Why is more energy use bad for the environment?
RESEARCH
-
Major AI Alignment Improvements Announced
By
–
Alignment improvements are massive: 80% reduction in sycophancy
Less deception in adversarial tests
Reduced power-seeking behavior
Stops encouraging delusional thinking First frontier model evaluated with mechanistic interpretability techniques. -
Deeper reasoning across all domains
By
–
Reasoning gains across every domain. Finance professionals, lawyers, medical researchers, STEM academics all tested it independently. Universal feedback: "dramatically better domain knowledge and reasoning" than Opus 4.1. Not faster answers. Deeper thinking.
-

GLM 4.6 Release: Enhanced AI Capabilities
By
–
@Zai_org is preparing to release GLM 4.6! "As the latest iteration in the GLM series, GLM-4.6 achieves comprehensive enhancements across multiple domains, including real-world coding, long-context processing, reasoning, searching, writing, and agentic applications."
-
Proprioceptive States in Visuomotor Policies Research
By
–
#PapersAccepted by Jiqizhixin
Our report: https://
mp.weixin.qq.com/s/9W7rBdtTtW0W
cmSyodtPow
… Do You Need Proprioceptive States in Visuomotor Policies? Shanghai Jiao Tong University, Spirit AI, Tsinghua University, and others
Paper: https://
arxiv.org/abs/2509.18644
Project: https://
statefreepolicy.github.io -

State-free Robot Policy: Vision-only Visuomotor Learning Breakthrough
By
–
Do You Need Proprioceptive States in Visuomotor Policies? New robot manipulation breakthrough: Ditching proprioceptive inputs in imitation learning, the State-free Policy relies purely on visual observations. Full task-relevant vision via dual wrist cameras Huge spatial
-

Claude Sonnet 4.5 Claimed Best Coding Model on Benchmarks
By
–
Claude annonce Sonnet 4.5, présenté comme le meilleur modèle de code au monde. Du moins, sur les benchmarks… On attend de voir. Mais connaissant Claude, ça ne paraît pas déconnant.
-
xAI to Open Source Results for All Human Knowledge
By
–
xAI is doing this for all human knowledge. We will open source the results.
-

Program Synthesis via Test-Time Transduction: Active Learning Approach
By
–
Seoul National University & Adobe Present: Program Synthesis via Test-Time Transduction A formulation that explicitly leverages test inputs during synthesis, unlike prior approaches that only generalize from training examples. Key idea: treat synthesis as active learning over a
-

CAKE: Adaptive Kernel Design for Bayesian Optimization
By
–
Adaptive Kernel Design for Bayesian Optimization Is a Piece of CAKE with LLMs The Chinese University of Hong Kong, Shenzhen, and others
Paper: https://
alphaxiv.org/abs/2509.17998
v2
…
Code: https://
github.com/richardcsuwand
i/cake
…