Oriol Vinyals

Overview

Oriol Vinyals is a Spanish computer scientist, born in Sabadell, Catalonia, in 1983.^[1] His public career runs through Google Brain, DeepMind and Google DeepMind, and covers neural sequence models, the AlphaStar research group, and the Gemini model family.^[1]

Education and early career

He received his PhD from the University of California, Berkeley, in 2013; his advisor was Nelson Morgan.^[1] After Berkeley he joined Google Brain, and in 2016 he moved to DeepMind in London.^[1] He has spent most of his career at DeepMind.^[1]

Sequence-to-sequence learning

In 2014 he co-invented sequence-to-sequence learning with Ilya Sutskever and Quoc Le, published as "Sequence to Sequence Learning with Neural Networks".^[1][2] The name describes the idea fairly directly: train a neural network on pairs of sequences so that it learns to turn one sequence into another. A translation system built this way reads a sentence in one language and produces a sentence in another, rather than relying on a separate hand-built stage for each part of the job. The same framing can be applied to any task that can be written as one sequence in, one sequence out, which is why the formulation travels well beyond any single application.

DeepMind and AlphaStar

At DeepMind he led the AlphaStar research group.^[1] The group is one of the research efforts he is associated with during his years at the company, alongside his work on sequence models.

Emergent abilities of large language models

On 15 June 2022 a paper titled "Emergent Abilities of Large Language Models" was submitted to arXiv, with Vinyals as a co-author together with Jason Wei, Jeff Dean and others.^[4] The subject is a phenomenon familiar to anyone who watches models grow: some abilities are simply absent in small language models and appear once a model passes a certain size, instead of improving gradually along the way. The paper puts a name to that pattern and treats it as something worth measuring and arguing about, which matters because it makes the relationship between scale and capability hard to predict in advance.

Gemini

As of August 2024 he was a technical lead on Gemini, along with Noam Shazeer and Jeff Dean.^[1] In May 2026 he co-signed the Gemini 3.5 announcement as a vice president at Google DeepMind.^[3] On 5 August 2026 he announced that he was leaving Gemini to start a company, Discovery Loop.^[1]

Place in the field

Vinyals's record is a fairly compact illustration of how the field moved over roughly a decade. Sequence-to-sequence learning gave a general way to state a learning problem: pair up inputs and outputs and let a network learn the mapping.^[2] Work on emergent abilities asked what happens when models are scaled up, and what can and cannot be expected from that scaling.^[4] His later role sat inside Gemini, a large model family developed at Google DeepMind, where he shared technical leadership with two other researchers.^[1] The sequence of moves — from a general-purpose learning recipe, to questions about scale, to running a frontier model project, and finally to founding a company of his own — traces a path that many senior researchers in the field have followed in the same period.^[1]