BestBlogs.dev Highlights Issue #64

GPT-5-Codex, OpenAI's latest breakthrough, is a GPT-5 model revolutionizing agent programming. It excels in real-world software engineering tasks, responding quickly to interactive sessions and independently completing complex tasks for up to 7 hours, including project construction, feature development, test writing, debugging, and large-scale refactoring. GPT-5-Codex excels in code review, proactively identifying critical vulnerabilities. It has already reviewed most of OpenAI's internal PRs. The model outperforms GPT-5 in SWE-bench Verified and Code refactoring tasks (key software engineering benchmarks), and dynamically adjusts thinking time based on task complexity. The article also introduces a series of upgrades to the Codex platform, including a newly designed open-source Codex CLI (supporting image input, to-do lists, tool calls, and permission management), plugins for IDEs such as VS Code (providing context awareness and seamless cloud-local switching), and deep integration with GitHub. OpenAI also emphasizes Codex's security measures, such as the default sandbox environment, permission mechanisms, and configurable security settings. Codex is included in various ChatGPT subscriptions, with API access coming soon, opening up new possibilities for developers.

This article introduces Tongyi DeepResearch, Alibaba's first open-source Agent model for deep research. Featuring a lightweight 30B-A3B architecture, it surpasses competitors like OpenAI Deep Research and DeepSeek-V3.1 with SOTA performance on benchmarks including HLE, BrowseComp-zh, and GAIA. The article explores key factors driving its capabilities: a multi-stage data strategy that uses large-scale, high-quality generated data without manual labeling, and the innovative IterResearch Paradigm and Research-Synthesis Framework, which address cognitive bottlenecks and noise in complex, multi-step tasks. Also discussed are the novel end-to-end Agent training process (Agentic CPT → Agentic SFT → Agentic RL) and policy-based reinforcement learning, highlighting the critical role of data quality and training environment stability for Agentic RL success. Finally, the article showcases Tongyi DeepResearch's successful applications in Amap and Tongyi Legal, providing open-source access to the model, framework, and solutions for practical use.

The article details the latest Digital Human technology, Kling-Avatar, released by the Kuaishou Kling team. This technology enables Digital Humans to perform vividly based on user intent, moving beyond simple lip-sync. At the core of Kling-Avatar is a two-stage generation framework powered by a Multimodal Large Language Model. First, a MLLM Director integrates audio, images, and text prompts into a structured storyline, generating a globally consistent blueprint video. Second, based on the keyframes of the blueprint video, the system employs a cascaded generation approach with parallel synthesis, combined with an audio-aligned interpolation strategy, to efficiently generate minute-level long videos, ensuring lip synchronization and identity consistency. The article also elaborates on training and inference strategies for lip synchronization, text controllability, and identity consistency. It introduces a high-quality training data pipeline and an evaluation benchmark containing 375 samples. Experimental results show that Kling-Avatar surpasses existing advanced products such as OmniHuman-1 and HeyGen in multiple dimensions, including overall effect, lip synchronization, picture quality, instruction response, and identity consistency, especially excelling in complex pronunciation and long-duration video generation. Currently, this feature has been launched for public testing on the Kling platform, marking a significant breakthrough in the expressive depth of Digital Human technology.

The article details the latest spatial intelligence model, Marble, launched by World Labs, a startup founded by Stanford University Professor Fei-Fei Li. The core capability of this model lies in its ability to achieve 'infinite exploration' of 3D worlds with just a single image or a text prompt, generating persistent, freely navigable 3D worlds. The article emphasizes Marble's advantages over existing technologies, such as the permanent, deformation-free, and consistent nature of its generated 3D worlds, as well as its larger scale, more diverse styles, and higher-quality geometric structures. Users can not only freely explore these worlds from any perspective in a browser but also export them as Gaussian Splatting (a 3D scene representation technique) and seamlessly integrate them into downstream projects such as Three.js through the open-source rendering library Spark, thereby building Web-based 3D experiences. Marble currently focuses primarily on the creation of 3D environments and does not yet support the generation of individual objects (such as people or animals). The article also provides a whitelist application address, inviting users to experience the preview version.

