Miju Labs

← All buyers · Luma AI

Jiaming Song

Chief Scientist (his own site now reads 'former Chief Scientist at Luma AI' — status uncertain) · Luma AI

medium confidenceTier B · Map and watchImage, video & audio modelsvideo · image

Song has been Luma's Chief Scientist, leading research from 3D through video to multimodal models (Ray, Uni-1), and is the author of DDIM — one of the foundational diffusion sampling methods. Important caveat: as of September 2026 his own website describes him as former Chief Scientist at Luma AI, while third-party listings still show him in post. If he has left, Luma's research leadership and its data philosophy are both in flux, which matters for anyone planning an approach.

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

  1. c. 2022 – present or recently departedChief Scientist, Luma AI — tsong.me describes the role in the past tense as of 2026-09
  2. c. 2021 – c. 2022Research Scientist, NVIDIA Research
  3. c. 2016 – 2021Postdoc with Stefano Ermon; PhD, computer science, Stanford University — DDIM
  4. c. 2016–2021Stanford University, PhD, computer science

On the record

there is no right dataset · 2025-05

Says it is genuinely hard to say what the right dataset is, and that theory does not settle it. source

multimodality · 2025-05

Argues real multimodality needs a unified latent space rather than retrofitting a language backbone to every modality. source

Interviews & talks

Why it matters here

Important but unstable. If he is still at Luma he is the person who can bless a data experiment; if he has left, that authority has moved and the Ray3 evaluation pair (Day, Motwani) plus whoever inherits research become the route. Either way this needs verifying before anyone spends effort on him.

How to reach

Verify employment first — one look at his X (@baaadas) or LinkedIn settles it. Speculation: if still in post, approach with an ablation proposal rather than a services pitch; he responds to empirical claims and has a public record of saying data questions are settled by experiment. If he has left, redirect the effort to Day, Motwani and Luma's RL hires.

What we could not establish
  • Whether he has actually left Luma AI, when, and for what — his own site uses the past tense but no announcement was found.
  • Who leads Luma research if he has gone.
  • Whether Luma has ever used an external data or rater vendor.

Why they are on the map

Chief Scientist; publicly discusses data curation and ratios; senior author on TVM (2025-11). source — evidence dated 2025-11

Needs the data for a model; writes the spec.

How Luma AI buys data

Ray3 Evaluation Report (2025-10-14, by Sabrina Day and Simran Motwani) says outputs were 'rated by trained evaluators with diverse domain expertise' and that the framework was built with 'scientists, filmmakers, artists, and creators' plus 'independent review'. The vendor is not named. 2026 hiring: RL researcher to 'build reward infrastructure… reward-model serving, LLM-as-judge', a World Model researcher to 'define and own the metrics', and a PM for Applied Research. Leadership stresses curated, quality-over-quantity data.

Founder-led (CEO Amit Jain, CTO Alex Yu) with Chief Scientist Jiaming Song over research. Visible teams: Foundation Models, Training Compute (RL), Data Infrastructure, Inference, Forward Deployed Creative, and Dream Lab LA. Evaluation reports come from a small team (Day, Motwani).

Sources

  1. Jiaming Song — personal site
  2. Luma Labs' Diffusion Revolution — Amit Jain and Jiaming Song (Cognitive Revolution, 2025-05-11)
  3. Jiaming Song — MIT Tech Review Innovators Under 35
  4. Ray3 Evaluation Report — State-of-the-Art Performance for Pro Video Generation (Luma AI, Oct 2025)