BestBlogs.dev Highlights Issue #71

The article reports on Moonshot AI's latest open-source model, Kimi K2 Thinking, which is based on the core concept of 'Model as Agent' and possesses powerful thinking and tool-use capabilities, capable of executing 200-300 consecutive tool usages without human intervention. Technical details include 1TB parameters, 32B activation parameters, INT4 quantization, and a 256K context window. Kimi K2 Thinking refreshes the SOTA in benchmarks such as HLE, BrowseComp, and SEAL-0, outperforming GPT-5 and Claude Sonnet 4.5. The article emphasizes the comprehensive improvement of its agent capabilities, autonomous search and browsing capabilities, and Agentic programming capabilities, and points out that its INT4 quantization technology enhances compatibility with Chinese accelerated computing chips while improving reasoning speed. The model has been launched on the Kimi platform and the API is open-sourced under the MIT license, which is of great significance for promoting the development of open-source AI.

The article details the differing technical approaches and competition between MiniMax and Kimi (Moonshot AI) regarding attention mechanisms in LLMs, two major AI companies in China. MiniMax M2 chose to revert to Full Attention and provided a detailed explanation for abandoning Efficient Attention (such as Linear/Sparse Attention), citing the complexity of the engineering pipeline, limitations of the evaluation system, and immature infrastructure. MiniMax believes that the rapid advancement of GPUs will solve the cost problem of Full Attention. Conversely, Kimi (Moonshot AI) released Kimi Linear, featuring an innovative hybrid architecture of Kimi Delta Attention (KDA) and Multi-head Latent Attention (MLA). This architecture achieves significant KV Cache reduction and decoding throughput improvement, claiming to outperform Full Attention across various scenarios under fair comparisons for the first time. The article concludes that these two routes represent different explorations in the industry regarding the balance between efficiency and performance, and points out that this technical debate is also a competition for dominance in attention architectures.

The article delves into the efficiency challenges faced by AI Agents when using the Model Context Protocol (MCP), specifically, the excessive consumption of Tokens by tool definitions and intermediate results, leading to increased costs and delays. To address this issue, the article proposes treating the MCP server as a 'Code API,' allowing Agents to interact by writing code rather than directly calling tools. This 'Code Execution' mode allows Agents to load tool definitions on demand and process large amounts of data in the execution environment, avoiding unnecessary Token consumption. The article elaborates on the significant advantages of this approach in terms of saving context, protecting privacy, implementing complex logic, and enabling state persistence and skill development, while acknowledging the additional infrastructure costs associated with security sandboxing.

This article delves into the practical experiences of building and optimizing AI Agents, particularly within the 'Yun Xiaoer Aivis' project, an AI-powered customer service agent developed by Alibaba Cloud. The author argues that unpredictable Agent outputs stem from vague expectations and inadequate technical context management. Focusing on optimizing Context Engineering and Multi-Agent architectures, it summarizes ten practical experiences: clarifying expectations, accurate context delivery, defining system identity and historical execution, structured expression logic, custom tool protocols, effective use of Few-Shot learning, context optimization, memory management, balancing controllability and flexibility in Multi-Agent systems, and adhering to Human-in-the-Loop (HITL), a process where human feedback is integrated into the AI system's learning and refinement. These insights, based on real-world challenges and solutions, offer valuable guidance for Agent developers.

The article provides an in-depth analysis of how Perplexity AI engineered its 'AI Google' to deliver direct, cited answers rather than a list of links. It highlights Perplexity's core idea of combining real-time web search with powerful AI to overcome challenges like outdated information and hallucinations in traditional LLMs. The technical backbone is a meticulously implemented Retrieval-Augmented Generation (RAG) pipeline, encompassing query intent parsing, live web retrieval, snippet extraction, synthesized answer generation with citations, and conversational refinement. A key architectural strength is its model-agnostic orchestration layer, which intelligently routes queries to a mix of in-house fine-tuned Sonar models and third-party frontier models, balancing performance, cost, and strategic flexibility. For retrieval, Perplexity leverages Vespa AI for massive-scale, real-time indexing and hybrid search capabilities, ensuring freshness and fine-grained content understanding. The generation engine utilizes a hybrid approach, combining custom Sonar models with advanced third-party LLMs via Amazon Bedrock. Finally, the inference stack, powered by the custom-built ROSE engine, Rust, PyTorch, NVIDIA H100 GPUs, and Kubernetes on AWS, is hyper-optimized for speed and cost-efficiency. This integrated, end-to-end system is Perplexity's true competitive advantage.

