BestBlogs.dev Highlights Issue #65

The article details the seven Tongyi series large language models (LLMs) released and upgraded at Alibaba's Yunqi Conference, comprehensively covering the full range of applications from text, vision, speech, and video to code and images. Among them, Qwen MAX, a trillion-parameter flagship model, excels in code generation and tool calling capabilities, achieving high scores in SWE-Bench Verified and AIME25 evaluations. Qwen3-Omni, a new generation multimodal large language model, adopts the Thinker-Talker Mixture of Experts (MoE) Architecture, realizing seamless integration of audio and video, image, and other multimodal capabilities with text intelligence, outperforming competitors in various speech and image tasks. Qwen3-VL focuses on visual understanding, supporting ultra-long video analysis, visual programming, and 3D spatial perception. In addition, Qwen-Image-Edit enhances multi-image editing and consistency maintenance. Qwen3-Coder improves project-level code understanding and repair capabilities. Wan2.5-Preview enables audio-visual synchronized video generation. Tongyi Bailing, an enterprise-level speech foundation large model, significantly solves the core pain points of hallucination output and cross-lingual speech in speech recognition through the Context Enhancement Architecture. By detailing the core capabilities, key upgrades, and evaluation data of each model, the article showcases the comprehensive progress of the Tongyi Large Model family in general intelligence and vertical applications, noting that all models are online, supporting one-click deployment or API calls.

This article announces the full API release of OpenAI's gpt-5-codex model, previously limited to a CLI tool. It details the pricing, which mirrors gpt-5, and emphasizes the significant 90% discount for cached input tokens, crucial for agentic workflows. The model is accessible via the Responses API, requiring the llm-openai-plugin for LLM integration, with new tool support largely self-authored by GPT-5 Codex. The article highlights the model's specialized nature for agentic and interactive coding, advocating a 'less is more' prompting principle due to its built-in coding best practices. Practical demonstrations include a pelican benchmark and a successful multimodal image description, showcasing its versatile capabilities.

This article details the release of updated versions of Google's Gemini 2.5 Flash and Flash-Lite models, now accessible via Google AI Studio and Vertex AI. The updates prioritize delivering higher quality and improved efficiency. Key enhancements for Gemini 2.5 Flash-Lite include significantly better instruction following, reduced verbosity for lower token costs and latency in high-throughput applications, and stronger multimodal and translation capabilities. The updated Gemini 2.5 Flash model features notable improvements in agentic tool use, demonstrating a 5% gain on SWE-Bench Verified, and enhanced cost-efficiency. To streamline access to preview versions, Google also introduces -latest aliases for each model family, while recommending stable versions for production applications. This release aims to gather user feedback to inform future stable model iterations.

This article announces Google DeepMind's latest advancements in embodied AI with the introduction of Gemini Robotics 1.5 and Gemini Robotics-ER 1.5. These models aim to create intelligent, general-purpose robots capable of solving complex, multi-step tasks. Gemini Robotics-ER 1.5 functions as the high-level brain, excelling in planning, logical decision-making, and state-of-the-art spatial understanding within physical environments. It can natively call digital tools like Google Search and generate detailed, multi-step plans. Gemini Robotics 1.5, a vision-language-action (VLA) model, translates these high-level plans into specific motor commands, allowing robots to "think before acting" and even explain their reasoning processes. A significant breakthrough is Gemini Robotics 1.5's ability to learn across different robot embodiments, accelerating skill transfer without specialized training. Gemini Robotics-ER 1.5 is now accessible to developers via the Gemini API in Google AI Studio, while Gemini Robotics 1.5 is available to select partners. The article also highlights Google's commitment to responsible AI development in robotics, implementing safety measures and releasing an upgraded ASIMOV benchmark for semantic safety evaluation.

