Daniel Y. Fu

Worth following because he brings a real theory background into the model-systems layer, especially where structured linear algebra and sequence methods end up mattering for practical modern architectures.

Important because he sits at a productive seam between machine learning, data systems, and model infrastructure, with work that ranges from weak supervision to some of the most important efficiency breakthroughs in modern training stacks.

One of the clearest researchers to follow for efficient sequence-model systems, especially the line of work that made frontier training and inference materially faster rather than merely cleaner on paper.

A high-signal figure for understanding how DeepMind turned ambitious research systems into durable products, especially across reinforcement learning, speech, and code generation.

Foundational less for any single public paper than for shaping the infrastructure, engineering culture, and systems thinking that make frontier-model research possible.

A strong person to follow if you care about open-weight language models and retrieval-heavy NLP systems, especially the line from RoBERTa and RAG into LLaMA-era model development.