This article introduces the concept and principles of ReAct (Reasoning and Acting). It addresses the challenges of traditional AI, where reasoning and action are often separated in complex tasks. ReAct achieves information acquisition, transparent reasoning, and complex collaboration through a cycle of 'thinking-acting-observing-adjusting.' Next, the article details how the LangGraph Framework abstracts the execution process of AI Agents into a directed graph, and efficiently implements ReAct through core mechanisms such as state-driven, conditional routing, and tool binding. Finally, the author combines a real-world project case of PPT Outline generation in an intelligent solution system, deeply analyzing the architectural evolution, tool design principles, prompt engineering practices, and performance optimization strategies of ReAct in addressing issues such as inaccurate, overloaded, and rigid output. It provides readers with valuable practical experience and solutions in tool design principles, prompt engineering practices, architectural evolution, and performance optimization strategies.

This article explores how Chrome DevTools MCP and AI coding assistants (like Gemini CLI) are revolutionizing front-end debugging. By granting AI direct access to the browser environment for DOM inspection, console output reading, and JavaScript execution, AI can effectively 'see' within the browser. Through detailed examples, the article demonstrates MCP server configuration and the use of AI to diagnose and fix console errors (e.g., 'React is not defined'), debug styling and layout, simulate user behavior for interactive testing, and assess network and CPU performance. The conclusion highlights MCP's potential to significantly improve debugging efficiency and usher in an era of AI-driven automated testing and intelligent error detection, underscoring its transformative impact on development efficiency and quality.

This article, the translated eighth chapter of 《Agent Design Patterns》, explores the challenges AI agents face without effective memory mechanisms. These include the inability to perform complex tasks, maintain coherent interactions, and continuously learn. To address this, the article systematically elaborates on the core concepts and practices of agent memory management. First, it clarifies that agents require a two-layer memory architecture, combining short-term contextual memory and long-term persistent knowledge, to achieve coherent interaction and continuous learning. Short-term memory resides within the context window of large language models. Long-term memory is implemented using external databases or vector databases, enabling semantic search. The article details the usage of the three major components of the Google ADK framework: Session (Session Management), State (Temporary State), and MemoryService (Long-Term Knowledge Base), as well as different storage options (such as InMemory, Database, Vertex AI). Next, it explores how LangChain and LangGraph manage short-term and long-term memory through tools such as ChatMessageHistory, ConversationBufferMemory, and BaseStore. It further divides long-term memory into semantic memory, episodic memory, and procedural memory. Finally, the article also mentions the Vertex AI Memory Bank service as a managed long-term memory solution. This article offers a comprehensive and practical guide to building intelligent systems that learn and remember.

This article introduces Spring AI's new Recursive Advisors, available from version 1.1.0-M4, designed to facilitate complex, iterative AI workflows that traditional single-pass advisors cannot handle. It begins by explaining the role of ChatClient and standard Advisors in intercepting and modifying AI interactions. The core innovation, Recursive Advisors, allows the advisor chain to loop multiple times, enabling use cases such as sequential tool calling, output validation with retries, and autonomous agentic loops. The article provides a clear implementation pattern for recursive advisors, highlighting the use of chain.copy(this).nextCall(...) for controlled iteration. It also details two built-in recursive advisors: ToolCallAdvisor, which brings tool execution within the advisor chain for greater control, and StructuredOutputValidationAdvisor, which validates LLM responses against a generated schema and retries on failure. Important considerations like the experimental nature, cost implications, and best practices for setting termination conditions are discussed. This feature significantly enhances Spring AI's capability for building sophisticated, agent-driven applications.

This article is written by a heavy Claude Code user. The author has accumulated rich experience in personal projects and the construction of enterprise-level AI-IDE tools. The article systematically analyzes almost all the functions of Claude Code, including the maintenance of CLAUDE.md, context management (avoiding /compact), custom slash commands, sub-agents (recommending a 'trunk-clone' architecture, where the main agent delegates tasks to clones, rather than a 'leader-expert' model), session recovery, hooks (recommending commit-time blocking), planning mode, Skills (considered more important than MCP), MCP (positioned as a security gateway rather than a bloated API), Claude Code SDK (for parallel scripts, building internal tools, and rapid prototyping), and GitHub Actions (GHA, for production operations). The author emphasizes that the key to using Claude Code efficiently is setting clear context and guardrails, enabling autonomous agent decision-making. The article also provides practical tips and highlights common pitfalls to help readers avoid detours.

