���ʵ���

How did you end up back at Aalto?

One trigger was the news that was being established. I thought that the new institute will bring new experts and create a positive buzz around artificial intelligence in Finland and at Aalto. And as a family we had decided that it is time to return to Finland from Switzerland.

In my 2018 dissertation I studied the linkage of historical records and how they can be structured into family trees. Though I used machine learning and AI methods in my doctoral research, there was still a lot of manual work. I had to clean up data and get it into a certain format. Large language models can help automate a lot of processes. You can give a model raw data and ask it to find connections.

I think artificial intelligence (AI) and language models will change our society a great deal, and I want to be shaping that change to be as positive as possible. I also want to help raise up the next generation of AI experts. That’s why I’m interested in teaching. In my free time I already taught one course a few years ago.

I principally work at Google DeepMind four days a week. One day a week I spend as an adjunct professor at Aalto and an affiliate of ELLIS Institute Finland. In both cases I study large language models, but in academia, I will focus more on interdisciplinary applications, like using AI in the fields of social sciences and humanities. There’s a lot to do in this space, because the companies developing language models rarely focus on applications such as the study of history.

What do you expect from the future at Aalto?

I’ll start teaching a seminar course on large language models soon. It’s a special course for master’s and doctoral students covering the technical side and the latest publications. The course is meant for students who have already completed some machine learning courses.

How is the rap algorithm doing?

It was published at Aalto about ten years ago. The generator has been used by over a million people and it’s still going strong. If there’s been a service break, I’ve gotten a lot of frustrated messages. I’ve also done one follow-up study on generating lyrics using AI.

Hobbyist rappers use the generator to support their creative process. It gives them ideas for good lines. They don’t just press a button and use the output as is, but they use it to bounce around ideas and find what rhymes with their own text.

At Google my first interviewer said they had read a news article about the rap generator. I don’t think that it influenced the hiring decision, but it gave me confidence in the interview. When I moved to the Gemini team at Google, my first project was improving its lyrics generation capabilities. The rap algorithm was a pivotal side project in my career that led to me studying language models at Google, a path I’m still on.

What do you think about AI and creativity? 

I don’t think that AI is developed to replace people but rather to support creativity and intelligence. AI can be a chat companion, an expert in many fields that most people wouldn’t otherwise have access to.

AI can be seen as an opportunity, in the same way as the synthesizer. At first, some musicians feared being replaced by it, but it has opened completely new possibilities to express yourself. In a similar fashion, AI creates new opportunities to express yourself for example through vibe coding without having to practice coding for years.

What do you think are the most important traits for researchers and teachers?

For the researcher, curiosity about new things and a desire to go below the surface. For teaching, the ability to put yourself in someone else’s shoes and trying to see the problem from another point of view, to tailor the message suitably for another person.

What has been the highlight of your career?

At Google DeepMind I lead a team of researchers and software engineers that focuses on training Gemini models. More specifically other teams try to get as much information as possible into the models in the pre-training phase, and in the post-training phase we try to make the models as useful as possible for people. This allows the models to provide more useful answers and communicate more effectively.

Lately, for example, we’ve studied how language models can give better answers given more time to think. At Google DeepMind we recently applied this to chess and other board games. Language models have traditionally been poor at games that require long deliberation, but we succeeded in developing a grandmaster-level chess AI based on language models. It isn’t yet better than the best human players, but I don’t stand a chance against it, even though I’ve played chess at the national level when I was younger.

Some didn’t think language models could develop deep reasoning skills in the first place. Critics kept repeating that the models can’t even play legal moves. 

I was asked to lead a team of 15 for this chess project, which was a technically really challenging problem. The first time I lost to our language model I got the feeling that we had succeeded in creating something bigger than ourselves. I got a chance to demonstrate it for the CEO of Google DeepMind.

Contact information:

Eric Malmi
Adjunct professor
Aalto University
eric.malmi@aalto.fi