This article introduces SRPO (Semantic Relative Preference Optimization), the latest research achievement from Tencent Hunyuan. SRPO enhances the realism of portraits generated by text-to-image models, addressing the "oily skin" issue in the open-source model Flux. It innovatively uses positive and negative control prompts to adjust the reward model online, effectively avoiding reward hacking. The team also proposed "Direct-Align" to optimize the early generation trajectory, solving the overfitting problem on high-frequency information. This method achieves SOTA level with only 10 minutes of training, reducing training time by 75 times compared to DanceGRPO.

The article delves into the concept of the General Validator as a key advancement in LLMs, aiming to address the limitations of Reinforcement Learning with Verifiable Rewards (RLVR) in handling complex and subjective domains (beyond simple “right/wrong” judgments). The article elaborates on two major technical paths. The first path is “LLM-as-a-Judge,” which involves training powerful models to serve as evaluators. Among them, ScaleAI's RaR (Rubrics as Rewards) framework generates detailed, multi-dimensional scoring rubrics through a meta-framework defined by human experts, solving the scalability problem. Ant Group and Zhejiang University's Rubicon further refines the scoring system on this basis and introduces phased reinforcement learning to solve the Seesaw Effect in multi-skill training, improving the model's performance in humanities, creative, and other fields, and even reducing the artificiality of AI-generated responses. The Alibaba Quark team's Writing-Zero focuses on strengthening the referee model itself, forcing it to conduct critical analysis before scoring to improve the discrimination and reliability of the evaluation and avoid reward cheating. The second path is “Believe in the Power of the Model Itself and Let It Self-Evaluate” (Self-Evaluation). SEALab's VeriFree uses the model's “confidence” in standard answers as a reward signal but still relies on standard answers. UC Berkeley's INTUITOR goes a step further by calculating the self-determination (KL Divergence from a uniform distribution) of the model when generating each word to provide unsupervised internal rewards, without the need for external labels or standard answers, demonstrating a significant improvement in cross-domain generalization reasoning ability. The article points out that although these two paths are promising, they still have limitations: the “referee model” approach depends on manually created frameworks and is difficult to achieve full-domain coverage; the “self-evaluation model” path is limited by pre-trained knowledge and cannot verify external facts or create new knowledge. Ultimately, the article connects these explorations with the OaK Architecture vision proposed by Richard Sutton, the father of reinforcement learning, and believes that current efforts are building key components for future general intelligent agents that can learn and self-verify autonomously.

The article aims to correct the common misunderstanding of Model Context Protocol (MCP) among AI engineers, which is to simply regard it as "a more advanced Function Calling." The author, through rigorous "hypothesis-validation" logic, analyzes MCP from three perspectives—architecture analysis, SDK source code inspection, and Host dissection of the open-source project CherryStudio—arguing that MCP is essentially a model-agnostic engineering protocol for building interoperable AI applications. The article clearly distinguishes the responsibilities of MCP's Client-Host-Server (CHS) three components, emphasizing that the Host is the only component that carries AI intelligence (Prompt construction, LLM invocation), while the Server and Client are purely RPC middleware. Subsequently, the article deeply distinguishes the hierarchical relationship between MCP (infrastructure protocol) and Function Calling (model decision-making capability), and demonstrates the engineering advantages of MCP in decoupling, standardization, and interoperability through pseudo-code comparison. Finally, the article discusses the key factors that determine the application effect of MCP (tool quality, Prompt Engineering, LLM capability) and its inherent challenges (high Token cost, stability of intent recognition), offering AI engineers a comprehensive understanding and practical guidance on MCP.

