分享李飞飞对机器人最新的判断，训练机器人和训练大模型是两回事。

TL;DR - A transcript of an a16z conversation (Martin Casado with Fei-Fei Li and Yunzhu Li) on World Labs' acquisition of SceniX, arguing that robot training requires a fundamentally different recipe than LLM training because internet-scale action data doesn't exist. It matters because it lays out a concrete real-to-sim-to-real infrastructure bet for scaling robot learning and evaluation.

• World Labs' focus is "spatial intelligence" via Large World Models; its Marble foundation model takes text/single or multi-image prompts and outputs geometrically consistent 3D worlds (Gaussian Splat, Mesh). SceniX was originally a paying Marble customer before the merger.
• SceniX contributes a real-to-sim-to-real pipeline built on tensor reconstruction — capturing appearance, 3D geometry, and environment dynamics — which is compute-heavy today and would be made cheaper by Marble's sparse reconstruction and generative 3D capabilities.
• Their case against pure video-model approaches: robot policies need global consistency across space, time, viewpoints, and interactions; video predictors that make pushed objects vanish give no usable learning signal. The proposed path is a hybrid physics + data-driven data flywheel, physics-weighted early, data-weighted as real deployments accumulate.
• Two claimed core use cases are training and evaluation. Evaluation is framed as the underrated bottleneck (distinguishing a 90% vs 92% success-rate checkpoint costs enormous real-world labor hours), and simulation enables randomization over lighting, friction, geometry, and material to cover the state space. Waymo's reported billions of simulated hours is cited as precedent. The platform is positioned as model-agnostic and embodiment-agnostic infrastructure — no robot hardware — with deployment expected to follow structured → semi-structured → unstructured environments.