EM: First of all, the primary thing I'm doing is studying. University is my number one objective at the moment. Before that, I worked at two startups – really different jobs and lived experiences that I was fortunate to have.

In the first one, I did research and development aimed at speeding up computations necessary for AI using various mathematical trickery. Apart from research, a lot of my work also involved standard AI development. So basically, looking at algorithms, looking at stuff that could be used with those algorithms and trying to combine them and apply them to new data. The work is very multidisciplinary, so at times I have to get a working grasp of fields in a short amount of time that I have little or no prior experience with, like dynamical systems and control theory.

The second start-up is an IT startup with a focus on machine learning. One project in particular that stuck and was successful was a cattle monitoring system. The idea is to make extensive agriculture, which is a more ethical but less cost-effective type of agriculture, more efficient. The system is used to track and monitor the weight of cows and make predictions about future weight trends. The system is completely non-intrusive, meaning we don’t tamper with the cows in any way.

The working principle of the project was that a cow would come to its waterer and step on the weight scale inadvertently. We use cameras and data from the weight scale to identify which cow is on the scale at any given time and log this information to the cloud. When the cow comes up to drink the water, it steps on the weight scale and tilts its head towards the camera. At that moment, the camera starts collecting images, it recognizes there is a cow, collects the images and basically identifies the cows by their snouts. Cow snouts have an intricate pattern–just like our fingerprints.

Each cow has its own online portfolio containing weight trends and predictions about future weight behaviour. When the farmer or rancher sees that a cow is performing well, like gaining a lot of weight in a very short amount of time, it means that it makes sense to breed this cow. Or maybe on the other hand when a cow has a drastic decline in weight, it can be a signal of disease.

So the rancher can be proactive and actually see what's happening instead of waiting until something happens, like the cow gets sick, dies or something worse, like infects every cow on the ranch. Usually, getting this information is very costly by traditional methods or impossible because they weigh cattle monthly or bimonthly (or even rarer).

The two start-ups took a lot of my time, but I'm happy because it's fun to do and fulfilling seeing my effort have real-life positive outcomes.

EM: I was probably like 12, or 11 – around fifth grade. There aren't any formal competitions until seventh grade where I live. So the first time I could actually compete was in seventh grade. Although I started practising a bit earlier, there were some mathematics programs, basically clubs running some lectures weekly, which I'd go to every Sunday and Saturday. It was a sacrifice in the sense that most of my weekends were spent with math and programming which I didn’t like in the beginning.

Later on, I grew to like them and I really enjoyed the feeling of solving problems. My mom is a math teacher, so she gave me some books and materials to learn from initially. By sixth grade or seventh grade, you would only have yourself to help you and your books, since school teachers weren’t well-informed about competitive math and programming.

Programs and organizations running some weekly lectures helped a lot here, connecting you to a network of students who also liked math and competing. Inspired by that feeling, I co-ran a maths club in high school. My friends and I wanted to work on it together.

We had support from the Ministry of Education in Sarajevo. They recognized our efforts and gave us some funding. Now that I'm in university, I'm still connected to the network and the competitors, and I still enjoy solving problems from competitions.

EM: I did a lot of camps and a lot of international competitions, which gave me a lot of opportunities to go abroad. The most scenic place was the Olympiad in Japan in 2023. We went into the second tallest building in the world, Tokyo Tower. The whole city of Tokyo is so large that the horizon is basically just buildings. When you're so high up, the tiny buildings make the surroundings look like a computer's motherboard.

I wanted to mention the camps because having this community of people that are interested in similar things amplifies everyone's goals and skills. These international camps have helped me get a broader understanding of not only math but different schooling systems also.

EM: A fundamental guiding principle for me is that I want to have the biggest positive impact with my life on the world around me. I find myself invested in the AI safety issue for this reason and because it is an interesting, important, and neglected problem to work on. I've done multiple small projects on AI safety and I’m trying to develop myself more in this direction.

Something tangentially related to the first startup I mentioned is interpretability, the research area where you try to figure out how neural networks and these complex models do what they do. When you understand how something knows what it knows, you can discover malicious intent or unwanted behaviours early before it causes potentially bad or catastrophic situations. I explored this and related areas through the Non-Trivial Fellowship, a fellowship for talented high schoolers.

I worked on an AI safety project for eight weeks called “Constitutional AI via debate,” collaborating with another Non-Trivial fellow. We won the second prize! That was my first attempt at doing AI safety research, trying to invent something in this field and it felt good being recognized and rewarded for the work.

Since then, I've been more-or-less learning, not necessarily inventing things or trying to figure out new approaches to this problem because it's a really hard problem. A lot of people are working on it, like people from Google, Microsoft, and Anthropic, and nobody's sure what the future holds and how to crack the problem of potential AI misalignment. Some think it might be uncrackable, but we'll see where it goes in the coming years.