This Young Belarusian Stood at the Origins of ChatGPT. Here's What He Did

Even when Dzmitry Bahdanau was a teenager, he showed a great talent for programming. In 2009, the young man graduated from BSU Lyceum and went to the International Olympiad in Informatics in Plovdiv, Bulgaria, where he took 13th place out of 301 participants. 

The young man then studied applied mathematics at BSU and continued to participate in competitions. In 2012, Bahdanau, as part of a three-person team, went to the prestigious ACM-ICPC student programming championship. In the world final, the Belarusian team took fifth place, even outperforming competitors from Harvard — who were seventh.

After BSU, the Homiel native enrolled at Jacobs University in Germany, where he pursued a master's degree, and in 2014, he went for an internship at the University of Montreal. There he met Yoshua Bengio — a Canadian computer scientist and one of the «godfathers» of artificial intelligence, who had been developing the first neural networks back in the 2000s.

In 2014, together with Bengio and Kyunghyun Cho, a PhD student at the University of Montreal, Bahdanau wrote an article through which he entered the history of neural networks, with Dzmitry listed first among the authors. That article was published under the title «Neural Machine Translation by Jointly Learning to Align and Translate».

So what did Bahdanau and his colleagues come up with? It's about teaching computers to perceive text differently. 

Imagine you need to remember a long sentence. If you try to memorize every word individually, you will probably lose some words and won't be able to recall them later. Plus, there's a risk that the part of the sentence you lost contained some semantically important information. It's a completely different story if you try to understand the meaning of the sentence and the connections between words. Then, when you need to recall the sentence, you will do so taking into account the words that are important in the immediate context.

This is approximately what Bahdanau and his colleagues taught neural networks: to determine which parts of the original sentence need more attention.

Remember machine translation in the 2000s and early 2010s? Back then, Google Translate and similar programs could confuse word meanings, lose connections between parts of long sentences, or even change the meaning of statements. But in the mid-2010s, much changed: Google Translate began transitioning to neural network translation, and according to Google's data, translation quality improved by 55-85% over time compared to the previous technology. In explaining the transition to such translation, Google refers, among other things, to the article by Bahdanau, Bengio, and Cho.

However, the implications of that article extended far beyond machine translation, for which Bahdanau's mechanism was originally developed. In 2017, transformers emerged — neural networks that process information largely based on the principles of improved attention mechanisms. And it is precisely on transformers that modern language models, including ChatGPT, are based.

Since Bahdanau's time, however, that attention mechanism has been significantly improved, and this is understandable — as it is applied not only in machine translation. 

But both ChatGPT and Bahdanau share one common idea — the neural network itself chooses which words in the text are the most important. ChatGPT operates on the principle where the program itself asks words many questions and, thanks to training, knows which connections between them need to be considered.

In 2020, Dzmitry Bahdanau received his doctoral degree in Montreal, and in 2021, he headed the Conversational Assistant research program at the ServiceNow startup. He worked, among other things, on systems capable of writing programs and understanding human instructions. 

And from 2025, Bahdanau is a member of the founding team of the large startup Periodic Labs, where they are trying to combine AI with real scientific experiments. Concurrently, he is a visiting professor at McGill University in Canada, where he works with young AI researchers.

The Belarusian's contribution to AI development is significant, but it would be incorrect to say that without him, we wouldn't have ChatGPT and good machine translators. If it weren't for Bahdanau, Bengio, and Cho, other scientists might have arrived at the same solutions. But the world of technologies that have become so ingrained in current life might then have been different.

• 38
• 3
• 0
• 0
• 0
• 0
• 0
• 2