My name is Balázs, but everyone just calls me Balu. I am a Master Student in Statistics at ETH Zurich, interested broadly in Data Science, Machine Learning and AI, with a particular focus on Reinforcement Learning and Large Language Models.

The incredible odyssey I am about to tell you about all started one day back in the spring of 2025, when I saw on the Instagram story of ETH that UNDRR (United Nations Office for Disaster and Risk Reduction) was looking for Data Science and AI Interns in Bangkok. I immediately knew that I had to apply. While traveling around in Southeast Asia for two months two years ago, I fell in love with Asian culture. Considering that in my studies I put a strong focus on Data Science and AI, I thought this is the perfect opportunity for me!

Fast forward 6 months, I am standing at the gates of ESCAP (Economic and Social Commission for Asia and the Pacific) in Bangkok, ready to be picked up and start my internship. I quickly got to know my team and all the other interns. I was surprised to see how many other interns there were across all the UN agencies in Bangkok. This made it very easy to find friends and build a social circle right from the beginning. What I valued right from the getgo, was the diverse background of all the people around me, be it culturally or academically. Most of the other interns were in the same phase of life as me, hovering somewhere around finishing their Master’s Degree. The majority had a background in social sciences, such as economics or law. At ETH we sometimes forget that there is also a world out there that does not care about math as much we do! It was very nice to be surrounded by people that look at the world from another angle compared to most STEM students. This helped me broaden my horizons in many aspects of life.

The start of my internship was fairly slow. Because the other two interns (Bilal, also an applied maths student from ETH, and Siwei, studying in Stockholm) and I were the first three AI interns ever at UNDRR, it was a lot of figuring out on both ends. As per usual, the first few weeks I was busy clicking through all the online onboardings. Who knew that as an intern in the IT team, I needed to know how to behave in case I am trapped in a mine field in the desert. But I soon got introduced to my main project: DATUM, UNDRRs RAG-based LLM. In a nutshell, this is a language model like ChatGPT, that builds on internal data at UNDRR. The language model can retrieve information from documents from the database of UNDRR that is relevant for the question a user may ask.

To my big surprise, after some weeks I was named the Project Manager. It felt quite surreal being given this responsibility as an intern. But I quickly realised that this internship would not be like any other internship. I doubt that there are many other internships that give you the same amount of responsibility right from the start as we have been given. This was both, challenging and also rewarding, at the same time. Considering the fact that AI was not the main focus of our team and something most members were not deeply familiar with, we interns were the ones with the most expertise and hence we were the ones who built the foundation for future interns or even full time employees. This is one of the things that I am most proud of: setting something in motion that will hopefully live on way beyond the end of my internship. To educate the other members of our team on topics revolving around AI, we held bi-weekly AI-talks, where we interns gave presentations about AI-related topics that we deemed interesting and relevant.

One of my proudest moments was when Jeanette Elsworth (Chief of the Communication, Advocacy and Knowledge Management Section of UNDRR) came to visit us in Bangkok and I had the opportunity to present to her the progress of our work. Within the three to four months that I have been there at that point, we had a prototype running of our RAG-based LLM. It was really rewarding to see the engagement and excitement of the whole team. For the remaining part of my internship, we further improved our model and developed an interface for it for internal use.

Another thing that I appreciated and made use of, was, that I could attend most of the conferences held at ESCAP. One of these conferences was the “AI for Developing Countries Forum” (https://af.net/bangkok/). By attending this conference, I learned a lot about the human aspect of artificial intelligence, and the imbalance in power in the AI race that is ongoing worldwide. The two most striking learnings for me were:

1) 60% of web content is written in English, whereas only around 16% of the world population speaks English (https://pollicy.medium.com/the-digital-divide-extends-to-language-what-now-919da821389e). This leads to an immense bias towards the English language when training LLMs on all of the data available on the web.

2) The CLOUD Act: a law that allows American federal law enforcement agencies to access data stored on servers from US-based tech companies (Meta, Google, Amazon, etc.) regardless of the location of the datacenter, be it in the US or abroad (https://en.wikipedia.org/wiki/CLOUD_Act, https://www.congress.gov/bill/115th-congress/house-bill/4943, https://www.govinfo.gov/content/pkg/PLAW-115publ141/pdf/PLAW-115publ141.pdf). Learning about this law made me question all the privacy claims that these big tech giants make. Attending this conference really broadened my horizons and gave me a new perspective on AI and data sovereignty. 