The article details the experience of a seasoned developer, JokeGold5455, who spent six months using the AI programming assistant Claude Code to refactor a TypeScript Microservice project from approximately 100,000 lines of code into 300,000-400,000 lines. Despite the immense effort, even quipping that it “might shorten my life by five years,” he successfully transformed the project from a legacy system burdened with Technical Debt into a high-quality, efficient new system. The author delves into how to overcome the challenges AI faces in long-term, complex tasks, such as context loss, Skills failing to auto-activate, and inconsistent code quality. To tackle these issues, he built an innovative AI-assisted development system, including a Skills auto-activation system that uses Hooks for dynamic loading of skills, combined with PM2 for Backend Service management and debugging, and a Hooks system to ensure code quality and consistency. The article also shares planning processes, the evolution of the development documentation system, the application of Agents and Slash Commands, providing developers with a practical framework and valuable experience to maximize the effectiveness of AI programming assistants.

The article explores the evolving role of software engineers in an AI-driven future, moving from direct code implementers to managers of AI agents. It introduces two paradigms: "Conductor" and "Orchestrator." The Conductor guides a single AI agent synchronously and interactively on specific tasks, much like an AI pair programmer, with examples including Claude Code CLI and Cursor. The Orchestrator, conversely, oversees multiple autonomous AI agents asynchronously, delegating high-level goals and reviewing their outputs, often delivered as pull requests. This orchestrator model, exemplified by tools like GitHub Copilot Coding Agent, Google Jules, and OpenAI Codex, enables parallel development and significantly boosts productivity by abstracting away low-level coding. The article argues that this shift represents the next major leap in programming productivity, transforming engineers into strategic supervisors of AI teams and potentially leading to an explosion in software output, while emphasizing the continued importance of human judgment and oversight.

This article analyzes the Multi-Agent Public Opinion Analysis project '微舆 BettaFish', developed by a 20-year-old college student, which quickly topped GitHub Trending. The project surpasses traditional public opinion monitoring tools by using Insight, Media, Query, and Report Agents to automate data collection, in-depth analysis, and report writing. It extracts viewpoints, traces event origins, analyzes multi-modal content, and outputs complete, traceable reports. Using 'Wuhan University Brand Reputation Analysis' as an example, the article demonstrates the project's detailed insights into aspects such as emotional trends and event dissection. In addition, the article emphasizes the author's zero-to-one development process with the assistance of AI tools, as well as proactive promotion strategies, regarding it as a model for ordinary people to create miracles and realize personal value in the AI era, providing valuable practical references and career growth inspiration for technology practitioners.

The article analyzes in detail the evolution of AI Large Model evaluation systems, pointing out that traditional Benchmarks are gradually failing due to data contamination and the inability to simulate real interactions. Next, it focuses on LMArena (Large Model Arena) as a new type of dynamic evaluation mechanism, using user preferences to evaluate models through anonymous battles and Elo ranking systems. The article delves into the operating principles, technical innovations, and global impact of LMArena. At the same time, it objectively reveals the fairness crisis faced by LMArena, such as human bias, model manipulating the rankings, and the neutrality challenges that commercialization may bring. Finally, the article looks forward to the future evaluation system moving towards a hybrid approach, emphasizing the importance of high-quality expert data and the construction of reinforcement learning environments, pointing out that model evaluation is becoming the 'core science' of AI development.

The podcast features Melanie Perkins, co-founder and CEO of Canva, discussing her entrepreneurial journey. She introduces "Column B thinking," a philosophy of envisioning a "crazy big dream" and working backward to achieve it, contrasting it with incremental "Column A" approaches. Perkins emphasizes setting "crazy big goals" that inspire intense effort, even if the timeline is uncertain, and celebrating milestones. She recounts overcoming over 100 investor rejections by iteratively refining her pitch and maintaining a clear vision, alongside navigating a challenging two-year code rewrite that halted new product releases. The interview delves into Canva's product expansion, from graphics to documents, websites, and AI integration, driven by a mission to "empower the world to design anything, with every ingredient, in every language, on every device." A unique aspect highlighted is Canva's "two-step plan": building one of the world's most valuable companies while doing the most good, exemplified by significant philanthropic contributions and free educational products. Perkins also shares insights on leadership, the importance of work-life balance, community-driven product development, and her long-term vision for a better world by 2050, where basic human needs are met. The discussion underscores resilience, vision, and a deep commitment to both business success and global positive impact.