This article delves into the issue of unreliable code resulting from developers coding by intuition in the era of AI programming. Addressing this pain point, the article introduces GitHub's open-source toolkit, Spec Kit, advocating for Spec-driven Development. This model overturns the traditional habit of coding first and documenting later, emphasizing the writing of executable living documents—specifications—as the sole source of truth for AI agents to generate, test, and validate code. The article elaborates on the four core stages of Spec Kit: Specify, Plan, Tasks, and Implement, as well as the roles of developers as helmsmen and validators within them. Spec Kit can be used in conjunction with AI tools such as GitHub Copilot and Claude Code, transforming vague prompts into clear intentions through a structured process, thereby improving the accuracy and reliability of AI-generated code. This method is particularly suitable for new project development, new feature development for existing systems, and modernization of old systems. The key advantage is decoupling 'what to do' from 'how to do it,' which fosters iteration and experimentation.

This article provides an in-depth look at how OpenAI's internal teams leverage Codex, their AI coding assistant, across engineering domains from security to infrastructure. It summarizes seven core application scenarios, including accelerating code understanding, efficient refactoring and migration, identifying and optimizing performance bottlenecks, improving test coverage, speeding up development, helping engineers maintain flow, and assisting in exploration and ideation. In addition, the article shares six practical best practices, such as starting with 'Ask mode,' organizing prompts like writing a GitHub Issue, and gradually improving the Codex environment, to help developers maximize Codex's effectiveness. The article highlights Codex's transformative impact on OpenAI's development processes and envisions the potential for deeper AI integration in software development.

This article details Agentic Coding, a new paradigm for AI in software development, emphasizing AI agents' ability to autonomously plan, decompose, execute, and iterate on complex development tasks, rather than being limited to code completion. Taking Alibaba Cloud's CLI Tool Qwen Code as an example, the article delves into its core Prompt Engineering, including role definition, core specifications, task management, and specific workflows for Software Engineering tasks and new AI Development. Based on the capabilities of the Qwen3-Coder series models, through these Prompts, Qwen Code can achieve goal-driven automated development processes, manage tool invocations, and autonomously perform building, testing, debugging, documentation generation, and version control, improving efficiency, quality, and developer oversight.

This article delves into the application of the AI-assisted programming tool Cursor in enhancing development efficiency, with a particular focus on its impact on legacy projects like WebX. The article first elaborates on the 'efficient usage' philosophy of AI-assisted programming, which involves letting AI handle the main programming tasks while developers act as code reviewers or solution architects. It then details the product features of Cursor, including core functions such as the AI chat area, Composer, and Bug Finder, and emphasizes the importance of introducing contextual information through Notepad and Rules to improve the accuracy of AI code generation. In the practical demonstration section, the article showcases how Cursor intelligently generates code skeletons adhering to complex specifications based on project design documents and existing coding styles, through two specific scenarios: building new features in existing projects (such as generating SQL, Mapper, Bean, Controller, and HSF services) and code refactoring optimization. Finally, it provides tips for using Cursor and envisions its potential integration with MCP (Multi-Cloud Management Platform), highlighting the importance of continuous practice and context accumulation for maximizing AI-assisted programming effectiveness.

This article explores the revolutionary AI-Native Development paradigm of 'Intent-as-Code,' aiming to elevate abstraction by allowing developers to define business intents in natural language, while AI handles implementation, exploration, and verification. It details three core pillars: Intent Orchestration, managing business logic and implicit data flows via visual canvases and structured intent trees; Resource Discovery, building an AI-understandable map of the external world for dynamic tool utilization; and Intent Constraints, ensuring AI-generated code reliability and predictability through contracts and behavioral testing. The article illustrates a complete AI-Native Development workflow with a 'User Login' example, highlighting its potential to boost development efficiency, ensure software correctness, and enable agile development, envisioning a shift from 'code craftsman' to 'thought creator'.

