Musk steps aside as Tesla and SpaceX weigh a merger

Recent work collectively establishes that video generators fail measurable physical laws rather than just look implausible: simulator-grounded benchmarks from Shanghai Jiao Tong University, Zhejiang University, and Carnegie Mellon show systematic errors in causal event chains, mechanics, optics, fluids, and thermodynamics, while Stanford's velocity-field supervision and NYU's symplectic latent dynamics demonstrate that explicit physics inductive biases improve rollout consistency. The line has moved from perceptual plausibility toward geometric and dynamical constraints enforced during or after sampling, with Freiburg/Huawei's GeoART and KAIST/ETH's GeoFix refining camera and depth consistency and Zhejiang/NTU/Tsinghua/Alibaba's HA-Sampler separating sampling errors from physically implausible but model-likely outputs.

Recent work by Hartl and colleagues introduces the Kronecker Delta Rule and builds it into KronLSTM, a linear-attention architecture with an associative matrix memory. The update rule is proven to keep memory dynamics bounded, directly addressing the stability problem for long-context recall.

Recent work collectively shows that synthetic environments can be produced at scale from world models, video generators, or proposer agents and used to train agents in navigation, computer use, games, and racing (Tencent Hunyuan, UNC Chapel Hill with Microsoft, NUS with Tencent IEG, Toyota with UCSD). The line has moved from fixed simulators toward learned environment generation and validation: terminal agents build sandbox tasks from a solver's rollouts, and synthesized GUI or gameplay environments come with automated success checks.