Meituan's LongCat team officially released the efficient reasoning model LongCat-Flash-Thinking. While maintaining the high-speed performance of its previous version, this model significantly improves reasoning abilities in multiple fields such as logic, mathematics, code, and agents, reaching the State-of-the-Art (SOTA) level for global open-source models. It innovatively combines 'deep thinking + tool use' with 'non-formal + formal' reasoning capabilities, becoming the first LLM in China with this combination of capabilities. The article details its core innovative architecture, including the domain-parallel reinforcement learning (RL) training method for solving reinforcement learning stability, the Distributed Optimistic Resource Allocation (DORA) system for achieving efficient training on a ten-thousand-GPU cluster, and the dual-path agent reasoning framework and expert iterative formal reasoning framework for enhancing agent and formal reasoning capabilities. Multiple authoritative evaluation results show that LongCat-Flash-Thinking excels in general reasoning, mathematics, code, agents, and formal reasoning, with some indicators even surpassing or matching top closed-source models. This model has been fully open-sourced on HuggingFace and Github and provides an online experience.

This article details OpenAI's GDPval evaluation system, designed to quantify the economic value and practical application potential of Large Language Models (LLMs) by simulating 1320 real-world tasks across 9 industries and 44 professions. This evaluation moves AI beyond 'passing exams' to a GDP-based assessment, reflecting its impact on the economic system. Research shows that leading models like Claude Opus 4.1 and OpenAI's GPT-5 match or exceed human expert performance in nearly half the tasks, with AI completing tasks faster and at a lower cost. The article also explores AI's structural impact on the labor market, suggesting it can free humans from repetitive tasks, enabling them to focus on more creative endeavors and drive economic growth. The GDPval's open-source task set and evaluation platform aim to promote the widespread adoption of AI tools and industry development, helping humanity adapt to the changing times.

The article delves into the 80-point dilemma commonly found in AI Programming, where AI can complete most of the basic code but, due to the lack of project-specific business rules, coding standards, and other private knowledge, the generated code is difficult to use directly, requiring developers to invest a lot of time in fine-tuning. The author compares AI to new employees who are technically strong but lack business experience. To address this, they propose a three-part solution of Private Knowledge Engineering: First, conduct onboarding training for AI through Code Deconstruction and Business Analyst Prompt to establish a private knowledge base containing architecture, data models, business rules, and development specifications; second, combine Development Expert Prompt and private knowledge base for intelligent programming, enabling AI to generate code that meets project specifications in one go; finally, achieve automatic incremental updates of private knowledge through Document Auto-Maintenance Expert Prompt, forming a self-evolving knowledge ecosystem. The article demonstrates the significant effects of private knowledge engineering in improving code quality and development efficiency by comparing data before and after the transformation, and provides directly usable Prompt templates.

This article delves into how developers can shift from relying on single tools to establishing efficient collaboration models in the age of AI programming. The author points out that AI's most underestimated ability is 'reading code.' Through a structured four-element Prompt, the time to understand unfamiliar codebases can be reduced from days to hours. The article then proposes an 'Explore-Plan-Build-Commit' four-stage workflow, emphasizing the application of classic software engineering principles to AI collaboration to avoid instinct-driven programming. In terms of efficiency, the author redefines 'efficiency' as the total time to deliver robust solutions, rather than lines of code, pointing out that high-quality upfront design can significantly reduce later debugging costs. Finally, the article presents a four-quadrant decision framework based on the 'importance' and 'urgency' of tasks, guiding developers to choose the appropriate AI collaboration model in different scenarios, and emphasizes that the core competence of engineers will shift from 'solving problems' to 'defining problems' and 'designing solutions.'

This article provides an in-depth technical analysis of Claude Code, a terminal AI programming tool developed by Anthropic. It begins by introducing Claude Code's system architecture, centered around the Interaction Layer, Execution Layer, and Core Engine, and elaborates on the complete execution process from user command submission to result rendering. Subsequently, the article delves into key components: how the Interaction Layer handles user input and renders AI responses; how the Core Engine manages messages, queries AI models, and schedules tools; how the powerful Tool System interacts with the external environment through a unified interface; and how Context Management utilizes strategies such as LRU Cache, On-Demand Loading, and Result Truncation to provide the most relevant information within a limited context window. The article also shares technical insights, including the Binary Feedback testing mechanism, MCP Tool layered management, AI-Assisted Security Detection, Context Compression, and efficient file system strategies. Finally, the article introduces iFlow CLI 2.0, inspired by Claude Code and based on Gemini CLI adaptation and incorporating its features, detailing its installation method, multiple running modes, SubAgent functionality, Open Marketplace resources, and applications in scenarios such as code development, website creation, and DeepResearch.

