Articles

This article deeply explores the core mechanisms and industry status of Interleaved Thinking in Large Language Models. It traces the evolution of Interleaved Thinking, from Anthropic's Extended Thinking to implementations by models such as OpenAI, MiniMax, Kimi, and DeepSeek. The core principle is that the model retains and accumulates reasoning states during multi-turn tool calling and reasoning, instead of simply discarding them. This significantly improves the performance of complex Agent tasks. The article points out that the lack of native support for Thinking fields in the OpenAI API has led the industry into a lack of standardization. Finally, the article affirms MiniMax's contributions to promoting Interleaved Thinking standardization and ecosystem building, discussing it within the broader context of Agentic AI development.

This article analyzes the current status and development trends of China's humanoid robot industry by thoroughly reviewing 60 representative domestic humanoid robot companies. Firstly, it points out that over half of the surveyed companies were established between 2023-2025, primarily driven by the ChatGPT boom and the pioneering influence of Tesla Optimus, with new companies emphasizing large models and embodied AI. Secondly, enterprises are highly concentrated in cities like Beijing, Shanghai, Suzhou, Hangzhou, and Shenzhen, areas known for their mature industrial systems and strong talent attraction. Lastly, despite variations in technical routes among companies, downstream application scenarios (factories, scientific research and education, hazardous working conditions, household chores) show substantial overlap, signaling intense future market competition. The article also details the background, core products, and technical highlights of each company, providing a clear overview of the humanoid robot industry for readers.

The article details how ByteDance's DataMind project builds an "AI + Data" unified data engine based on Apache Doris. The DataMind project addresses the limitations of traditional solutions in handling unstructured data and integrating with AI models. It integrates hybrid search capabilities, including Faiss-based vector indexing and Tablet-level BM25 scoring functions. At the same time, it complements AI functions such as AI_QUERY and TEXT_EMBEDDING, and supports Python UDF to meet custom model requirements. The article also delves into the implementation of GraphRAG (Graph Retrieval-Augmented Generation) construction and query mechanisms on DataMind, enhancing complex knowledge retrieval capabilities. Finally, it shares the implementation of an enterprise AI-powered data query platform, effectively solving practical challenges such as massive data volume, security, and query latency by optimizing data agent routing and connecting data lake permission systems, providing a solid foundation for the application of AI in enterprise data consumption.

The article introduces LatentMAS, a multi-agent inference framework proposed by Princeton University and other institutions, aimed at addressing the issues of low efficiency and information loss in traditional text-based multi-agent systems. LatentMAS innovatively allows agents to collaborate directly in the latent space. This is achieved by exchanging hidden layer representations and KV-cache working memory, rather than relying on text communication. Experiments demonstrate that LatentMAS achieves the highest accuracy improvement of 14.6% compared to single models and traditional text-based MAS on nine benchmark tasks, including mathematics, scientific reasoning, and code generation, while reducing output token usage by 70.8–83.7% and accelerating end-to-end inference speed by 4–4.3 times. This framework requires no additional training, is versatile, compatible with any HuggingFace model, and reveals that latent thinking has richer and more expressive semantic capabilities than discrete tokens.

The article focuses on the implementation of H5 animations for Taobao Live's 2025 Double 11 flagship interactive game, 'Huahuale,' delving into its high-performance, high-availability, and maintainable practical application. The author begins by analyzing the business context and animation requirements, proposing a hybrid technical solution that combines React, Phaser, Spine, and Lottie. It particularly highlights the advantages of Spine animation and its application in complex deformations and interactions. The article comprehensively explains how red packet flight path animations achieve smooth effects by exporting frame-by-frame data via AE scripts and dynamic frame interpolation, as well as the layer management for break-through visual effects. More importantly, the author shares engineering assurance practices for animations in mega-sale scenarios, including material delivery standards, loading control, animation degradation strategies, exception scenario handling, and a comprehensive monitoring system, offering valuable experience for running complex H5 animations in large-scale production environments.