This article details the newly launched PPT generation feature of the Google Gemini APP. The author emphasizes that this feature is based on a client-side implementation, allowing users to control the style and content with high precision through prompts, and points out that its generation quality is better than other competing products. The article first provides a detailed user guide, including how to enable Canvas Mode in Gemini, use its built-in search capabilities to automatically fill PPT content, and seamlessly export the generated PDF file to Google Slides for editing, and finally convert it to PPTX format for Office. In addition, the article also mentions the potential of combining Gemini's in-depth research capabilities to generate PPTs. The core highlight is that the author explores through practice and shares four advanced PPT style prompt templates, covering Bento Grid, minimalist neutral colors, fluorescent green Swiss Internationalism design, and minimalist black and white style, and describes the design requirements for each style in detail. The article concludes by pointing out the current limitations of the feature in terms of the number of PPT pages, and suggests that users can use it as a high-quality PPT template generator.

This article delves into the current state, challenges, and successful strategies for implementing AI Agents in enterprise-grade production environments. Through a survey of over 30 AI Agent startups in Europe and interviews with more than 40 practitioners, the author reveals that the biggest challenges in current deployments are not technical, but rather workflow integration, human-computer interface (HCI), employee resistance, and data privacy and security issues. The article defines the key attributes of AI Agents (goal-oriented, reasoning, autonomy, persistence) and distinguishes them from traditional LLM chatbots and RPA. The survey found that while AI Agent deployment is shifting from innovation budgets to core business line budgets, daily usage among employees remains limited. In terms of pricing strategies, 'hybrid' and 'per-task pricing' are the most common, while 'pay-per-result' is less common due to attribution and measurement challenges. The article also analyzes suitable scenarios for different accuracy and autonomy configurations. To overcome challenges, the article proposes a 'think small' use case promotion strategy, providing 'hands-on' consulting services (such as the FDE and 3E frameworks), and flexible product positioning (augmentation rather than replacement, emphasizing practicality rather than novelty). The article concludes by looking at the future development direction of proactive and ambient agents, and previews that subsequent research will focus on context engineering, reliable execution, and robustness.

The article starts with the author's lament about the decline of Bilibili's Remix Zone, introducing the 'Renaissance' phenomenon brought about by AI music tools like Suno V5. It showcases the powerful expressiveness of AI music through hit cases like 'Black Bear Spirit'. Subsequently, the article breaks down the process of creating AI music videos using Suno V5, Gemini, and Jiemeng Digital Human platform, including lyric creation, style selection, music generation, and video production. Finally, the article elevates the theme, exploring how the reduced technical threshold in the AI era allows creators to return to the essence of expression, rekindle the Remix spirit, and celebrates this 'punk and romantic' Renaissance, emphasizing the importance of creativity and emotion.

This article summarizes the core findings of the '2025 Artificial Intelligence Index Report' by Stanford University's Fei-Fei Li team. The report indicates that the industry has become the dominant force in AI Research and Development (R&D), with the US and China engaged in intense competition in the AI sector. The US leads in model releases and high-impact research, while China leads in the number of research papers and patents granted. In terms of AI Performance, NVIDIA GPU performance has significantly improved, and open-source models are approaching the performance of closed-source models. Large Language Models (LLMs) have met or exceeded human benchmarks in language understanding, visual reasoning, code programming, and mathematical problem solving. Simultaneously, global attention to Responsible AI is surging, but unified standards are lacking. Economically, global AI investment is rapidly growing, with the United States far ahead in private investment and the number of companies receiving funding, while China dominates in the installation of industrial robots. The report also explores the application of AI in science and medicine, global AI legislative progress, and public perception and concerns about AI, noting that English-speaking countries express greater anxiety regarding AI.

This podcast, based on a 170-page in-depth PPT presentation, comprehensively analyzes the current status and future trends of the AI industry in 2025, covering technology, products, capital, and bubbles. At the technology level, the host delves into the evolution and challenges of natural language models, multi-modality, and Agents, including the dominance of reasoning models, the widespread application of reinforcement learning, difficulties in data acquisition and memory expansion, and the obsolescence of traditional Benchmark evaluation standards. The podcast also compares the competitive landscape and valuation differences between Chinese and American AI model companies. At the product level, using OpenAI as an example, it analyzes the dual strategy from technology R&D to productization, including ChatGPT's “all-in-one” application vision, market traffic changes of AI Coding tools, the development of general Agents with a focus on the Chinese market, and the competitive integration of browsers and AI assistants. At the capital level, the podcast deeply analyzes the significant capital investments of tech giants such as NVIDIA, Google, and Meta, market concentration, the surge in primary market financing, the emergence of AI unicorns, and changes in exit mechanisms, especially the dominant role of large funds in the market. Finally, it conducts an in-depth discussion on the AI bubble, analyzing its formation mechanism, the similarities and differences with historical bubbles (such as the Internet bubble), pointing out that the current bubble is more of a productive equity bubble, and proposes multiple dimensions for evaluating the bubble, emphasizing the importance of AI-human collaboration and rational investment. The overall content is highly informative and professional, providing listeners with forward-looking industry insights.