This article, written by an AI voice product entrepreneur, deeply analyzes the technical practices and philosophical considerations from Context Engineering to AI Memory from a phenomenological perspective, with the core being how AI mimics human cognition and existence. The article first defines Context Engineering, emphasizing that it transcends Prompt Engineering and is the core of building a dynamic memory system for AI Agents, aiming to simulate human attention and memory mechanisms. Subsequently, by comparing the limited context window of LLMs with the similarity of human attention mechanisms, it points out that 'focused context' is superior to 'long context'. The article details the four strategies of Context Engineering: 'Write, Select, Compress, Isolate', and analogizes them to the construction of human consciousness. Next, it elaborates on the short-term and long-term, explicit and implicit mechanisms of human memory, and compares them with AI Memory, revealing the essential differences between carbon-based and silicon-based memory in terms of biological nature, emotion, consciousness, and forgetting. Finally, through a virtual dialogue with the philosopher Husserl, it explores whether AI Memory possesses true temporality, subjectivity, and emotional experience, and urges AI engineers to integrate philosophical considerations into their technological advancements to develop conscious AI that better reflects human existence.

This article delves into the development paradigm of AI applications in the Agent era, from LLM-driven simple dialogue and RAG approaches to the evolution of complex AI workflows and Agent-based paradigms. It elaborates on the AI application architecture within the Agent paradigm, including user interaction, core LLMs, environment modules, planning, execution, perception, reflection loops, and memory management. It then compares the differences between AI applications and traditional applications in the delivery process, focusing on key R&D infrastructure such as MaaS, memory solutions, MCP, AI gateways, Sandboxes, AI observability, and AI evaluation. Finally, the article analyzes the new security challenges and protection strategies faced by AI applications, such as prompt injection, tool usage security, identity authorization, and large model supply chain security, providing developers with a comprehensive practical guide.

This article is based on the "Dao, Methods, Techniques, and Tools" framework, deeply analyzing the changes and challenges faced by software development in the age of AI Programming. The author first points out common misunderstandings of developers when using AI tools, such as over-reliance or complete rejection, and the vague understanding of the boundaries of AI capabilities. Subsequently, the article elaborates on the three core elements of AI Programming: models, tools, and people, emphasizing the core position of people as commanders and decision-makers. The core content revolves around "Dao" (eternal software engineering principles, such as value, abstraction, simplicity, evolution, trade-offs, collaboration) and "Methods" (verified methodologies, such as Agile Lean, Design Patterns, TDD/BDD, Continuous Integration, Contract-Driven, Context Engineering), and deeply explores how these principles and methodologies are revitalized under AI empowerment and guide human-machine collaboration. The article emphasizes that AI Programming is not to replace humans, but to liberate developers from tedious code implementation, prompting the role to evolve from code implementers to intention-driven designers and intelligent orchestrators, and the core competitiveness shifts to business understanding and system design. Finally, it calls on developers to embrace change with a calm, exploratory, and intelligent attitude, and dance with AI to create greater value.

The article details ChatGPT Pulse, a new feature launched by OpenAI. This feature aims to change ChatGPT's previous passive mode of answering questions by proactively focusing on user needs, providing personalized updates while users sleep, and delivering a well-organized set of cards every morning. Pulse learns from users' conversation history, associated calendars, emails, and other mobile activity data, providing relevant information, creative inspiration, and action guidelines without prompting, such as itinerary planning, dinner suggestions, and fitness plans. OpenAI CEO Sam Altman praised it highly, calling it his 'favorite feature' and a 'competent personal assistant,' but some netizens worry that it could become an advertising platform. The design concept of this feature is an 'experience with an endpoint,' designed to solve problems efficiently rather than endlessly scrolling, and it is clearly stated that user feedback is only used to optimize personal exclusive experiences. Currently, the Pulse feature is only available to ChatGPT Pro users.

