Summary
Atli Kosson is a PhD student and researcher specializing in optimization algorithms for large-scale neural network pre-training, with eight years of industry and research experience across EPFL, Cerebras, Tesla, and Amazon. His work has produced influential papers (ICML, NeurIPS, MLSys) that reframed common practices—showing weight decay behaves like a learning-rate scheduler, explaining learning-rate warmup, and enabling multiplication-free transformer training. He blends deep theoretical insight with hands-on systems engineering, from custom CUDA modules and low-precision simulations to distributed training optimizations for thousands of GPUs. Based in San Francisco but trained at EPFL and Stanford, he focuses on making training both faster and more principled, and has helped design optimizers that remove the need for warmup while improving hardware efficiency.
8 years of coding experience
4 years of employment as a software developer
Bachelor of Science - BS Electrical and Computer Engineering, Bachelor of Science - BS Electrical and Computer Engineering at University of Iceland (Háskóli Íslands)
Doctor of Philosophy - PhD Computer Science, Doctor of Philosophy - PhD Computer Science at EPFL
Exchange Student Electrical and Computer Engineering, Exchange Student Electrical and Computer Engineering at UC Santa Barbara
Master of Science - MS Electrical Engineering, Master of Science - MS Electrical Engineering at Stanford University