VP of Research at Google DeepMind, Gemini co-technical lead
Oriol Vinyals
Profile
Oriol Vinyals is VP of Research at Google DeepMind and co-technical lead for Gemini, Google’s flagship foundation model line. If you use Google Translate, Google Search, or anything in the Gemini family, you’re touching the downstream effects of his work. His Google Scholar page shows over 335,000 citations — the kind of number that only lands when your papers become load-bearing infrastructure for an entire field.
Before he was running Gemini, Vinyals was quietly co-authoring some of the papers that made modern LLMs possible. In 2014, with Ilya Sutskever and Quoc Le, he published Sequence to Sequence Learning with Neural Networks — the seq2seq paper that showed neural networks could map input sequences to output sequences end-to-end. That unlocked modern neural machine translation and is a direct ancestor of every encoder-decoder transformer we use today. The same year, with Samy Bengio and others, he co-authored Show and Tell, one of the first neural image captioning systems. He also co-invented Pointer Networks, which showed up years later inside copy-attention mechanisms across NLP.
After moving to DeepMind from Google Brain, he led the AlphaStar project — the first agent to reach Grandmaster level in StarCraft II, published in Nature in 2019. AlphaStar is worth studying not because you’ll ever need to play StarCraft, but because the playbook (large-scale self-play, league training, imitation learning from human replays) keeps resurfacing in modern RL post-training for LLMs. He worked alongside David Silver and Demis Hassabis during DeepMind’s run of game-playing breakthroughs.
Now his day job is pushing Gemini forward. In late 2025 he publicly pushed back on the “pre-training is done” narrative, claiming that Gemini 3’s gains came from substantial pre-training and post-training improvements — a useful counterpoint if you’re trying to calibrate how much scaling headroom actually remains. For developers learning AI today, Vinyals is a good person to follow precisely because he’s a practitioner-researcher: he writes papers and ships models, and he’s candid about what’s working.
Key Articles & Papers
Sequence to Sequence Learning with Neural Networks Show and Tell: A Neural Image Caption Generator Grammar as a Foreign Language Pointer Networks A Neural Conversational Model Matching Networks for One Shot Learning StarCraft II: A New Challenge for Reinforcement Learning Grandmaster level in StarCraft II using multi-agent reinforcement learning Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic CapabilitiesVideos
YouTube
Spotify Podcasts