The article delves into practical strategies for AI collaborative programming, aiming to help developers leverage AI as a powerful tool. The author, through interviews with multiple founders utilizing AI for coding, summarizes a comprehensive 'AI Collaborative Programming Guide.' Key content includes: developing detailed plans at the project's outset, controlling the project scope, and adopting a small-step, incremental development approach; emphasizing the importance of version control, treating Git as a lifeline, and decisively resetting code when AI produces unexpected results; recommending prioritizing high-level integration testing and using it as a guardrail for AI's work to catch potential regressions. Furthermore, the guide provides methods for efficiently fixing bugs, such as effectively using error messages, adding logs, and switching models; optimizing AI tool configurations, creating instruction files, and leveraging local documentation to improve accuracy; and simplifying complex feature development by creating independent prototypes and modular architecture. The article also points out that selecting a mature and modular tech stack is crucial for AI performance and extends AI's applications beyond coding to DevOps automation, design assistance, content creation, and learning. Finally, it emphasizes the importance of continuous improvement and understanding the strengths of different models.

The article features an in-depth interview with Mati Staniszewski, CEO of AI Voice unicorn ElevenLabs, revealing how the company achieved rapid growth to $200 million ARR. By continuously exploring early product-market fit efforts, ElevenLabs shifted from movie dubbing to narration and voiceover, identified the real needs of creators, and achieved rapid user and business growth. Its success lies in a top-notch R&D team, rapid execution, and a deep focus on AI Voice application scenarios, while focusing on self-developed models and deploying multi-modal technology. In terms of team management, ElevenLabs adheres to a 'small team model,' emphasizing precise talent matching and rapid execution, avoiding the constraints of traditional hierarchies and titles. In terms of financing strategy, the company closely integrates financing announcements with product dynamics and user milestones, and emphasizes acquiring real users through community and vertical channels rather than over-relying on traditional media PR, the company values direct engagement. The article also explores the enormous commercial potential of AI Agents and how to choose investors who can truly provide assistance, offering valuable experience and inspiration for entrepreneurs.

The article reports in detail on the three major new products at the Meta Connect 2025 conference: Meta Ray-Ban Display (the first AI glasses with a display), Ray-Ban Meta (Gen 2) (consumer-focused upgrade), and Oakley Meta Vanguard (sports-oriented model). Among them, the Ray-Ban Display presents AI information intuitively and achieves precise air gesture control through the color waveguide HUD and Neural Band electromyographic neural interface wristband. The article also introduces the “Hyperscape” technology, which can scan real spaces into the Quest VR Headset, as well as updates to the Horizon platform's game engine and AI assistant. Despite the setbacks in the conference demo, the article believes that Meta is reshaping the computing platform through the combination of AI and hardware, demonstrating its ambition in the field of smart glasses and the Metaverse.

The article reveals popular AI Applications in a "parallel world" that is ignored by the mainstream AI community by analyzing the API call volume data of OpenRouter, an API aggregator that connects hundreds of LLMs. This list excludes products with self-built services or those tightly integrated with major platforms, focusing on the real needs of open-source projects, independent developers, and agile teams. The top ten are mainly divided into two categories: Coding Agents that serve developers and role-playing and entertainment applications that provide emotional value. The article details the functions, features, business models, and market performance of representative tools on the list, providing readers with a unique perspective to gain insight into the cutting-edge trends of AI applications.

The article explores the innovative applications of AI for mental health and self-healing. By introducing a Reddit community called 'Therapy GPT,' the author demonstrates how to use Large Language Models (such as ChatGPT) as private, non-judgmental, and always-online partners to help people cope with their 'Inner Critic,' process emotions, and explore their inner selves. The article selects and interprets 10 popular Structured Prompts, covering various psychological applications such as life coaching, Acceptance and Commitment Therapy (ACT), Trauma Transformation, Self-diagnosis, and Emotion Management. These Prompts are designed to guide users in deep self-reflection and emotional processing. The article emphasizes that AI cannot replace professional psychological counseling but can provide a safe, low-cost space for self-exploration and emotional relief, helping users better understand themselves and find self-acceptance.

