Articles

The article examines the core business essence of 'features' versus 'products', clearly outlining their key characteristics. It emphasizes that a product must establish a closed-loop workflow and build data/ecosystem barriers, rather than being a mere, easily imitable feature. The cautionary tale of Sunrise Calendar, crushed by platforms, highlights the inherent risks faced by feature-based products. Conversely, the successful evolution of Dropbox and Midjourney from single functions to comprehensive ecosystems demonstrates effective strategies for moat-building. Ultimately, the article provides three crucial insights for AI entrepreneurs—'from single-point solutions to workflow lock-in', 'building data/relationship moats', and 'solving mission-critical (P0-level) pain points'—aimed at guiding them to create sustainable and valuable products.

This article thoroughly analyzes the eight major economic tasks for 2026, unveiled at the 2025 Central Economic Work Conference, and contrasts them with past policies. It highlights the 'Year of Intensive Cultivation' characteristic of current Chinese economic development. The core transformation involves a strategic shift from short-term 'external support' to long-term 'self-sustaining growth,' from a pursuit of speed to a focus on quality, and from reliance on subsidies to the activation of internal dynamics. Specifically, domestic demand stimulation is evolving from 'consumption subsidies' to 'increasing residents' income'; investment is transitioning from 'government-led' to 'stimulating private sector vitality'; innovation-driven development is moving from 'dedicated research' to 'accelerated implementation,' with a regional and entity focus on the Beijing-Tianjin-Hebei region, the Yangtze River Delta, the Greater Bay Area, and enterprises; tackling tough reforms is progressing from 'establishing rules' to 'removing obstacles,' emphasizing a unified national market and addressing detrimental internal competition; opening up to the outside world is deepening from 'opening up projects' to 'opening up rules,' promoting institutional opening-up and service exports; coordinated development is shifting from 'each stabilizing independently' to 'linked restructuring,' with county towns as carriers and urban agglomeration collaboration; the dual carbon goals are advancing from 'industrial upgrading' to 'systemic restructuring,' building a new energy system and a carbon market; people's livelihood policies are elevating from 'ensuring basic needs' to 'investing in people,' focusing on population size, education upgrading, and high-quality employment; and risk prevention is transitioning from 'short-term stabilization' to 'long-term systemic reconstruction,' promoting real estate model transformation and resolution of local government debt. The article envisions a more equitable, predictable, and robust future growth model and identifies corresponding market opportunities.

This article focuses on the inference optimization of Large Language Models (LLMs), providing an in-depth analysis of the critical role and technical details of Key-Value Cache (KV Cache). It first introduces how KV Cache improves LLM generation efficiency by reusing the Key and Value matrices of historical tokens, thus avoiding redundant computations. It then explains why only K/V are cached, rather than Q, and distinguishes between Prefilling (computation-intensive) and Decoding (memory bandwidth-intensive) as two key stages. The article points out that KV Cache becomes a bottleneck for GPU memory bandwidth in scenarios with long sequences and multi-batch processing. Based on this, it compares the differences between traditional ML Infra and AI Infra in terms of hardware stability, memory bandwidth, and model hardware evolution. Finally, it introduces optimization techniques such as PD/PS Separation, Speculative Decoding, and micro-system-level optimizations like DeepSeek V3's CPU+GPU Hybrid Addressing to address the challenges and improve performance in AI Infra.

The article introduces Agent Lightning, an open-source framework developed by Microsoft Research Asia – Shanghai, designed to integrate Reinforcement Learning (RL) into AI agents without extensive code modification. Traditional LLM-based agents often struggle with complex, multi-step tasks, and while RL can improve their decision-making, its adoption has been hindered by the need for significant code rewrites. Agent Lightning solves this by treating agent execution as a sequence of states and actions, converting experiences into a standardized format usable by RL algorithms. It employs a hierarchical RL approach, where the LightningRL algorithm assigns rewards to individual LLM requests within a multi-step task, allowing for efficient training with existing single-step RL algorithms. The framework acts as middleware, featuring an agent runner, an algorithm module, and the LightningStore for data management. This decoupled design offers algorithmic flexibility, resource efficiency, and scalability, as demonstrated by consistent performance improvements across text-to-SQL, RAG, and mathematical QA tasks.

