AI离“理解万物”还有多远？先拿癌细胞和行星轨道试试水
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR — JEPA-Anything 是一种跨领域潜在世界模型：通过领域专用编码器与共享预测核心，在七类系统中学习可迁移、可解释的动态规律。它不仅提升了多数基准任务的预测精度，还从潜在表示中发现了经肝癌模型验证的干预方案，并近乎精确地复现开普勒第三定律。
- 核心方法“正交预测因子分解”将目标状态拆分到互补且受约束的子空间中，减少不同尺度信号之间的相互干扰。
- 在数据、架构和算力一致的条件下，该方法在 10 项预测任务中的 9 项取得提升；其中两项 PDEBench 任务的 MSE 约降低 40%，WeatherBench 2 的误差降低 10.5%。
- 在四种分子系统上，它取得了已报告的最低单步及 100 步滚动预测误差，但长时间跨度下优势有所收窄，在控制任务中的表现也并不一致。
- 研究者利用学到的潜在因子提出“IL-18 联合 CD73 阻断”的候选癌症干预方案，并在肝癌模型中完成验证。
- 模型学习到的轨道模式给出 −1.4991 的拟合斜率，几乎精确复现开普勒第三定律所预测的 −1.5 标度关系。
注：第一组来源聚焦 JEPA-Anything；第二组关于华为昇腾芯片、超节点与 CANN 的内容属于另一项工作，与本标题所述研究无直接关联，因此未合并。
Sources (2)
AI离“理解万物”还有多远？先拿癌细胞和行星轨道试试水
TL;DR - JEPA-Anything is a cross-domain latent world-model framework that shares one predictive core across seven system types while using domain-specific encoders. Its orthogonal factorization improves prediction on most benchmarks and yields interpretable latent structures that supported a validated cancer intervention and recovered Kepler’s third law.
- Orthogonal Predictive Factorization splits target states into complementary, constrained subspaces, reducing interference between signals at different scales.
- Under matched data, architecture, and compute, it improved 9 of 10 prediction tasks, including roughly 40% lower MSE on two PDEBench tasks and 10.5% lower error on WeatherBench 2.
- It achieved the lowest reported one-step and 100-step rollout errors across four molecular systems, though gains narrowed over longer horizons and were inconsistent across control tasks.
- Researchers used learned factors to identify an IL-18 plus CD73-blockade candidate validated in liver-cancer models, while latent orbital modes reproduced the theoretical −1.5 Kepler scaling with a fitted slope of −1.4991.
华为汪涛：华为要打造AI算力底座，只做好一颗芯片远远不够
TL;DR - Huawei outlined a system-level AI infrastructure roadmap centered on annual Ascend chip upgrades, large-scale supernodes, optical interconnects, and the CANN software ecosystem. The strategy aims to offset manufacturing constraints by improving aggregate computing efficiency and reliability rather than relying on chip performance alone.
- Ascend 960DT is slated to deliver 2 PFLOPS FP8 or 4 PFLOPS FP4, with up to 288GB HBM and 9.6TB/s bandwidth; Huawei plans annual generations through Ascend 980 in 2029.
- The Ascend 960 supernode connects 4,096 NPUs with unified memory addressing, offering up to 8 EFLOPS FP8 and more than 1PB of HBM.
- Huawei says its NPO-based Hi-ONE optical engines replace roughly 48,000 conventional 800G modules, cutting power by over 550kW while targeting 99.8% system availability.
- CANN now reports over 5,200 monthly active developers and is entering regular open-source operation, while Ascend is moving onto PyTorch’s official support roadmap.