Yu Wang
Member Of Technical Staff at xAI
Mountain View, California, United States
Summary
🤩
Rockstar🎓
Top SchoolYu Wang is a research-focused member of technical staff specializing in memory systems for Large Language Models and LLM-based agents, with 12 years of experience bridging cutting-edge research and applied engineering. Currently at xAI and with prior roles at Amazon and IBM, Yu built MemoryLLM, long-term latent memory pools, and Sleep-Time Compute techniques that materially boost retention without extra GPU cost. Their work spans self-updating models, agent lifespan cognition, multimodal agent memory, and multi-agent architectures, and is grounded in production-oriented contributions to high-profile open-source tooling like Facebook Research’s fairseq. Notably, Yu’s internships produced a reinforcement-learning framework for learned memory policies and a multimodal agent memory that improved performance by 244% over text-only baselines. Trained at UC San Diego and USTC, they combine deep academic rigor with measurable engineering impact in scalable memory augmentation for LLMs.
12 years of coding experience
2 years of employment as a software developer
University of California, San DiegoBachelor's degree Big Data Science and Technology, Bachelor's degree Big Data Science and Technology at University of Science and Technology of China
Github Skills (13)
pytorch10
machine-learning10
eval10
artificial-intelligence10
nlp10
trainings10
python10
evaluation10
natural-language-processing10
modeling10
data-augmentation9
quantization8
jit8
Programming languages (8)
Github contributions (5)
facebookresearch/fairseq
Dec 2019 - Nov 2021
Facebook AI Research Sequence-to-Sequence Toolkit written in Python.
Role in this project:
ML Engineer
ML EngineerContributions:1 review, 25 commits, 2 PRs in 1 year 11 months
Contributions summary:Yu primarily contributes to the development and improvement of the fairseq toolkit, focusing on machine learning tasks within the realm of sequence-to-sequence models. Their work includes adding features for evaluating model performance, such as saving predictions for MAP and MAUC calculations. The user also implemented teacher-student learning for TALNet, incorporating both static and dynamic teacher models. Furthermore, they made updates to enable conversion of the model to JIT format and quantization.
aipythonsequence-to-sequencepytorchartificial-intelligence
MaigoAkisame/maigoakisame.github.io
Jan 2019 - Sep 2021
Contributions:15 commits, 15 pushes in 2 years 8 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.