Oriol

Oriol Vinyals is an AI researcher who coauthored sequence-to-sequence learning research with Ilya Sutskever and Quoc V. Le, and neural-network distillation research with Geoffrey Hinton and Jeff Dean. He also contributed to DeepMind’s AlphaStar team. These projects addressed translating sequences, compressing model capabilities for deployment, and learning strategies in an imperfect-information game.[3][4][5]

On August 6, 2026, Latent.Space reported that Vinyals, Jeff Dean, Sanjay Ghemawat and Quoc Le were departing DeepMind to cofound Discovery Loop, a public benefit corporation aimed at automating machine learning, science and engineering.[1]

Vinyals’s earlier collaborations provide context for that transition: in 2014–2015, he worked on methods for learning sequence transformations and making ensembles easier to deploy.[3][4]

In Sequence to Sequence Learning with Neural Networks, first submitted in September 2014, Sutskever, Vinyals and Le used a multilayer LSTM to encode an input sequence into a fixed-dimensional vector, then another deep LSTM to decode the target sequence. A consequential optimization choice was to reverse the source words, leaving target order unchanged. The authors found that this created shorter dependencies between source and target and made optimization easier.[3]

On WMT’14 English-to-French translation, the paper reported 34.8 BLEU across the entire test set, with a penalty for out-of-vocabulary words, compared with 33.3 for a phrase-based statistical machine translation system. A separate experiment used the LSTM to rerank that system’s 1,000 hypotheses and reached 36.5 BLEU.[3]

Vinyals, Hinton and Dean coauthored Distilling the Knowledge in a Neural Network, identified with the NIPS 2014 Deep Learning Workshop and submitted to arXiv in March 2015. Building on ensemble-compression work by Caruana and collaborators, they developed a different technique for transferring an ensemble’s knowledge into a single model. They reported improving a commercial acoustic model and introduced ensembles combining full models with specialists trained to distinguish fine-grained classes that the full models confused.[4]

In a February 2026 interview, Dean recalled the practical motivation: an ensemble of perhaps 50 specialist image models was cumbersome to serve. Distillation compressed that capability into a deployable form. He explained that supervision from a larger model’s logits can guide a smaller model toward behavior it would not learn from hard labels alone.[2]

Vinyals was a member of the AlphaStar team announced in January 2019, alongside Igor Babuschkin, Junyoung Chung, Michael Mathieu, Max Jaderberg, Wojtek Czarnecki and other collaborators. StarCraft II required agents to scout for hidden information and balance immediate actions with long-term planning. The team first trained AlphaStar by imitation from human replays, then seeded a reinforcement-learning league. New competitors branched from existing agents and learned by playing others, expanding the strategies encountered during training.[5]

The team reported a 5–0 victory over professional player MaNa in Protoss-versus-Protoss games on CatalystLE. That demonstration used a raw interface exposing visible-unit attributes without requiring camera movement. MaNa subsequently defeated a camera-interface prototype trained for seven days in an exhibition match.[5]

In a December 27, 2025 interview with Cristina Sáez, presented in ARA’s explicitly unverified English translation, Vinyals described choosing unfamiliar research areas: deep learning at Google Brain, reinforcement learning after moving to DeepMind in London, and leadership work at Gemini.[6]

In that interview, Vinyals said he used an AI assistant with access to shared team documents to flag promising ideas and suggest conversations with colleagues. He also described it reading scientific studies, summarizing innovative ideas and suggesting research directions. He argued that students still needed subject knowledge to question and correct AI outputs, and said he did not feel the models he interacted with were conscious at that time.[6]