Oriol Vinyals

Some ideas seem obvious after someone makes them happen, and that's exactly the trap. In 2014, Oriol Vinyals, a Catalan researcher who had just finished his doctorate at UC Berkeley, published alongside Ilya Sutskever and Quoc V. Le a paper that now has more than 30,000 citations: "Sequence to Sequence Learning with Neural Networks." The idea: a neural network that reads a complete sequence, compresses it into a vector, and then reconstructs it as another sequence, even in another language. It sounds simple. Before that paper, nobody knew how to do it well with neural networks.

That encoder-decoder model described in the 2014 paper is the foundation of practically everything that followed: automatic translators, virtual assistants, and with certain variations, the large language models that generate text today. In 2015 he extended the idea with "pointer networks," a variant that allowed the network to point back to parts of the input, something crucial for tasks like summarizing or copying information. In 2024, the NeurIPS Test of Time Award recognized that original work as one of the most influential of the past decade in machine learning.

But if there is one project that captures Vinyals's ambition in its full magnitude, it is AlphaStar. In January 2019, the DeepMind team he led presented an AI that defeated professional StarCraft II players, one of the most complex real-time strategy games in the world, at Grandmaster level. It was not a simplified game or a controlled scenario: it was the full screen, against the best human players, in near-real conditions. The Nature paper that year documented it for history. Vinyals arrived at that project already serving as VP of Research at Google DeepMind and co-technical lead of the Gemini project, the model that today competes with the best in the world. He was also named by MIT Technology Review as one of the 35 innovators under 35 years old in 2016.

To me, the lesson from Oriol Vinyals is that the most powerful ideas in AI are not usually the most complicated ones. Sometimes it is an elegant abstraction, like "train a network to convert sequences into sequences," that unlocks years of progress. Every time you ask an AI to translate a paragraph, summarize a document, or generate a response, a little piece of that intuition is working in the background. And that, coming from someone who grew up in Barcelona and studied at Berkeley before changing everything, feels like a reminder that geography does not define what you can contribute to the world.