The article critically examines the obsolescence of the traditional "3-in-a-box" (PM, Dev, UX) hand-off model in AI-first product development, terming its failures as "Dev-Star" syndrome, which results in late, bloated, and off-target AI features. It argues that AI's probabilistic nature, minimal UI, and the necessity of real-user feedback invalidate upfront specification and siloed workflows. The author highlights how this old model leads to a "Telephone Pictionary" effect, distorting requirements and reducing user feedback to superficial UI preferences while core AI behavior remains untested. Citing high-profile AI project failures like IBM Watson and Zillow, the article asserts that AI is too critical to be confined to developers and data scientists. The proposed solution is the "Snowball model," emphasizing unified team collaboration, continuous iteration, and early, direct user involvement. This model prioritizes building working code over extensive documentation and advocates for a data-first approach, where initial prototypes simulate AI behavior directly (e.g., LLM + RAGs) before traditional UI design. The article issues a "code or die" rallying cry to UX designers, urging them to embrace "vibe coding" with AI assistance to become product catalysts, driving innovation by getting hands-on with implementation and validating real solutions rapidly.

This article presents a deep dive interview by Founder Park with Plaud China CEO Mo Zihao, revealing the secrets to Plaud's success and future vision as a leading AI hardware startup. Plaud's innovative AI recording cards (Note and NotePin) have enabled remarkable product-market fit (PMF) and rapid growth, generating hundreds of millions of US dollars in annual revenue. Mo Zihao emphasizes that Plaud's success stems not only from its unique product form but also from positioning the product as a sensor for users' real-world context. The article details core upgrades in Plaud 3.0, including multi-modal recording, Press to Highlight for capturing user intentions, and versatile template-based summarization, aimed at unlocking deeper value from conversations beyond human cognitive abilities. Plaud's product philosophy centers on aligning large language models (LLMs) with human intentions, leveraging AI's ultra-long memory, multi-faceted thinking, and proactive questioning to guide users toward better decisions. Plaud has elevated its product positioning from simple voice recorders and note-taking tools to a true work companion, serving "three-high" users (those with a high proportion of language-based tasks, high industry knowledge demands, and high decision-making leverage). Looking ahead, Plaud plans to cultivate a "petri dish" environment for agents to evolve autonomously based on user context, even embracing "hallucinations" to spark innovation. Mo Zihao also shares Plaud's strategies in team building, hardware advantages, and expansion in the Chinese market, while envisioning the future of AI-Native hardware.

The article delves into how CapCut has deeply integrated AI technology to become a comprehensive full-stack AI productivity tool. The author demonstrates CapCut's powerful capabilities in audio processing (such as strong noise reduction and vocal enhancement), video generation and editing (AI one-click transitions, image-to-video, AI text-to-video, one-click video from footage, album diary), content enhancement (video super-resolution, AI frame interpolation, AI expansion, AI object removal, AI lip sync, video translation), and AI music creation (AI-generated music, smart lyrics, lyrics rewrite and cover) through several specific examples. The article emphasizes CapCut's low barrier to entry, integrating ByteDance's internal Seedance and Seedream AI models, and points out that its monthly subscription price of 59 yuan is more cost-effective compared to many single-function AI products on the market. The author believes that CapCut has deeply integrated AI into its product DNA, and exemplifies the new generation of AI-powered applications with a vast user base and deeply integrated AI capabilities, far surpassing many native AI products.

The article deeply analyzes the product design philosophy of NoteBookLM, pointing out its "napkin philosophy" (referring to its simple, intuitive design) - solving the user's pain points of organizing materials when switching between multiple windows through a simple three-column structure (source of information, AI conversation, notes). The author emphasizes NoteBookLM's spirit of focus, not pursuing a large and comprehensive knowledge base, but concentrating on providing accurate and sourced answers based on user-provided materials. The article criticizes the current phenomenon of "feature stacking" in AI products, and believes that true innovation lies in prioritizing simplicity, that is, finding the essence of the product and having the courage to reject unnecessary functions. Through the case of NoteBookLM, the article calls on AI product developers to focus on user needs, maintain product tenacity, and achieve the design concept of "Think Smarter, Not Harder".