This article introduces DingTalk v8.1.10's new AI features designed to significantly boost user efficiency in communication and meeting scenarios. Key updates include: **AI Smart Replies**, which intelligently generate high-EQ reply suggestions based on conversational context and relationships; **AI Meeting Notes Q&A**, evolving the AI assistant from mere transcription to intelligent information retrieval and Q&A based on content; and the **upgraded DingTalk A1 Recording Card**, offering real-time speech-to-text, multilingual simultaneous interpretation, and high-speed file transfer. These enhancements not only streamline individual workflows—such as HR interview analysis and student study reviews—but also provide robust solutions for cross-border business collaboration, fostering the widespread and deep integration of AI in daily office operations.

This article explores BabyAGI founder Yohei Nakajima's latest assessment of AI Agent development trends, highlighting that the ultimate goal for Agents is not prompt engineering, but rather establishing an "autonomous improvement loop" through self-improving mechanisms. The article elaborates on three advanced architectures for self-improving Agents: synthetic data generation, automated fine-tuning, and tool building, and stresses the importance of reliable evaluation mechanisms. Through this analysis, the article reveals the profound impact of this trend on technology, business models, and capital, foreseeing that future Agents will transform from "static" entities into "growable" "living systems." It also provides concrete action recommendations and screening criteria for AI entrepreneurs and investors.

This article introduces Zhipu AI's newly released GLM-ASR series of speech recognition models and their application products. GLM-ASR-2512 is a globally leading cloud-based speech recognition model, achieving an industry-leading low Character Error Rate (CER) in complex testing environments. Building on this, Zhipu also open-sourced the 1.5 billion parameter edge-side speech model GLM-ASR-Nano-2512. This model maintains high accuracy while offering stronger privacy protection and lower interaction latency, achieving SOTA performance in the open-source domain. The article highlights the "Zhipu AI Input Method" built upon these models. This input method not only provides precise speech-to-text functionality but also deeply integrates large model capabilities, supporting various smart operations such as translation, rewriting, sentiment transformation, and Vibe Coding. It also features whisper capture for public settings and efficient hotwords, aiming to elevate voice interaction from "turning speech into text" to an efficiency tool that "directly completes tasks with voice."

The article focuses on the paradigm of combining Vision-Language-Action (VLA) models with Reinforcement Learning (RL) in Embodied AI. It first points out the limitations of pure imitation learning and emphasizes the importance of Online RL in improving robot robustness and exploring better solutions. Next, the article dissects the three major technical difficulties faced when applying RL to VLA models: environmental discrepancies, model collapse and instability, and computational burden. Subsequently, it highlights and elaborates on the iRe-VLA two-stage cyclic iterative learning method proposed by Tsinghua University and Singularity Computing (a Chinese company). This method effectively overcomes the aforementioned challenges through an online reinforcement learning phase of exploration, where only the action head is trained (likened to "freezing the brain, training the limbs only"), and a supervised learning phase for consolidation, involving full model fine-tuning with mixed data. Experimental results demonstrate that iRe-VLA significantly improves training stability, performance in simulation environments, and real-world generalization capability, emphasizing the immense potential of this architecture for continuous robot learning and practical deployment, showcasing the leading role of Chinese teams in cutting-edge technologies.

The author organizes recent thoughts on the AI field, proposing several core viewpoints for AI product development. He believes that AI products should concentrate on fundamental capabilities and task orchestration, rather than current forms like Agents, which are susceptible to rapid iteration. Simultaneously, the article stresses the importance of 'aesthetic judgment' and attention to detail amidst the homogenization of AI products, arguing that AI currently struggles to mimic human sensibility and discernment. The author advocates for restrained product development, expanding functionality through a 'platform foundation + plugin ecosystem' model to avoid creating a 'jack-of-all-trades and master-of-none' product. He further suggests that the biggest difference between AI coding and human programmers lies in the generation of 'taste' or 'aesthetic judgment'.

Ant Group Digital Technologies announced the open-sourcing of the full set of papers, code, models, and user guides for its data analysis agent, Agentar SQL, at the second CCF China Data Conference. This technology aims to empower business users with natural language capabilities for data querying and analysis, significantly enhancing the data analysis capabilities of non-professionals. Agentar SQL currently leads the BIRD-SQL evaluation, surpassing many international vendors, and maintains a top position in both execution accuracy and efficiency rankings. The article elaborates on the four major challenges faced by NL2SQL in industrial implementation (understanding ambiguous spoken language, injecting industry knowledge, parsing complex databases, and generating accurate SQL). It highlights how Agentar SQL addresses these issues by developing a complete capability stack (online scaling, offline scaling, human-computer interaction, self-evolution). Ant Group Digital Technologies plans to gradually open-source more capability modules to promote the application of data agent technology in enterprise digital transformation.