The article records the fireside chat between OpenAI CEO Sam Altman and Microsoft CEO Satya Nadella, and deeply discusses the unique cooperative relationship between the two companies, especially OpenAI's innovative structure with a non-profit foundation overseeing a for-profit entity, which aims to ensure that AGI benefits all mankind. The dialogue covers the inside story of Microsoft's 10 billion investment in OpenAI, equity structure, revenue sharing, and new agreement details such as exclusive model authorization. The two giants also had a candid exchange on core issues such as the definition and prospect of AGI, the planning of up to 1.4 trillion US dollars in computing power investment, the bottleneck of power and data center construction, the impact of AI on employment and productivity, and the reshaping of software and business models in the AI era, and expressed concerns about the challenges posed by varying AI regulatory laws across US states. They emphasized that computing power is the biggest constraint on current AI development, and are full of expectations for AI-driven scientific discoveries and new computing devices, demonstrating a deep understanding of both the opportunities and challenges in the future of AI.

In this a16z Runtime conference keynote, Marc Andreessen and Ben Horowitz delve into the current state and future implications of AI. They challenge common perceptions of AI's creative and intellectual limits, arguing that even if AI primarily 'remixes' information, its ability to surpass 99.99% of human capabilities makes it profoundly impactful. The discussion highlights that effective leadership and success involve more than just high IQ, emphasizing emotional intelligence, 'theory of mind,' and situational awareness—qualities AI currently struggles with. Addressing the 'AI bubble' concern, Horowitz dismisses it, citing immense current demand and the psychological nature of true bubbles. Andreessen adds that fundamental technological viability and customer willingness to pay are the key indicators. They foresee significant platform shifts beyond current chat interfaces and anticipate future 'oversupply cycles' in talent and chips, driven by economic incentives. A critical segment focuses on the US-China AI race, where China excels in implementation and scaling. Andreessen warns that while the West leads in AI software, China's manufacturing dominance poses a severe threat in the coming 'embodied AI' (robotics) phase, advocating for a reversal of Western de-industrialization to maintain competitiveness.

ElevenLabs CEO and co-founder Mati Staniszewski discusses the company's rapid delivery of research-grade AI products, including text-to-speech, AI music, and real-time voice agents. He highlights voice as the next human-computer interface and elaborates on their organizational model of small, autonomous teams, a global remote-first hiring strategy, and a 'no titles' policy. The interview covers balancing AI research with product releases, navigating the creative industry's adoption of AI through initiatives like their Voice Marketplace (which has paid creators over $10 million), and the complex transition from a creator-first approach to an enterprise platform. Staniszewski also shares lessons learned in hiring specialized talent, managing long sales cycles, and aligning internal incentives during company scaling.

Based on CB Insights' 'AI Agent Bible' report, this article provides a comprehensive overview of the AI Agent industry in 2025. It highlights that AI Agents have become a standard feature of AI products and predicts six major trends in the next 1-2 years: voice becoming the mainstream interaction method, a wave of mergers and acquisitions by large companies, severe inference costs, the criticality of AI Agent shopping payment infrastructure, competition for user data, and the growing need for AI Agent monitoring tools. The article further analyzes the distribution of AI Agent startups, pointing out that nearly half are focused on general-purpose tools, with customer service and software development showing the most significant commercialization progress. In terms of commercialization, coding AI Agents have the highest revenue, with a revenue per employee of $1.4 million, while customer service AI Agents have the highest valuation premium, with an average valuation multiple of 219x. The article also delves into the impact of high inference costs on business models and the industry's exploration of workload-based task pricing. Finally, the article provides a technology stack map of AI Agents and emphasizes the opportunities in the two emerging fields of payment and voice AI, as well as YC's four major bets on Agentic AI and the implementation status and trends in three vertical fields (shopping, manufacturing, finance/insurance).