Almost a year ago, I stumbled upon an Instagram post from ETH Zurich about an internship offer for the United Nations Office in Bangkok, more precisely at UNDRR (Disaster Risk Reduction). The 21-year-old, applied Maths master’s student I was at the time, with close to no professional experience, did not think for a minute I had a chance. For some reason I decided to I read the offer and noticed met all the requirements, so I gave it a shot.

I applied internally with the ETH Studios, passed this first screen, then with the UN, got the interview, did the interview, got the position, each step being a complete surprise for me and I still was a bit baffled (and jet-lagged) on my first day in office in September. Everything was new (and a bit scary) to me and the onboarding process was a lot to assimilate.

I would describe this internship as a double way learning and teaching experience: I learned from my supervisors and my team (the Tech & Design team of UNDRR) how a big international agency works, and what current technologies are used in disaster mitigation. In turn, I got to teach them a bit about I learned at ETH: neural networks, LLMs, reinforcement learning… and how these tools can be used in the organisation.

As I was the first ever data and AI intern, I had to build from scratch with minimal resources and no pre-existing workflows. We figured out a lot of things on the go. The arrival of two other new interns at the same position as me, one of whom is also an ETH student, helped to speed up our work.

I primarily worked on two projects: a machine learning based catastrophic model for fragility assessment during seismic events, which I did alone, and a RAG chatbot based on UNDRR, with 3 other people. We were able to deliver working Proof Of Concepts for both, which was a big success given our modest beginnings. I was also the first presenter of a series of webinars about AI related topics, which are still delivered weekly to the entire UNDRR staff to this day. 

This goes in hands with what I would consider as my biggest accomplishment during this internship: planting the seeds of a proper machine and deep learning service at UNDRR (which was the vision our team’s boss had in mind when creating the internship offers). By the end of my six months, the proof of concept opened up opportunities for potential partnerships and funding to continue the work. A new ETH student is also starting her internship right when I finish and I personally see a real potential in a closer ETH/UNDRR partnership.

On a personal level, those past 6 months were an amazing adventure, full of new people from all over the world and travels around Thailand and Asia in general. Bangkok ended up feeling like home and coming back to Switzerland and its 9 million people (half of the population of metropolitan Bangkok) will certainly be quite the change; that and the price of groceries!

It feels like I took off for BKK airport only two weeks ago but in the meantime, the person I am now is fundamentally different from the student who arrived a bit late to work with his brand-new suit in September. I guess this is how growing up feels like.

ETH Studios

ETH Studios are global innovation outposts where ETH students tackle real-world challenges alongside companies, researchers, and startups. Whether in New York, Heilbronn, Davos, or at the experimental StudioX, students gain entrepreneurial and interdisciplinary experience while contributing to sustainable and tech-driven solutions.

Contact eth-studios@ethz.ch if you are interested in more information on and about the ETH Studios.

In order to understand how I, a Master student in mathematics, ended up working with real trees (not just the mathematical objects, we have to go back in time to my very first lecture at ETH Zurich. The subject was “Mathematics of Data Science”, and I attended it by coincidence, just because I was told it should be amazing – and it was! But not only was the lecture great, but I also discovered that this topic had exactly the flavour of mathematics I loved.

However, much to the dismay of the computer science students, the lecture stuck strictly to its title, meaning we only learned about the MATHEMATICS behind Data Science and never applied the techniques to real data. When I saw an ad for a short data science internship at the SLF (Institute for Snow and Avalanche Research), I got immediately excited. I wanted to use my newly acquired knowledge on real-world data, and, to my even greater excitement, the data would be in the field of environmental science, a field that I’ve always been interested in. So, just a few mails, a short video call and six months later I was sitting in a train that climbed its way up to the mountain town of Davos for the start of my internship.

The project I would spend the next five weeks working on, was about detecting slow movements of mountain slopes by using tree ring data. You might wonder why we are not using a more straightforward approach, like remote sensing techniques, and instead work with trees? These are wonderful questions!

 The first thing one has to know about slope movements is that they are processes that happen over hundreds to thousands of years, whereas remote sensing data is often only available for the last decade.