Outside of work, I had the opportunity to make Bangkok, the concrete jungle of Thailand, my home and engage in all the craziness and chaos it has to offer. I also had the privilege to travel around the country and explore it’s beautiful beaches, its underwater world while scuba diving, its majestic temples and most importantly, get to know the immense kindness of the Thai people. Many weekends and long evenings well spent that I will keep in good memory for the rest of my life.

At this point, I want to say a big thank you to all the people that have made this experience so amazing and unforgettable. Some people I want to mention explicitly include my direct supervisors, Anders Hansen and Revati Mani, my coworkers with whom I worked together on DATUM: Ryan Sandigan, Bilal Bounajma and Siwei Zhang and Camilla Andreozzi; and all the other interns that I’ve met along the way.

Lastly, none of this would have been possible without the support of ETH Studios. I am immensely grateful to have been given the financial support necessary to embark on this once in a lifetime opportunity.

What does it take for organisations and individuals to keep innovating when technology cycles speed up, hype multiplies, and AI can generate “answers” in seconds?

Those were the questions at the heart of the ETH Global Lecture “Empowering Innovation: Thriving in the Age of AI”, where ETH alumna and executive Petra Ehmann and ETH Professor Stefano Brusoni explored how innovation actually happens in practice, and what changes when AI enters the picture. The conversation was moderated by Chris Luebkeman, Head of ETH Zurich’s Strategic Foresight Hub.

Innovation is not a vibe. It’s staying power.

Petra Ehmann offered a definition that cuts through the buzzwords: innovation is not about novelty for its own sake — it’s about longevity.

“To me, innovation is giving longevity to a company, making it invincible so that it will thrive in the future.”

She also dismantled a few familiar myths:

“Innovation is not a bunch of geniuses who sit in the room and they tinker something crazy… Innovation is not a fun, colourful park with nice cushions… It’s hard work.”

The subtext is important: real innovation is less about theatre and more about customer value, iteration, and the discipline to keep going when outcomes are uncertain.

The winners aren’t occasional innovators, they’re systematic ones.

Stefano Brusoni highlighted a difference many organisations underestimate: the gap between having one good idea and building the capability to innovate repeatedly.

“Very few companies are really good at becoming systematic innovators.” 

And he described the unglamorous reality of how that capability is built:

“It’s this process of trial and error and then error and then again an error where you stay persistent… that really set apart the occasional innovators from the systematic innovators.” 

Petra echoed this, adding that innovation is fundamentally a learning process — and learning is rarely “efficient” in the short term because it involves mistakes.

Culture change starts at the top.

One of the most concrete practical lessons came through in the discussion about culture: if leaders demand innovation but punish missteps, experimentation dies.

“Permission to fail… is super important and it has to come from… the board.” 

This matters even more in the age of AI, when output is cheap and prototypes appear instantly. Evidence becomes the antidote to building impressive things nobody wants.

AI won’t change the fundamentals, but it will change the speed.

Both speakers agreed: innovation still depends on building something people need, but AI changes the velocity.

“What will change is the velocity, how fast you’re getting to that point.”

Petra also predicted a shift towards hyper-personalisation, to which Stefano added a thoughtful distinction: unlike older tools (such as calculators), generative AI works best when humans share the why, not just the what, which turns problem-solving into a dialogue.

“Suddenly we are engaged in a conversation… that is actually shaping the problem.” 

But with speed comes risk: the internet can fill with “cringe” content at scale, and organisations can confuse quantity with impact. Petra put it bluntly:

“The big question is not how much you do, but what is the impact of it.”

The real competitive edge: skills, critical thinking, and the courage to upskill.

A recurring theme was that thriving in AI-era organisations requires more than familiarity with the latest tools. Petra’s advice: cultivate depth and judgement.

“The most important one is having some very deep knowledge that AI cannot have… [and] critical thinking.”

Stefano broadened that into a challenge for universities:

“Our job is… making sure that… our students have more opportunities than at the beginning of their studies.” 

In the end, the lecture made one thing feel refreshingly clear: AI can help you move faster, but it can’t tell you where to go. Direction still comes from a North Star, from critical thinking, and from teams that have the permission and the desire to be curious and continue to learn.

