Congratulations on becoming tenured CISPA-Faculty! What does this milestone mean to you?
Thank you. To me, becoming tenured Faculty is a sign that I’m getting older, not only literally, but intellectually as well. It’s truly an important milestone in my academic career because it gives me the freedom to pursue ambitious, long-term research questions and to do so more boldly.
What has CISPA given you as a researcher that you might not have found elsewhere?
Working at CISPA is special to me because it provides an environment that allows me to devote my most valuable resource, time, to the research I believe can have the greatest impact. I decided to join CISPA four years ago because its vision and unwavering commitment to research excellence strongly resonated with me.
On top of that, as a young and rapidly growing research center, CISPA offers me a unique opportunity not only to pursue ambitious research, but also to help shape its future and contribute to something far greater than myself and my own research group. Equally important are my colleagues, both researchers and members of the administrative staff, whose support, collaboration, and positive spirit make CISPA an exceptional place to work.
I am also especially grateful to the current and former members of the Rational Intelligence Lab, who have turned what could have been a rather solitary research journey into one that has been both enjoyable and deeply memorable.
Looking ahead, what are the fundamental challenges in machine learning that you want your research to address?
I often compare the way we train today’s machine learning models to the prisoners in Plato’s Allegory of the Cave. Like the prisoners, our models learn from only a limited view of reality and must infer how the world works from incomplete observations. Yet, unlike the allegory, they rarely have the opportunity to step outside the cave and verify whether what they have learned truly reflects reality.
This is what motivates my research. In the coming years, I want to better understand how we can build machines that learn rationally when knowledge is incomplete, incentives are misaligned, and the world itself refuses to stay still. If we want AI systems to operate reliably in the real world, they must be able to reason under uncertainty, anticipate strategic behavior, and adapt to environments that change over time.
Traditional machine learning often assumes that we have complete knowledge of the learning problem at hand. Imagine an AI system trained to detect skin cancer using images collected from one hospital. If we assume that this dataset represents every patient, the system may become overconfident when diagnosing people with different skin tones or images from different cameras. The result can be missed diagnoses or unnecessary treatments.
At the same time, as machine learning systems become more powerful and widespread, we need to think carefully about the incentives of the people who interact with them. Imagine an AI system that helps decide who should receive a loan. If it is trained on historical decisions that contain human biases, it may simply reinforce those biases, making it harder for certain groups to obtain a loan. But once people know that an AI is making the decision, they may also learn how to game the system to improve their chances of approval. As a result, the AI is no longer operating in the same world it was trained on.
Many AI systems no longer operate in isolation; they are embedded in larger social and economic systems involving millions of users. Their predictions and decisions can influence people’s behavior and well-being, for better or worse, while also creating opportunities for strategic manipulation.
Finally, the world itself doesn’t stand still. Sometimes it changes precisely because of the predictions our models make. Imagine a navigation app recommending the fastest route. If millions of drivers follow the same advice, the route becomes congested, and the prediction is no longer correct.
This challenges one of the fundamental assumptions in machine learning: that the environment stays the same. In reality, AI systems increasingly influence the very world they are trying to predict. If we want them to perform reliably, we must design them to understand and adapt to these changing interactions.
Ultimately, my goal for the next few years is to develop the foundations for AI systems that can reason about what they do not know, understand the incentives around them, and adapt rationally to a changing world.
Your research group is called the Rational Intelligence Lab. What makes an AI system not just intelligent, but rational?
To me, rational intelligence goes beyond pattern recognition. Today’s AI systems are remarkably good at identifying patterns in data, but intelligence should involve more than recognizing correlations. A truly rational system should also understand why those patterns arise and how they relate to the world.
At a deeper level, such a system should understand the limits of its own knowledge. It should recognize when the available evidence is insufficient and be willing to say, “I don’t know,” rather than respond with unwarranted confidence. This idea is often referred to as epistemic humility. It echoes a famous insight attributed to Socrates: true wisdom begins with knowing what you do not know.
For example, a rational system should recognize that the same observed pattern can often have more than one possible explanation. If ice cream sales and drowning incidents both increase in the summer, a rational system should not conclude that buying ice cream causes drowning (hence, ice cream sales should be banned). Instead, it should consider alternative explanations, such as hot weather driving both.
