, the bootcamp changed the way he understood AI. “It connected everything I already knew into a coherent whole,” Kai says.
Today, Kai combines his role as AI education coordinator and lecturer at UMCG with his work for exquAIro. He is a trainer and member of the editorial board of exquAIro, where he contributes to the development of the programme. He also teaches sessions on large language models (LLMs), helping researchers understand not only how to use these models, but also what happens under the hood.
From regression to neural networks
One of the most valuable insights Kai gained during the bootcamp was a better understanding of the different types of AI models and the principles they share. “Whether you are working with a simple model or a generative AI model, the underlying idea is similar: you have data, you train a model on that data, and the model learns patterns that can be used to make predictions.”
That perspective helped Kai connect techniques he had already encountered in his research with the broader field of AI. Linear regression, for example, can be used to identify relationships in data and make predictions. More complex approaches, such as tree-based models and neural networks, identify patterns in increasingly sophisticated ways.
The crucial difference lies in how those patterns are found. “With a regression model, you have a relatively simple way of fitting a pattern to the data. With a neural network, there can be extremely complex interactions between variables. Conceptually, we understand what the model is doing, but it becomes much harder to understand exactly how it arrives at a particular output.”
This distinction became particularly relevant as Kai explored generative AI and LLMs. Rather than producing a single numerical prediction, these models generate new content. Large language models do this by learning statistical patterns in vast amounts of training data and using those patterns to predict what comes next in a sequence of tokens.
Learning the possibilities and limitations of LLMs
LLMs were already emerging when Kai attended the bootcamp in 2024, and he had actually given a workshop on language models to his fellow participants. For several of them, discovering what these models could do was one of the biggest learning experiences of the bootcamp. That experience has since become part of Kai’s own teaching.
His approach is straightforward: understanding the technology makes it less intimidating. But understanding the technology also makes it easier to recognise its limitations.
One of those limitations is the reliability of the output. An LLM can produce fluent and convincing text without that text necessarily being correct. For Kai, this makes judgement a fundamental AI skill. “You need to make an assessment of how reliable the answer is. And if that reliability is low, you need to find a way to verify whether the output is actually true.”
That means checking sources, understanding where information comes from and critically evaluating the evidence. Even when AI tools provide citations or links, users still need to assess whether those sources actually support the answer and whether the answer is aligned with ones own values and beliefs.
Beyond prompting
Kai: “AI literacy therefore goes well beyond learning how to write effective prompts. Real learning starts when people begin working with AI themselves and develop an understanding of what constitutes a useful or unreliable output.”
This is particularly important because AI systems are evolving rapidly. The fundamental architecture of language models has not changed completely overnight, Kai explains, but the systems built around them have become increasingly sophisticated. Tools now combine multiple models, external sources and other capabilities. Agentic AI, for example, can involve multiple language-model-based components working together to perform tasks and make decisions.
“The underlying technology itself may not change completely, but the way it is organised and the capabilities it provides keep changing.” That makes continuous experimentation essential.
From participant to trainer
Kai’s message to researchers and healthcare professionals is not to expect a workshop to provide all the answers. A workshop can provide principles and a foundation. The real learning starts afterwards: by experimenting, finding out what works and what not, and learning from experience.
For Kai, this is also the challenge for AI education at UMCG. Over the next few years, he hopes AI becomes an integrated part of education and professional practice, not something that is either feared or blindly embraced. “The goal is to understand both the possibilities and the limitations of AI.”