The article features an in-depth interview with researchers Nicole Brichtova and Oliver Wang from the core team of the Google Gemini 2.5 Flash image AI model (nicknamed Nano Banana). The team points out that the current image generation quality is approaching its peak. The core challenge in the future is to improve the AI model's ability to understand user intent, transforming it from a creative tool into an information query tool. They emphasize that integrating the 'world knowledge' of Large Language Models (LLMs) into image AI models is crucial, enabling them to handle more complex requirements. The article explores the future trends of multimodal interaction, considering UI design and user intent recognition as key challenges, especially in solving the challenge of starting from a blank canvas. For aesthetic requirements, the solution direction is deeply personalized contextual interaction. The team also suggests that AI model evaluation should be guided by real user feedback and predicts that image and video AI models will merge and develop into 'Omni Models', coexisting with traditional professional tools in the long term to meet the precision and creative needs of different users.

The article cites the rigorous research of Seyed M. Hosseini and Guy Lichtinger, doctoral students in economics at Harvard University, and their advisor, Larry Katz, revealing the significant impact of AI on the job market. The study cleverly uses the Difference-in-Differences (DiD) method (双重差分法). It identifies companies that recruit 'AI Integrators' as the experimental group, establishing a causal relationship between AI adoption and the impact on entry-level positions. This proves a sharp decline in entry-level hiring by AI adopters, rather than mass layoffs. This impact is prevalent in various industries, with the wholesale and retail industries being the hardest hit. In addition, the study found that the protection of prestigious academic qualifications against the AI impact presents a 'U-shaped curve,' with upper-middle-tier university graduates being the most affected. Based on this, the author proposes strategies for accelerating career advancement, considering 'tacit knowledge' and 'meta-skills,' and seeking ROI from interests, emphasizing the urgency and direction of personal career development in the AI era.

This podcast explores how organizational structure drives competitive advantage for AI startups. Guest Ren Chuan, co-founder of Palona AI, shares his team's practical experience in building an AI-Native Organization. Key points include: defaulting to AI for all R&D work, such as 90% of code being written by AI, reducing code review time from days to 10 minutes, and optimizing go-to-market strategies; the application of AI tools such as CodeRabbit, Linear+Devin, and incident.io in efficiency improvement; and reducing interpersonal interactions by improving communication efficiency through digitalization principles. Regarding talent management, engineers in the AI era need three major characteristics: 'Context Provider,' 'Fast Learner,' and 'Full Lifecycle Owner.' Human-AI collaboration must outperform AI alone. In terms of organizational structure, it advocates outcome-based division of labor rather than process-based division of labor, encourages engineers to communicate directly with customers, and predicts that future organizations may shift to a flexible model of a small number of partners and a large number of contractors. The podcast also discusses the challenges of large companies transforming to AI-Native models, as well as the advantages of startups in organizational innovation, providing technology practitioners with forward-looking insights and actionable practical advice.

The article records an in-depth interview between Luo Yihang, founder and CEO of Silicon Star, and Zhu Xiaohu, managing partner of GSR Ventures. Zhu Xiaohu first pointed out that Chinese open-source models represented by DeepSeek will become the 'new infrastructure (foundational technologies and infrastructure supporting AI development)' for global AI, ensuring AI openness. He emphasized that AI application start-ups need to have extremely fast Go-to-Market speed and user retention capabilities, and warned that AI programming is a 'utility' business subsidized by tech giants, which start-ups should avoid. In the field of robotics, Zhu Xiaohu prefers 'workhorse robots' that can create actual commercial value and completely replace positions; AI hardware should focus on 'subtraction,' focusing on core functions to achieve mass shipments. Regarding globalization, he believes that Chinese entrepreneurs should confidently go abroad as Chinese companies, with advantages in the consumer market, while the business-to-business (B2B) sector requires localized sales teams. Regardless of their location, they should operate as proud Chinese companies. Finally, Zhu Xiaohu elaborated on the importance of user engagement and cash recovery time in early-stage investments and predicted that opportunities in the AI era would evolve at three times the speed, and entrepreneurs need to find sustainable opportunities outside of tech giants.

