Koray

Koray Kavukcuoglu leads generative AI model development and integration into Google products as SVP of Google DeepMind and Chief AI Architect at Google. His earlier work includes co-authorship of the Atari deep reinforcement learning and WaveNet papers; he also founded DeepMind’s deep learning team during his prior tenure as VP of Research.[2][3][4]

Latent.Space reported on August 6, 2026 that Kavukcuoglu was stepping up from CTO to SVP and taking operational control of DeepMind, overseeing Gemini, frontier research, and product and development teams as Demis Hassabis moved from CEO to Chair and Chief Scientist. Google’s biography lists Kavukcuoglu as SVP and describes his responsibilities as developing generative models and integrating them into products at scale.[1][2]

Before that reported transition, Kavukcuoglu’s public role already combined model leadership with Google-wide AI architecture. The November 18, 2025 Gemini 3 announcement identified him as DeepMind CTO and Google Chief AI Architect.[6][1]

In their joint announcement on behalf of the Gemini team, Kavukcuoglu and Hassabis framed Gemini 3 as progress toward artificial general intelligence: “Today we’re taking another big step on the path toward AGI and releasing Gemini 3.”[6]

They announced Gemini 3 Pro in preview across Google products while staging access to Deep Think, an enhanced reasoning mode. Safety testers would receive access before Google AI Ultra subscribers, with additional safety evaluations and tester input preceding subscriber availability. This was a jointly announced team release approach.[6]

Kavukcuoglu’s earlier research provides context for his later model-development leadership. Google credits the deep learning team he founded with breakthroughs including DQN (Deep Q-Network) and WaveNet, both represented in his co-authored research.[2][3][4]

Kavukcuoglu studied at NYU’s Computational and Biological Learning Lab. Its historical alumni directory lists MSc study in 2005–2007 and PhD study in 2007–2010, followed by a research scientist position at NEC Labs in Princeton.[5]

He co-authored Playing Atari with Deep Reinforcement Learning, submitted on December 19, 2013, with Volodymyr Mnih, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra and Martin Riedmiller. Their model used a convolutional neural network trained with a variant of Q-learning: it took raw pixels as input and produced a value function estimating future rewards.[3]

The team applied the same architecture and learning algorithm to seven Atari 2600 games without adjustments between games. The authors reported outperforming previous approaches on six of those games and surpassing a human expert on three.[3]

In 2016, Kavukcuoglu co-authored WaveNet: A Generative Model for Raw Audio with Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner and Andrew Senior. First submitted on September 12 and revised on September 19, the paper introduced a probabilistic, autoregressive neural network that generated raw audio waveforms by conditioning each sample’s predictive distribution on previous samples.[4]

A single WaveNet could switch between speakers by conditioning on speaker identity. For English and Mandarin text-to-speech, the authors reported that human listeners rated its output as significantly more natural sounding than the best parametric and concatenative systems in their comparison. Google credits these breakthroughs to Kavukcuoglu’s team; the papers document collaborative authorship.[4][2]