A rational system should also be able to reflect on its own reasoning and learn from its mistakes. Much like a scientist revises a hypothesis when new evidence appears, an AI system should be able to update its beliefs instead of stubbornly sticking to its original conclusions.
It should also be resilient to manipulation. For example, a self-driving car should not mistake a stop sign for a speed limit sign simply because someone has placed a few stickers on it. Likewise, an AI assistant should recognize when a user is trying to manipulate or exploit its reasoning.
Finally, it should be able to fill in missing pieces of knowledge in a coherent way. Just as a child who has learned how bicycles and motorcycles work can make reasonable predictions about an electric scooter without having seen one before, an intelligent system should be able to generalize from what it has learned to unfamiliar situations.
If you look beyond individual projects, what is the big question you ultimately want your research to help answer?
One question that keeps me awake at night is, as we continue to learn, will the gap between what we know and what we don’t know ever truly disappear? There are two possible attitudes to this question. One optimistic perspective, shared by several prominent philosophers, is that learning steadily reduces ignorance. The things we don’t know today are simply facts we haven’t discovered yet. So with more data, better models, and greater computing power, the unknown becomes known, and the gap between knowledge and ignorance keeps getting smaller.
A more skeptical view is that learning doesn’t eliminate ignorance. It makes us more aware of it. Every scientific breakthrough answers some questions, but it also opens many new ones. Every model simplifies reality, revealing assumptions we hadn’t noticed before. And every technological advance creates new challenges we didn’t anticipate. In this view, knowledge doesn’t just shrink the unknown, but also expands the frontier of what remains to be understood.
I believe the truth lies somewhere in between. Science helps us answer important questions, but every answer also uncovers new questions we didn’t know to ask.
Machine learning has followed the same pattern. As our models become more powerful, they solve problems that once seemed out of reach. But they also create new challenges that we didn’t have to worry about before. For example, AI systems can confidently make things up, become unreliable when the world changes, be manipulated by users, or even change the very environment they are trying to predict. In other words, as AI becomes more capable, the nature of its uncertainty also becomes more complex.
I believe this question has become increasingly important because we are entering an era in which AI is becoming an essential partner in how we acquire knowledge. It has made generating information faster, cheaper, and more accessible than ever before. But at the same time, it has made the boundary between genuine knowledge and convincing misinformation increasingly blurred.
The challenge, then, is no longer just acquiring more information. It’s knowing what to trust, and recognizing the limits of what we truly know.
Whether the gap between knowledge and ignorance will ever truly disappear remains an open question. I suspect it won’t. But if that’s true, then our goal should not be to eliminate uncertainty, but to build AI systems that can reason responsibly in its presence: systems that know what they know, recognize what they don’t, and act accordingly. To me, that’s the promise of Rational Intelligence.
Looking back at your time at CISPA so far, what are you most proud of?
I would say my greatest achievement has been finding research questions that I genuinely care about and am willing to dedicate my career to. Having a clear sense of purpose has given me the confidence to make long-term investments in challenging problems.
Equally important has been building a team around that vision. I’m grateful that many talented young researchers have chosen to join me at CISPA. To me, that’s one of the most rewarding aspects of academia: not only advancing knowledge yourself, but also creating an environment where others can grow and do their best work.
Does tenure change the kinds of questions you feel free to explore or how you choose which questions to pursue?
I honestly don’t know. The more I learn, the more I realize how much remains to be understood. Rather than following a fixed roadmap, I try to follow my curiosity. Right now, that curiosity is leading me toward philosophy, epistemology, metacognition, and perhaps even quantum computing. Where it ultimately takes me, I don’t know, but that’s one of the joys of being a researcher, and one of the greatest privileges that tenure provides.
Krikamol, thank you for these thoughtful answers and the insightful interview.
About the tenure track program at CISPA
CISPA has set up a special career development program to give aspiring scientists a long-term perspective. It is very similar to the tenure track program that the federal and state governments have been establishing at German universities to promote young academics since 2016. In the CISPA tenure track, researchers must prove that they can hold their own in international competition. This includes submitting publications to top conferences, supervising junior researchers, establishing collaborations and acquiring third-party funding. If they succeed, they are appointed as "tenured" Faculty, i.e., leading scientists "for life".