Therefore, the available data covers a too short period of time to understand these very long-term processes of slope movements well. And that is exactly where trees come in! Trees are widely available on slopes in Europe and can be centuries old. So, if we could use trees as sensors for slope movements, this would extend our data from just a decade to potentially centuries.

And how can we use trees as sensors for slope movements?

The idea behind this approach is that naturally trees always try to grow straight upwards. However, if the slope beneath the trees moves, this tilts the trees and to stand straight again, the trees grow asymmetrically in the following years. This asymmetry can be measured by comparing the differences in tree ring width from two different directions of the trees (don’t worry, there were no trees cut down to collect the data, only two small cores were collected per tree. Sounds like a crazy idea. It is! But it was successfully used by other research groups to detect rapid slope movements. Which leads us to the question: Can we do the same for slow slope movements?

During my five-week journey to answer this question, I tried to apply many of the techniques learned during my course of Data Science. I learned more than I ever thought I would about bark beetles and polar expeditions from chats during lunch. Unfortunately, regardless of what I tried, I could not find a clear relationship between the tree ring data and slow slope movements. However, unexpectedly, I discovered a more rapid event in the data that turned out to coincide perfectly with a documented mudflow in the corresponding area. So, even though I could not answer the initial research question, we could prove again that the method works in the setting of more rapid events.

In addition to my project itself, I had a wonderful summer time in Davos with many after-work hikes and bike tours, I got to know so many nice people from different fields and talking with them really broadened my horizon of what research apart from mathematics can be like. Great thanks go to my supervisors Dr. Alessandra Bottero and Dr. Andrea Manconi who gave me the opportunity to be part of their team and ETH Studio for enabling me this enriching experience! 

And if you got curious, then give it a try and apply as well😊!

ETH Studios

ETH Studios are global innovation outposts where ETH students tackle real-world challenges alongside companies, researchers, and startups. Whether in New York, Heilbronn, Davos, or at the experimental StudioX, students gain entrepreneurial and interdisciplinary experience while contributing to sustainable and tech-driven solutions.

Contact Francesca Caggiula at eth-studios@ethz.ch if you are interested in more information on and about the ETH Studios.

From the first day, the excitement was clear. This was a debut not only for us as participants, but also for ETH Zurich Campus Heilbronn, whose team poured heart and precision into every detail. Their warmth and constant support made us feel welcome. From diverse keynotes and hands-on workshops to thoughtful activities and perfectly timed Lichtfest, the week felt thoughtfully crafted.

Each morning began with Smart Start Breakfasts, coffee, food, and short “thoughts to start the day” that nudged us to see AI as a paradigm shift in how we think. Throughout the week I learned from researchers at ETH Zürich and beyond, seeing how AI touches climate, health, language, and human–AI work. Talks from TU Berlin, TU Munich, UC Irvine, the University of Tennessee, and the Institute of Science Tokyo reshaped how I see AI’s role across fields. One lecture that truly stayed with me was Professor Torsten Hoefler’s public talk; it felt like time travel through the pivotal moments of modern technology. He argued about how we’ve moved beyond the Age of Data into the Age of Computation, and offered an angle I hadn’t considered before: that LLMs’ “hallucinations” might be seen not as flaws, but as sparks of intelligence and creativity.

Workshops with mentors from ETH Zurich, NASA, and Oak Ridge National Laboratory on foundation models turned hard ideas into tools I can use. Much of that learning happened in small teams; mine, XplorAI, included people from Kosova, Portugal, Ecuador, and China. We worked hard, laughed often, and proved that curiosity and good humor help as much as code.

On a very hot afternoon, likely the longest day of the year, we visited LVWO Weinsberg. Among the vineyards we heard how AI helps growers and cellars, from monitoring vines to tuning the process, subtly transforming even the most traditional of crafts. That mix of tradition and innovation mirrored Heilbronn itself, a city changing fast. All week we heard about Dieter Schwarz’s big support to turn the city into an AI innovation hub. It felt like watching a place build its future in real time. 

Perhaps my biggest takeaway was how deeply human this “ML&AI” week felt. We were a global mix of students, scientists, engineers, and thinkers, from all over the world united by open questions and a shared drive to explore them. I left with practical advice, clearer views of the research world, ideas that will stick, and friends. In my farewell note, I wished everyone a symbolic Dieter Schwarz to back their dreams. May readers of this piece find theirs, too.

