--- title: "The ResNeXt and DiT author who is on leave from NYU to build LeCun's world-model lab" date: 2026-11-07 category: Builders site: NeuroAI canonical: https://neuroai.site/a/na-people-saining-xie-ami language: en --- # The ResNeXt and DiT author who is on leave from NYU to build LeCun's world-model lab > Saining Xie (谢赛宁), the Chinese computer scientist behind ResNeXt and the Diffusion Transformer that underpins Sora, took leave from NYU in 2026 to become chief scientific officer of AMI Labs with Yann LeCun. A researcher who helped design the architecture behind OpenAI's video model Sora is on leave from one of the world's top universities to help build something bigger: a "world model" lab run by his old mentor. Saining Xie (谢赛宁) is not a household name, but his papers sit inside almost every modern vision system. ## Key takeaways - Saining Xie (谢赛宁) earned a computer-science bachelor's from Shanghai Jiao Tong University (2013) and a PhD from UC San Diego (2018) under Zhuowen Tu. - He was a research scientist at Meta AI (FAIR) from 2018 to 2022, then joined New York University's Courant Institute as an assistant professor in September 2022. - As first author he proposed ResNeXt (with He Kaiming and others) and, with William Peebles, the Diffusion Transformer (DiT) — a foundation for OpenAI's Sora video model. - In 2026 he co-founded AMI Labs (Advanced Machine Intelligence Labs) with Yann LeCun, serving as chief scientific officer while on leave from NYU; the lab launched in Paris in January 2026. - His recent work, Cambrian-S (ICLR 2026), argues that scaling alone has hit a wall on spatial perception — a thesis that shapes AMI's direction. ## The papers you have already used If you have trained a vision model in the last seven years, Xie's name is probably in your code. As a PhD student and FAIR researcher he co-authored ResNeXt, which added a new "cardinality" dimension to He Kaiming's residual blocks; MoCo and MAE, the self-supervised pre-training landmarks, carried his byline too. In 2022 he and William Peebles published the Diffusion Transformer (DiT), a design that paired Transformers with diffusion models — the architecture OpenAI later built Sora on. Xie has been careful to note he was not a Sora inventor; his role was the underlying method. That body of work made him, at 30-something, a widely cited young vision researcher. ## From NYU back to LeCun Xie's career has circled Yann LeCun. After UCSD he joined FAIR, where LeCun was chief AI scientist; in 2022 he returned to NYU — LeCun's academic base — as a professor. In 2024 he co-released Cambrian-1, an open multimodal model, with LeCun's team. The circle closed in 2026. In December 2025 LeCun announced Advanced Machine Intelligence Labs (AMI Labs), and the lab formally launched in Paris in January 2026 with offices in New York, Montreal and Singapore. LeCun is executive chairman; Xie is chief scientific officer, on temporary leave from NYU. The founding team is heavy with ex-Meta talent, and the mission is a "world model" — an AI that predicts and understands physical reality, not just language. ## Why his latest paper matters Xie's 2026 paper, Cambrian-S (accepted at ICLR 2026), is more than an academic release; it is a research manifesto. With collaborators including LeCun and Fei-Fei Li, he argues that current multimodal systems only reach the first stage of "spatial hyper-perception" — parsing pixels into objects — and collapse on longer, continuous video. In one benchmark, model accuracy fell off sharply as the video lengthened. His conclusion: scaling data and size alone is nearing its limit for spatial understanding, and the next gain must come from predictive perception of the 3D world. That thesis is exactly what AMI is betting on. Xie's move from tenure-track professor to chief scientific officer signals how seriously the field treats the world-model race — and how willing top researchers are to leave academia for it. ## What readers can do now - **If you build with vision models, read DiT and ConvNeXt directly.** Xie's architectures are free, citation-rich starting points, not black boxes. - **Watch AMI Labs as a bellwether for the "world model" bet.** Its hires and papers will show whether predictive perception outperforms scaling. - **For students, note the loop:** Xie went industry (FAIR) → academia (NYU) → industry (AMI) — and each step published openly. Open research is still the fastest career path. ## Honest limitations This profile uses Baidu Baike (English), a long-form Chinese tech profile, and Jiemian News reporting on AMI Labs; I have not read AMI's founding documents or NYU's leave records. The biographical facts — SJTU bachelor's (2013), UCSD PhD (2018), FAIR (2018–2022), NYU from September 2022 — are consistent across Baidu and the tech profile. AMI Labs' December 2025 announcement, January 2026 Paris launch, and four offices are from Jiemian News; I have not confirmed the exact funding amount and therefore do not state one. The DiT–Sora linkage is Xie's documented contribution, with his own public clarification that he is not a Sora inventor. The Cambrian-S benchmark trend comes from the cited profile of the paper and is presented as reported. I have not compared Xie's citation totals across databases, which disagree. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-people-saining-xie-ami Free to quote with attribution and a link to the original.