Light Origins tweet 2080861398712946779 (content not retrievable at filing time)

TL;DR

A tweet from @LightOrigins_ (亮源新创 / Light Origins), the Singapore/Shenzhen/Beijing-based embodied-intelligence company founded in 2024 by Roger Jiang — co-inventor of InstructGPT and RLHF and a GPT-4 contributor (cited 8× in the OpenAI contributor list). The specific tweet body was not retrievable at filing time (X login wall, no embedded URLs, no Nitter cache), so this page records the account context rather than a specific claim: Light Origins’ stated thesis is that “embodied intelligence needs its own pretraining” — compressing human behavior into Universal Action Representations that transfer across bipeds, quadrupeds, wheeled robots, and digital humans, so that interaction, adaptation, and physical intuition can migrate across environments and embodiments and continue to evolve during deployment.

Key claims

• Company mission per lightorigins.com: unify navigation, locomotion, and manipulation so embodied systems accumulate cross-space, cross-object, cross-environment, cross-time interaction priors and become autonomous agents that infer intent and find new ways to participate in the world [lightorigins.com landing].
• Team backgrounds span AI, robotics, consumer hardware, and software, drawn from DJI, ByteDance, Tencent, Huawei, Roblox, and Oxford/Cambridge/Tsinghua/Peking/ETH [lightorigins.com — Team].
• Founded by Roger Jiang, a co-inventor of ChatGPT and InstructGPT, credited across GPT-4 foundational RLHF, data, and contamination-investigation work per OpenAI’s public contributor list [SCMP 2024-12; 36Kr 2024-12].
• Offices in Singapore, Shenzhen, and Beijing [lightorigins.com landing]; a 2026-05-22 Tencent News profile summarizes Jiang’s technical bet as bringing LLM-style scaling laws to embodied pretraining [news.qq.com 2026-05-22].

Method

Not applicable — this is a social-media pointer to a company account. No paper or method is attached to the tweet. The content not retrievable note in the title is intentional: a curator with a logged-in X session can revisit the tweet URL to recover the specific claim being made in this thread.

Results

Not applicable.

Why it’s interesting

Even without the specific tweet body, the account itself is a signal Luma should track: Light Origins joins a small set of ChatGPT-era foundation-model alumni betting that the pretraining recipe — not just the model architecture — is the missing piece in robotics. That framing puts them adjacent to the “scale embodiment-free UMI data” thesis in Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training (100K hours of Universal Manipulation Interface pretraining) and to the human-video-scaling thesis in EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data (20K+ hours of egocentric human data). Their “compress human behavior into Universal Action Representations that transfer across embodiments” phrasing is the same problem Cross-Embodiment Robot Manipulation via a Unified Hand Action Space (UHAS) is attacking from the action-space side, and the same one Human-to-Robot Retargeting catalogues as an open interface-design question.

See also

• VLA Models — Light Origins’ stated goal (unify navigation, locomotion, manipulation into one policy) is the modal VLA framing.
• Human-to-Robot Retargeting — Universal Action Representations are exactly this problem: a shared behavior manifold that transfers across morphologies.
• World Foundation Models — Light Origins positions itself as building the pretraining substrate for embodied agents, in the same slot as world-foundation-model efforts.
• Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training — closest filed analog for “scale embodiment-free interaction data at LLM-pretraining volume”.
• Cross-Embodiment Robot Manipulation via a Unified Hand Action Space — UHAS is a concrete attempt at a Universal Action Space for dexterous manipulation.