This article introduces Nexus, a novel AI Model architecture by Huawei Noah's Ark Lab, designed to overcome limitations of Transformers in complex reasoning. By introducing a Higher-Order Attention Mechanism, Nexus provides context-awareness and multi-hop reasoning via a recursive framework. A key innovation is the significant improvement in model performance on complex tasks like mathematics and science without increasing parameters. Results show Nexus provides stable gains whether training small models from scratch or modifying existing large models (e.g., Qwen2.5), highlighting the significance of architectural innovation over mere scaling of parameters.

This article details how Figma Co-founder and CEO Dylan Field, after the IPO, led the company out of the shadow of the failed Adobe acquisition, evolving it from a single tool into a comprehensive platform encompassing Figma Design, FigJam, Slides, and the latest AI tool, Make. The article unfolds across six core dimensions: Firstly, Figma combats “corporate malaise” and maintains startup agility by squeezing out inefficiencies, guarding against path dependence, and establishing “Maker Week.” Secondly, through deep insight into user behavior and a “playful” product philosophy, it successfully expanded FigJam from a single tool into a multi-product matrix. Next, it delves into Figma's positioning for its AI tool Make and the lessons learned from its early failures, emphasizing quality and editability. Subsequently, it elaborates on Figma's “shorten time to value” product refinement philosophy, balancing foundational experience with product vision. Finally, Dylan Field shares his understanding of “taste” and predicts the rise of “product builders” in the AI-era workflow and the designer's new role transitioning from a visual specialist to an architect.

The article details the new ReCode agent paradigm proposed by the MetaGPT team (Foundation Agents Open Source Community), which addresses the limitations of current AI agent frameworks in decision granularity. It points out that ReAct's fine-grained loops or the forced separation in planner-based agents limit adaptability and generalization. The core insight of ReCode is that "planning and action are essentially the same thing, just with different granularity," unifying them through recursive code generation. ReCode represents high-level plans as placeholder functions and low-level actions as executable functions, allowing agents to freely mix and recursively expand, achieving dynamic adjustment of decision granularity. Experimental results show that ReCode's average performance in complex decision-making environments increased from 47.4% to 60.8%, inference costs decreased by 79%, and the number of training samples required decreased by 27%, significantly outperforming existing solutions such as ReAct, revolutionizing agent learning and development.

The article details the latest EyeReal glasses-free 3D display technology from a Chinese team appearing in Nature. By deeply integrating computational optics and artificial intelligence, this technology successfully overcomes the physical limitations of traditional glasses-free 3D displays in terms of screen size and viewing angle. EyeReal achieves a desktop monitor-sized 3D image without glasses, covering a wide viewing angle of over 100 degrees, and provides a continuous and smooth viewing experience. Its core innovation lies in the dynamic space-bandwidth product (SBP) utilization strategy, combined with real-time eye tracking, which precisely projects limited optical information to the observer's eyes. Technically, it transforms complex viewing angle changes into standardized planar images through physically accurate binocular geometric modeling and eye geometry encoding, and then uses a lightweight fully convolutional neural network (FCN) for reverse rendering to control the pixel phase of three common TFT-LCD panels, using Malus's Law to synthesize a precise light field. EyeReal achieves 'full parallax' and 'focus parallax', effectively solving the '3D sickness' problem that plagues users, and provides a low-cost, high-quality new solution for immersive experiences. The research was led by Ma Weijie, a doctoral student at Fudan University, and was jointly trained by institutions such as the Shanghai AI Laboratory.

This issue of Tech Weekly delves into the 'Space Data Center Controversy,' detailing the concept proposed by Musk and the Beijing Science & Tech Commission. Citing NASA engineers, it analyzes the significant technical hurdles and high costs in energy supply, heat dissipation, particle radiation protection, and data communication, highlighting the current challenges to its realization. Additionally, it features Qiniu Cloud's Large Language Model (LLM) access service, including details on supported models, request limits, and free tier. The issue also covers cutting-edge tech like the Icelandic ocean current crisis, Martian lightning sounds, artificial leaves, various software tools, AI applications, and industry perspectives on code quality in big companies and the potential for LLMs to homogenize personal voices.