Miju Labs

← All buyers · Meta (Meta Superintelligence Labs)

Jiahui Yu

Former Multimodal Lead, MSL (left Aug 2026 to found a startup) · Meta (Meta Superintelligence Labs)left

medium confidenceTier B · Map and watchFrontier labsimage · video · multimodal

Led multimodal at Meta's TBD Lab and shipped Muse Spark, Muse Image, Muse Video and Voice Mode — then announced on 14 August 2026 that he was leaving Meta to start a new company. Previously OpenAI's perception lead and, before that, a Google Brain researcher on PaLI and Parti. He is no longer a Meta buyer; he is a future one, once his company is named.

Background

Yu did his PhD at the University of Illinois Urbana-Champaign on image generation and inpainting, then worked at Google Brain on large vision-language and text-to-image models (PaLI, Parti, CoCa). He moved to OpenAI as perception lead, contributing to the multimodal capabilities of GPT-4o and later models, and was recruited to Meta in June 2025 during the Superintelligence Labs hiring wave.

At Meta he built and led the multimodal direction inside TBD Lab alongside Zuckerberg and Alexandr Wang. His team's shipped output — Muse Spark (April 2026, the first model under Alexandr Wang, powering Meta AI), Muse Image, Muse Video, Voice Mode — is Meta's visual-generation line. Muse Spark 1.2 shipped eight days before he announced his departure. He said he was drawn to a problem that would shape humanity's future and remains largely unexplored, without naming it.

What they run now

Until August 2026: Meta's multimodal model line. Now: an unnamed new company. Nothing public about its domain.

What they buy

Nothing, currently. As Meta's multimodal lead he would have specified image and video preference data, aesthetic tuning and instruction-following judgements for the Muse line. If his new company is in generative media he becomes a founder-stage buyer who moves fast and has capital — a different and often better customer than a lab procurement process.

Career

  1. 2023Research scientist, Google Brain — PaLI, Parti, CoCa
  2. 2023 – 2025Perception lead, OpenAI — Multimodal capabilities incl. GPT-4o
  3. 2025-06 – 2026-08Multimodal lead, Meta Superintelligence Labs (TBD Lab) — Muse Spark, Muse Image, Muse Video, Voice Mode
  4. 2026-08 – presentFounder, Undisclosed startup — Company and domain not announced
  5. —University of Illinois Urbana-Champaign, PhD, Computer Science
  6. —University of Science and Technology of China, Bachelor's

On the record

Leaving Meta · 2026-08

He said he is leaving to work on a problem that will profoundly shape humanity's future yet remains largely unexplored. source

Interviews & talks

Why it matters here

Low today, potentially high in six months. The immediate value is corrective: he should be marked as departed in the index so Julian does not pitch a Meta role that no longer exists. Set a watch on the company announcement; a well-funded ex-OpenAI, ex-Meta multimodal lead founding a company is exactly the kind of early buyer who needs taste data and has nobody internal to produce it.

How to reach

Monitor. Speculation: when the company is announced, approach immediately and directly — founders at that stage answer their own messages, and the window before they build an internal data function is short. His X account and personal site are the channels.

What we could not establish
  • What his new company does.
  • Whether Meta replaced him and with whom (Huiwen Chang is the most senior remaining image person identified).
  • Departure details come from a secondary summary of his X post; the post itself was not readable in this pass.

Why they are on the map

Led development of Muse Image and Muse Video (Jul 2026); ex-OpenAI Perception lead. source — evidence dated 2026-07

Needs the data for a model; writes the spec.

How Meta (Meta Superintelligence Labs) buys data

Aug 2025: TBD Lab researchers reportedly preferred Surge and Mercor data over Scale's, despite Meta's $14.3B Scale stake. Meta has also partnered with Mercor on RL environments (TechCrunch, Sep 2025). In Aug 2025 Meta licensed Midjourney's 'aesthetic technology' to make its models 'more visually appealing', which is a direct signal of paying for visual taste. Muse Image/Video (Jul 2026) are marketed on human-preference Elo from internal ablations and Arena. SemiAnalysis (Jul 2026) reports about 3,000 engineers working on RL task/environment creation.

Alexandr Wang (Chief AI Officer, org context only) heads MSL. Since Aug 2025 it has four parts: TBD Lab (frontier models), FAIR (Rob Fergus), Products and Applied Research (Nat Friedman) and MSL Infra (Aparna Ramani, Amir Frenkel). Shengjia Zhao is Chief Scientist and owns the Muse roadmap; Jiahui Yu leads multimodal, including Muse Image/Video. No human-data owner is public; data-vendor decisions appear to sit with TBD Lab researchers and ex-Scale staff.

Known vendors: scale-ai, surge-ai, mercor, lmarena, Midjourney (aesthetics tech licence)

Sources

  1. Meta core AI researcher Jiahui Yu departs to launch startup
  2. Jiahui Yu departure post
  3. Jiahui Yu — personal site
  4. Former OpenAI Perception Lead Jiahui Yu Joins Meta's AGI Team