What challenges arise at high-scale training that you couldn’t imagine, having only experienced lower scale? How can conditional diffusion be brought out to domains where being exact is more important than being creative? Why is it so important to consider uncertainty in modelling?

And more crucially: could Heilbronn, a city of 100,000 inhabitants in Germany that I’m almost sure you never heard of, become the next AI research and development centre of Europe?   

At this very moment, I’m the one taking the lead on this investigation, so let me briefly introduce myself.

Hi, I am Loric Herbé, a regular master student in Data science at ETH Zurich, and for the next few minutes, you will have the opportunity to step into my shoes and discover the backstage of an event that could very well be our most tangible piece of evidence.

Let us first draw some context, so that the investigation board gets clearer in your heads.

On this end of June 2025, a soberly titled “Interdisciplinary School on Machine Learning and Artificial Intelligence for Science” was hosted at the emerging ETH Zürich Campus in Heilbronn. Supported by a range of sponsors, it’s part of a broader effort by ETH to establish a serious research and innovation presence in the region. The goal of this event: gathering talents from around the globe, from master to industry, to deepen their intuition and understanding of the latest applications of machine learning in science. How? By alternating between theoretical lectures, talks, and hands-on sessions that polished their mastery, while it was still hot.

My role regarding this event? With a few other colleagues, we were tasked with coming up with hands-on material, under the shape of Jupyter Notebooks. A challenging experience, as we had to find the right balance between tackling High Performance Computing (HPC) problems, along with giving the right amount of theory and intuition, all the while fitting into time constrained blocks.

What I just described took place mostly prior to the event, which means that on-site, I almost became a regular participant, and this is under this prism that I’m willing to unveil with you the fantastic colours of this five-day-long life fragment.

It would be an absolute crime not to start by giving a glimpse of the talks that we have been given. Without going into details, these talks were pure gold from a Data Science student’s perspective. The amount of insights and creative solution tracks to common problems we face in the field was just surprisingly high, especially given the duration of the talks. And this applies to a very broad spectrum of interest, such as the double-edged sword that is given by constrained optimization, large scale training techniques effects on stochasticity, or even techniques to bring in model interpretability for fields that can’t accept treating a result as a black box’s answer, and this is only to name a few.

Among all these thought-provoking talks, I want to give a special shoutout to the one of Professor Andreas Krause, which gave an overview of the active learning and reinforcement learning fields, notably through the lenses of optimism in the face of uncertainty. The reason I’m mentioning this talk is that as a Data science student, the in-depth lecture “Probabilistic Artificial Intelligence” given by Professor Krause was probably one of the most valuable perspectives on Machine Learning that I was ever given, and I’ll even go the extent that this also changed some of my perception and modeling assumptions on my life in general.

On another note, I can’t forget to mention what it felt like to have such brilliant minds buzzing around you for a week. And there I’m referring both to the organisers as well as the participants. To give my words some concrete pictures, imagine that in the first 10 minutes sitting in the bus to Heilbronn, I was discovering the current state of the art quantization technique for LLMs at inference time through some talk with another master student whose thesis built around that, and that my first meal revolved around learning a new perspective on model validation through Posterior Agreement with the eminent Professor Joachim Buhmann.  However, don’t get me wrong, although this could be a brain racking experience at some times, this also retained the social development aspect that you could find in a regular summer camp, where you have good laughs around a nice diner and talk about the near future under a nice sunset.

Taking back half-seriously my investigator costume, I would say that it’s hard to answer whether Heilbronn will truly become tomorrow’s AI centre of Europe, but it’s worth pointing out that it’s through such efforts, organising such events, that it will definitely earn some key position in this fast-paced field. 

What a journey we made together! Now, I believe that the most acute readers will feel that there is a missing piece to conclude this investigation. And they would be right, what happened to the first few questions? Sadly, I don’t yet pretend to be able to answer those questions at the same level those eminent professors did. However, I think I might have a solution… can’t wait to walk, alongside you this time, in a future second edition!