Background
Song did his PhD at Stanford with Stefano Ermon and stayed for a postdoc, producing DDIM and a body of work on generative modelling and inference-time scaling. He was a research scientist at NVIDIA Research before joining Luma as Chief Scientist, where he was named to MIT Technology Review's Innovators Under 35.
At Luma he led the transition from Genie (3D) to Dream Machine and the Ray video line, and then to multimodal work — Uni-1, a multimodal reasoning model for image generation, and Terminal Velocity Matching and Inductive Moment Matching on the sampling side. He is publicly thoughtful about the limits of theory in data selection.
What they run now
Luma's research agenda: video and world simulation (Ray3, Ray3.2, Ray3.14), the Uni-1 multimodal image model, and sampling efficiency. His public position on data is unusually candid — he says there is no clean answer to what the right dataset is, that the IID assumptions behind statistical learning do not match real distributions, and that an empirical, objective-driven approach beats a theoretical one. That is a person who would rather test a data supplier than argue about it.
What they buy
Inference. As Chief Scientist he approves research data spend rather than running procurement. His empiricism suggests the way to sell him is a measurable ablation — show that a model post-trained on expert-rated preference data beats one trained on crowd data on Luma's own seven dimensions. Luma's 2026 RL hiring (reward infrastructure, reward-model serving, LLM-as-judge) is the concrete demand signal: judges need human labels to be trained and validated against.
Career
- c. 2022 – present or recently departedChief Scientist, Luma AI — tsong.me describes the role in the past tense as of 2026-09
- c. 2021 – c. 2022Research Scientist, NVIDIA Research
- c. 2016 – 2021Postdoc with Stefano Ermon; PhD, computer science, Stanford University — DDIM
- c. 2016–2021Stanford University, PhD, computer science
