Vincent Hellendoorn
Research Scientist at Google DeepMind
United States
Summary
🤩
Rockstar🎓
Top SchoolVincent Hellendoorn is a research scientist at Google DeepMind and former assistant professor at CMU with 11 years of experience at the intersection of AI for code, reinforcement learning, and thinking-efficiency research. He brings a strong academic background (PhD UC Davis, MSc/TU Delft) and a track record of applied ML engineering, including contributions to high-profile open-source work like GPT-NeoX where he optimized attention mechanisms with Flash Attention and dynamic Alibi caching. At DeepMind he works on Gemini and projects that push both model performance and developer productivity, while his prior academic roles blended rigorous research with practical tooling for software engineering. Known for moving seamlessly between theory and production, he combines deep algorithmic insight with hands-on systems improvements that scale.
11 years of coding experience
1 year of employment as a software developer
PhD, Computer Science, PhD, Computer Science at University of California, DavisVWO, High School degree, VWO, High School degree at College het LooBachelor's Degree, Computerscience, 8.2, Bachelor's Degree, Computerscience, 8.2 at TU Delft
Dutch, English, German, French
Github Skills (10)
transformers10
flash10
pytorch10
multihead-attention10
gpt10
language-model10
self-attention10
python10
parallel-computing9
cuda9
Programming languages (8)
Github contributions (5)
EleutherAI/gpt-neox
Oct 2021 - Dec 2022
An implementation of model parallel autoregressive transformers on GPUs, based on the Megatron and DeepSpeed libraries
Role in this project:
ML Engineer
ML EngineerContributions:1 review, 6 commits, 2 PRs in 1 year 2 months
Contributions summary:Vincent contributed to the implementation and improvement of a model parallel autoregressive transformer, specifically within the context of GPT-NeoX. Their work included integrating Flash Attention, a technique for optimizing attention mechanisms. They also made adjustments to Alibi matrix caching for dynamic sequence lengths and refined other attention-related configurations. Further contributions include the addition of maximum tokens to the interactive mode.
pytorchlanguage-modeltransformersdeepspeedmodel-parallel
DeepTyper/DeepTyper
Nov 2018 - Feb 2020
Contributions:2 commits, 3 PRs, 3 pushes in 1 year 3 months
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.