This podcast features in-depth interviews with Bret Taylor, Chairman of the Board at OpenAI, and Clay Bavor, former senior executive at Google. These two legendary figures co-founded the AI company Sierra and discuss whether AI is a world-disrupting revolution or simply 'better software.' Bret Taylor argues that AI is revolutionizing the world by making 'intelligence' abundant, similar to how electricity and food became widely available. He believes this will fundamentally reshape socioeconomic structures and challenge human self-identity. The guests predict that the 'Agent' will become the core technological paradigm of the AI era, just as 'website' was for the Web era and 'application' for the mobile era. Agents are digital entities that can work autonomously, possess reasoning, and take action, becoming the primary interface for future interactions between people and businesses.

The podcast also highlights Sierra's disruptive 'pay-per-outcome' business model, where fees are charged only when the AI Agent successfully solves a problem for the client. This contrasts sharply with the traditional SaaS model, deeply aligning the interests of suppliers and customers. Additionally, they share the principle that AI-driven companies should 'focus on improving the AI system, not just correcting individual errors' and strongly oppose applied AI companies building their own foundation models. They argue that foundation model investments are huge and depreciate rapidly, and application-layer companies should focus on integrating and leveraging the best models to create exceptional user experiences. The podcast covers the far-reaching impact of AI on the speed of technology adoption, internet economic models, social structures, and personal identity, interspersed with valuable experiences and behind-the-scenes insights from the two guests' work at tech giants like Google, Facebook, and Salesforce.

The InfoQ AI, ML, and Data Engineering Trends Report 2025 provides a comprehensive overview of emerging technologies and their adoption trajectory, drawing insights from expert discussions on an accompanying podcast. Utilizing Geoffrey Moore's "Crossing the Chasm" model, the report categorizes trends into Innovators, Early Adopters, Early Majority, and Late Majority. Key innovations in the "Innovators" category include the rise of AI Agents transforming complex workflows, multi-modal language models for richer data understanding, and the significant emergence of Physical AI, embodying intelligence in robotics. New protocols like Model Context Protocol (MCP) are highlighted for enabling interoperability between AI systems, alongside evolving Human-Computer Interaction (HCI) driven by AI. In "Early Adopters," the report notes continued advancements in Language Models (e.g., GPT-5, SLMs, Vision LLMs) and the increasing commoditization of Retrieval Augmented Generation (RAG) in enterprise applications. The report also tracks the maturation and widespread adoption of key data engineering technologies, with Vector DBs, MLOps, and Synthetic Data moving into "Early Majority," and established technologies like Lakehouses, Stream Processing, and Distributed Computation entering "Late Majority." The report concludes with predictions for the next year, emphasizing the continued development of AI agents, a focus on practical utility, challenges in video RAG, and AI's increasingly subtle integration into daily life.

This 'Cyber Monthly' comprehensively reviews key AI industry dynamics in August 2025. Regarding language models, it notes the decreasing returns of capability improvements in new models like GPT-5. The industry focus is shifting towards reducing inference costs and exploring the potential of small models tailored for specific vertical applications. In contrast, image generation technologies like Nano-Banana have achieved breakthroughs in consistency and have become powerful productivity tools, indicating AI will reshape image editing. Video and audio fields are developing steadily, and digital human technology is accelerating, though application prospects remain uncertain. Progress in 3D and embodied intelligence is slow, while Agent technology evolves towards multi-agent collaboration and cloud and edge integration. AI Coding exhibits cloud-based, edge-based, and evolving interaction models. Application-wise, domestic firms favor internal integration, while overseas firms prioritize external collaboration. The article details August's technology releases, funding, and policy news, highlighting China's 'Artificial Intelligence+' action plan, providing practitioners with industry information and profound insights.