这段语境勾勒出一条定量时间线：1月发布、1943年起源、1990年代回潮、2022年JEPA研发、2025年11月创业转向、以及2023年可解释性证据，显示出从“数十年一跳”到“逐年推进”的加速。它也将该领域归纳为三种主要技术路径，其中Project Genie代表视频优先的世界模型路线。

关键性能数字凸显了当前约束与规模：Genie的会话最多只能稳定60秒，而实用机器人学习可能需要数十亿小时训练数据，仿真比真实世界采集更便宜地提供这类数据。统计趋势体现为数据体量与保真度的权衡，即2D/视频模型在生成规模上更强，但会丢失隐藏状态与跨场景持续性。

对比信号表明方向正从短时预测转向更广泛的一致性：完整生成的3D世界追求可供多用户共享的全局持久性，而JEPA试图在不逐秒模拟的情况下建模长期或然情境。另一种估计认为LLM已通过将互联网规模信息压缩到数百GB而内含世界结构，这一观点受到2023年Othello与Anthropic神经元簇发现支持，但仍被批评为缺乏落地锚定的语言推断。

The context traces a quantitative timeline: a January release, a 1943 origin, a 1990s revival, JEPA development since 2022, a startup pivot in November 2025, and 2023 interpretability evidence, showing acceleration from decade-scale intervals to year-by-year milestones. It also organizes the field into three main technical approaches, with Project Genie as the video-first world-model exemplar.

Key performance numbers highlight current constraints and scale: Genie runs remain coherent for a maximum of 60 seconds, while practical robot learning may need billions of training hours that simulation can provide more cheaply than real-world collection. The statistical trend is a tradeoff between data volume and fidelity, where 2D/video models scale generation but lose hidden-state information and cross-scene persistence.

Comparative signals show movement from short-horizon prediction toward broader consistency: fully generated 3D worlds target complete multi-user persistence, and JEPA targets long-range contingencies without second-by-second simulation. A competing estimate argues LLMs already encode world structure by compressing internet-scale information into a few hundred gigabytes, supported by 2023 Othello and Anthropic neuron-cluster findings but challenged as ungrounded language-only inference.

Source: AI tools are being prepared for the physical world

Subtitle: The race to build world models is o

Dateline: 2月 26, 2026 08:24 上午