Kimi K3 arrives with 2.8T parameters
Moonshot AI has launched Kimi K3 with 2.8 trillion parameters, a 1M-token context window and published API pricing, while the open-weight proof still depends on weights and model cards.
Page 23
View archive →RLinf v0.3 closes the robot training loop
RLinf v0.3 turns embodied AI training into a fuller pipeline, linking data collection, SFT, reinforcement learning, evaluation and real-robot deployment.
DeepMind turns AI toward biosecurity
Google DeepMind and Isomorphic Labs laid out a bioresilience program that uses frontier AI to prevent misuse, detect outbreaks and speed medical countermeasures.
Kingsoft launches two office AI agents
Kingsoft Office introduced Lingxi Pro and WPS Comate, shifting WPS AI from chat-style assistance toward editable documents, enterprise knowledge and workflow delivery.
Galbot lets robots learn from human video
Galbot, Peking University, CASIA and Tsinghua propose WAM-TTT, a deployment-time method that adapts robot behavior from ordinary first-person human videos without retraining the base world-action model.
MIT tests AI reasoning with 121 unfamiliar games
A Nature study from MIT, Cambridge, Princeton and collaborators suggests humans reason about new games with fast, shallow, goal-directed simulation rather than deep search.
Meta faces a 26-plaintiff lawsuit over algorithm-assisted layoffs
Meta faces a 26-plaintiff lawsuit over algorithm-assisted layoffs. The piece reviews the verified facts and why the signal matters beyond one announcement.
Dexmal DW0.5 moves robot post-training into a learned virtual trial field
Dexmal DW0.5 moves robot post-training into a learned virtual trial field. The piece reviews the verified facts and why the signal matters beyond one announcement.
Tencent open-sources two robot foundation models
Tencent released Hy-Embodied-VLM-1.0 and Hy-Embodied-RxBrain-1.0, splitting embodied AI cognition into scene-action understanding and multimodal subgoal planning.
Demis Hassabis proposes pre-release testing for frontier AI models
Demis Hassabis proposes pre-release testing for frontier AI models. The piece reviews the verified facts and why the signal matters beyond one announcement.