Koray Kavukcuoglu

Overview

Koray Kavukcuoglu is a computer scientist at DeepMind whose work runs from reinforcement learning research to the stewardship of Google's largest model releases. He co-authored "Playing Atari with Deep Reinforcement Learning", posted to arXiv on 19 December 2013, the first deep Q-network paper, together with Volodymyr Mnih, David Silver and other co-authors ^[1]. In the years that followed he moved into leadership roles at the organisation, becoming chief technology officer of Google DeepMind and chief AI architect of Google, and in that capacity announced and signed successive generations of the Gemini model family ^[2][3].

The deep Q-network paper

Reinforcement learning is the part of machine learning in which an agent learns by acting in an environment and receiving rewards, rather than by being shown labelled examples. Q-learning is one classical way of doing this: the agent tries to estimate how good each action is in each situation. A deep Q-network replaces the table of such estimates with a deep neural network, so that the agent can work from raw, high-dimensional input. The paper Kavukcuoglu co-authored is the one that introduced the deep Q-network name, and it appeared on arXiv in December 2013 with Volodymyr Mnih, David Silver and others on the author list ^[1]. It is the piece of research for which he is most often cited as a co-author.

Founding the deep learning team at DeepMind

Google's own blog credits Kavukcuoglu with founding the deep learning team at DeepMind, and points to DQN and WaveNet among the work that team has produced ^[7]. Both are research lines that came out of the same group: DQN from the reinforcement learning side, WaveNet from work on generative models of audio. That a single team is associated with both gives a sense of how the group's output was organised, and the blog treats the founding of that team as one of his contributions ^[7].

The Gemini announcements

Kavukcuoglu's public role in recent years has been closely tied to model launches. On 25 March 2025, as chief technology officer of Google DeepMind, he announced Gemini 2.5 ^[2]. On 18 November 2025 he co-signed the Gemini 3 announcement with Sundar Pichai and Demis Hassabis, by then holding the titles of chief technology officer of Google DeepMind and chief AI architect of Google ^[3]. In May 2026 he was the lead signatory of the Gemini 3.5 announcement, which was co-signed by Jeff Dean, Oriol Vinyals and Noam Shazeer ^[4]. By that autumn he was writing the launch posts himself as senior vice president of Google DeepMind and chief AI architect of Google: Gemini 4 Argon on 13 October 2026 and Gemini 4 Carbon on 28 October 2026 ^[5][6].

Responsibilities

In a message from Google's chief executive, Kavukcuoglu — then chief technology officer of Google DeepMind and chief AI architect — was named to step up as senior vice president of Google DeepMind, reporting to the chief executive and overseeing Gemini model development, frontier AI research, and the Gemini app and developer teams ^[8]. The remit combines research, the core models and the products built on them, which is roughly the span covered by the announcements listed above.

What the record shows

Read as a sequence, the announcements trace a shift in role: from a co-author on a 2013 research paper ^[1], to a named signatory on the Gemini 3 announcement in November 2025 ^[3], to lead signatory on Gemini 3.5 in May 2026 ^[4], to the author of the launch posts for the Gemini 4 models in October 2026 ^[5][6]. The same period saw his title move from chief technology officer of Google DeepMind to senior vice president of Google DeepMind, while he retained the chief AI architect title at Google ^[3][5][8]. Alongside that public-facing work stands the earlier institutional contribution Google credits him with: founding the deep learning team whose research includes DQN and WaveNet ^[7]. For readers following how frontier models are presented to the public, he is one of the recurring names on Google's announcements; for readers following reinforcement learning, he is one of the names on the paper that introduced the deep Q-network ^{[1][2][3][4][5][6]}.