Check out the pictures from the event here: Meet ETH Flickr

And in case you missed it, watch the talk here:

At the latest ETH Zurich Global Lecture, Liisa-Ly Pakosta, Estonia’s Minister of Justice and Digital Affairs, and incoming Vice President for Knowledge Transfer and Corporate Relations and digital ethicist Professor Effy Vayena, joined Chris Luebkeman to discuss how trust, transparency, and practical regulation are shaping the future of AI governance in Europe. Drawing on Estonia’s digital transformation and Switzerland’s sectoral approach, the panel explored the challenges and opportunities of implementing the EU AI Act.

How do we govern artificial intelligence in a world where technology evolves faster than regulation? This was the central question at the recent Global Lecture, “The EU AI Act in Practice: Cross-Border Perspectives on AI Governance,” hosted by Chris Luebkeman, Head of the Strategic Foresight Hub at ETH Zurich. At the start of the event the room was packed, making it one of the most well-attended lectures of the year, with a strong Estonian delegation and Estonian students keen to show support for their country’s inspiring representative, in addition to a curious ETH audience.

Liisa shared Estonia’s journey as a digital pioneer, explaining, “After regaining independence in 1991, we had to build an efficient state with limited resources. We chose a digital path, making transparency central to our approach.” Estonia’s investment in universal internet access and digital literacy laid the groundwork for a government where citizens own their data and can see exactly who accesses it—a model that has fostered deep public trust. “Trust is fundamental. People own their data and control who accesses it. This transparency builds trust and security,” she emphasised.

Effy reflected on the broader European context, noting that while many countries have the technical capacity for digital transformation, “regulatory and citizen engagement challenges remain.” She highlighted that transparency and clear benefits are key to building trust, regardless of a country’s starting point.

As AI systems become more embedded in daily life, the question of trust extends beyond government to technology itself. “We typically trust people, not machines,” Effy observed. “For AI, we seek reliability, safety, accountability, and transparency. Citizens want assurance that systems are safe and that someone is accountable.”

Liisa described Estonia’s evolving AI strategy: “We realised the need for full government control over sensitive data. We oppose backdoors in systems, even for law enforcement, to maintain trust and security.” She also stressed the importance of defining what AI should not do: “Declaring what is off-limits is crucial for maintaining trust.”

The EU AI Act aims to regulate AI in line with European values, but its complexity and the challenge of keeping pace with technological change were recurring themes. “The Act is complex, and simplification efforts are already underway,” Effy explained. In contrast, Switzerland, not an EU member, has opted for sectoral rather than horizontal regulation.

Liisa added, “Estonia advocated for more sectoral rules and a better balance between innovation and regulation. While rules are necessary for trust and clarity, the Act’s implementation is unclear, especially regarding prohibited and high-risk areas. Regulatory burdens often fall hardest on small and medium enterprises.” To address this, Estonia is developing a law to guarantee compliance for companies, shifting the burden from businesses to the state, and offering a compliance sandbox and state guarantees for AI products.

To close the discussion, Chris invited the speakers to a series of rapid-fire questions, offering concise insights on some of the other pressing dilemmas in AI governance:

• What builds trust faster: strong regulation or strong transparency?
  “Transparency,” said Liisa.
• How do you manage security and backdoor access?
  “Use separate, secure architectures like X-Road.”
• How will Estonia navigate differences with the EU?
  “With courage.”
• Why is there a discrepancy between attitudes toward pharma and AI regulation?
  “Pharma regulation has a longer history. We’re still defining our goals in AI,” Effy noted.
• Do we need a UN convention on AI?
  “Statements exist, but implementation and enforcement are key.”

Looking Ahead

The Global Lecture made clear that responsible AI governance is not just about rules—it’s about building systems that earn and deserve public trust. As Liisa concluded, “Despite technological possibilities, regulation must enable innovation while protecting individual freedom and imperfection. Estonia welcomes collaboration and research in this area.”

The conversation underscored that while the EU AI Act is a significant step, the journey toward trustworthy AI is ongoing—and will require courage, transparency, and a willingness to learn across borders.

Chris closed the event by thanking the panel for their “clarity, depth, and inspiration,” adding, “It’s been a joy, and I’ve learned a lot.” His words captured the spirit of the session: open dialogue, shared learning, and a commitment to shaping a trustworthy digital future.

Check out the pictures from the event here: Meet ETH Flickr

And in case you missed it, watch the talk here: