Regularized Newton Method with Global O(1/k^2) Convergence (One World Optimization Seminar) · One World Optimization Seminar / YouTube
Known for
Prodigy and Schedule-Free optimizers → Mistral, Oct 2026.
Why chosen
He connects optimization theory with the brute work of training frontier models. After a PhD at KAUST and a stint at Samsung AI Center, he co-authored D-Adaptation and Prodigy, optimizers that cut out manual learning-rate tuning. In 2024 his Schedule-Free AdamW submission won the self-tuning track of the first MLCommons AlgoPerf competition, and that October he joined Meta as a research scientist. In October 2026 he joined Mistral AI's pretraining team, saying model quality now comes before papers. Less tuning, more training.
Regularized Newton Method with Global O(1/k^2) Convergence (One World Optimization Seminar) · One World Optimization Seminar / YouTube
Role
AI Researcher, pretraining, Mistral AI
Status
JoinedLeft Meta in August 2026; joined Mistral AI's pretraining team in October 2026.