Brendan O’Donoghue is a director of research at Google DeepMind who leads its text-diffusion research team, including work on Gemini Diffusion and DiffusionGemma. His research spans convex optimization, reinforcement learning and language models that generate text through parallel, iterative refinement.
O’Donoghue studied computer and information engineering at the University of Cambridge and completed a Stanford electrical engineering doctorate in 2013 under Stephen Boyd. His research explored how convex optimization could produce practical control policies when exact solutions were computationally prohibitive.
He created and maintains SCS, the Splitting Conic Solver, an open-source system that uses first-order operator splitting to solve large-scale convex optimization problems or certify infeasibility. Related work on large-scale linear programming received the 2024 Beale–Orchard-Hays Prize.
Before joining DeepMind in 2015, O’Donoghue led a machine-learning team at Quantcast working on advertising targeting and real-time bidding. At DeepMind, he investigated how reinforcement-learning agents should explore when the consequences of their choices remain uncertain. His Uncertainty Bellman Equation extends value estimation to uncertainty, while K-learning combines expected rewards and epistemic uncertainty into knowledge values with formal regret guarantees. With Ian Osband and Cătălin Ionescu, he subsequently showed why probabilistic approaches must account for how an agent’s actions shape its future observations.
Language models that revise their answers
Gemini Diffusion generates text by repeatedly denoising blocks of tokens instead of committing to each token sequentially. Its successor, DiffusionGemma, is an Apache 2.0 open-weight model adapted from the Gemma 4 family through supervised fine-tuning, sampler distillation and reinforcement learning.
O’Donoghue’s approach emphasizes four consequential properties:
- Bidirectional reasoning and self-correction: Tokens within a generation block can attend to one another, allowing later reasoning to correct an earlier answer before the response is finalized. The same mechanism supports in-place code editing and context-aware infilling.
- Adaptive computation: Easier requests can finish after fewer denoising steps, while difficult programming or scientific-reasoning problems receive additional computation.
- Latency–throughput tradeoffs: Refining multiple tokens together reduces repeated transfers across constrained accelerator-memory bandwidth, improving individual response times. Additional forward passes can nevertheless lower throughput and increase serving costs for heavily batched deployments.
- Interactive software: Fast generation enables responsive code editing, voice-driven programming and interfaces generated dynamically as users interact with them. Local devices and robotics are especially relevant because they need low latency without serving thousands of simultaneous requests.
His AI Engineer Europe presentation connects these capabilities to a precise engineering constraint: the value of text diffusion depends on whether faster responses outweigh the computational cost of producing them.
