“It is extremely fast to test an idea today. In the past, weeks could pass from the moment I had an idea in my research until I knew whether it worked. With the latest AI tools and the compute power I now have access to here at Stanford, I can often get an answer within a few hours. The tools help writing the code based on the instructions I give.”
That is according to Anders Gjølbye Madsen, a PhD researcher who is currently on a research stay at Stanford University in Silicon Valley, California.
Although he is only 25, Anders is already a well-known profile within DDSA. He is a member of DDSA’s Young Academy Panel, and a former participant in the Pre-Graduate Retreat. He has also received funding from DDSA, including a fellowship grant for his ongoing PhD at the Section for Cognitive Systems, DTU Compute.
His research is broadly about understanding the decisions behind large AI models, including language models such as ChatGPT, Mistral or Copilot, with a focus on how they can be used more reliably in healthcare. More specifically, the title of his PhD project is:
“Causal Approach to Trustworthy Artificial Intelligence in Healthcare.”
At Stanford, he works in a building with hundreds of other AI researchers and developers, many with the same screen setup in front of them: two large monitors. On one screen, several AI tools are often open at the same time, including Claude Code, which is used for coding alongside other tools depending on the task. Together, they make it possible to move faster in research, from coding and experiments to literature searches and testing new ideas.
Still Way to Go for Developing AI in Healthcare
“The goal of my PhD is to help increase trust in AI in healthcare, which is an area I care deeply about, especially within neuroscience. I hope AI can become a trustworthy support tool for doctors, for example in diagnosing epilepsy or assessing brain tumours. Of course, it must always be a human being, and ultimately a doctor, who holds responsibility. There is still some way to go for developing better AI in the Danish healthcare system, but progress is being made. That is also reflected in the fact that the course about Medical Data Literacy is now mandatory for medical students at the University of Copenhagen,” says Anders Gjølbye Madsen.
His PhD is expected to be completed in the summer of 2027, and he has spent most of the programme in Denmark in contact with Rigshospitalet.
“It is important to involve doctors and other domain experts. AI should be a tool, and a central part of my work right now is understanding what is happening inside these systems, in other words why, they arrive at the answers they do, for example in a diagnostic context. I am investigating whether AI can better understand causal relationships and provide explanations that are meaningful to doctors. Here at Stanford, I am working to improve AI in an environment with many strong computer science and data science researchers, including researchers with experience from companies such as Google, OpenAI and Anthropic,” says the Ph.D. Student.
Access to Compute Power is a Parameter
According to Anders Gjølbye Madsen, Denmark is strongly positioned in data science and AI. But his stay at Stanford has also made it clear to him that the pace of research is not only about talent. It is also about access to tools, compute power, and an environment where new ideas can be tested quickly.
“The level in Denmark is high, but we also need to recognise that access to the right tools requires investment. Over the course of my PhD, I will personally spend significant resources on AI tools. My advice is that Denmark should be more curious about the newest opportunities and not be afraid to invest in them. If you are working with open research and non-sensitive data, there are good reasons to make use of the latest technologies as well.”
This applies especially to access to the kind of compute power that allows large computations to be run quickly and new ideas to be tested without long waiting times.
“We have Gefion in Denmark, and in Finland there is LUMI, but access is more cumbersome because you have to apply and describe the experiments you want to carry out. That takes time. Here at Stanford, I can often test something over the course of an afternoon. In the lab I am in now, I have access to larger machines and more powerful GPUs than I have personally had access to before,” he says.
A GPU, or Graphics Processing Unit, is a specialised chip that can perform many calculations at the same time, and is therefore central to modern AI, where large amounts of data need to be processed quickly.
For Anders Gjølbye Madsen, the point is not that Denmark lacks skilled researchers. The point is that strong research environments also require infrastructure and a culture in which ideas can quickly be turned into experiments.
He has a personal appeal to Danish data scientists:
“Denmark already has a strong AI and data science community, and it could become even stronger if we make greater use of one another to exchange experience. That can happen through DDSA’s initiatives, but also by getting involved in student-led initiatives. Cross-disciplinary collaboration is especially important if data science is to help strengthen many different industries and sectors in Denmark.”