The article delves into how to use the Lovart platform and its integrated Seedream 4.0 model to quickly transform any long article or document into image sets suitable for social media platforms such as Xiaohongshu (a popular Chinese social media and e-commerce platform). Through a series of practical cases, including thesis introduction cards, Chinese classical prose illustrations, biographies, and Xianxia-style (a genre of Chinese fantasy) science popularization, the author demonstrates in detail how to achieve precise control and iterative optimization of image style, text content, and layout through carefully designed prompts (prompt engineering) and Lovart's Magic Canvas local modification function. Lovart also features online search and information organization. This allows it to generate content directly from topics, such as the biography of Su Shi. The article emphasizes the great potential of this AI-assisted content creation method in improving the efficiency and diversity of social media content production, and mentions Lovart's current promotional activities to attract users to experience it.

The article reviews 10 innovative AI products from Product Hunt's latest monthly list, detailing their core functions, the problems they solve, and their application scenarios. These products include the workflow automation tool Trace, the AI job search tool Indy AI, the personal AI notebook Recall, Macaron AI (a personal AI assistant from a Chinese team), Anything and Floot (AI APP generation tools), the IDE Qoder with Agent functionality, the AI digital mentor nFactorial AI, the AI email processing tool Mocke, and the AI image processing tool X-Design. The article aims to help readers quickly understand the latest applications and product trends of current AI technology, inspire technologists and entrepreneurs, and demonstrate AI's broad potential in improving efficiency and enhancing life.

This article analyzes the latest AI usage reports from OpenAI and Anthropic. The OpenAI report indicates that ChatGPT has surpassed 700 million weekly active users with 18 billion weekly messages as of July 2025. Its core uses are practical advice, information retrieval, and writing, with non-work-related messages growing significantly while technical uses like programming decline. The report also reveals higher ChatGPT usage among individuals with higher education and income, with a narrowing gender gap. Anthropic's Economic Index Report highlights Claude's advantages in code writing and automation, with automated task delivery rising to 39%. Enterprise-level API customers, in particular, show a strong inclination towards automation, with up to 77% of tasks being fully automated. The article further explores the relationship between AI usage and regional economic structure and income levels, raising concerns about unequal distribution of AI benefits and potential widening of the wealth gap.

The article delves into key AI industry issues via an interview with OpenAI Chairman Bret Taylor. Taylor notes the 'performative' nature of many AI applications and a significant AI bubble, while affirming AI's long-term economic potential. He advocates for solution-centric AI companies, rather than AGI pursuit or self-developed models. Furthermore, he anticipates reduced fine-tuning importance due to increasing context windows and improved rule adherence. Taylor is particularly optimistic about AI Agents disrupting customer service, revolutionizing digital interaction with voice. He also introduces Sierra's pay-for-results model, shares insights on GPT-5's advancements, the evolving AGI definition, and perspectives on 'super intelligence' and safety.

This podcast invites A16Z co-founder and author of 'The Hard Thing About Hard Things,' Ben Horowitz, to share his profound insights on leadership, entrepreneurship, and artificial intelligence. Horowitz emphasizes that the worst choice for a leader is indecision, and the real value lies in making tough decisions that most people dislike, and the need to develop the mental muscle to confront the unknown. He uses personal experiences and pilot examples to illustrate that success is accumulated through a series of small but correct decisions. He also reveals for the first time the creation background of his classic article 'Good Product Manager, Bad Product Manager,' pointing out that a product manager is essentially a 'mini-CEO' who needs to lead the product to success through influence rather than authority. In terms of investment philosophy, Horowitz elaborates on A16Z's philosophy of 'investing in strengths rather than the absence of weaknesses,' using the controversial case of investing in Databricks and WeWork founder Adam Neumann as an example, emphasizing the importance of identifying and supporting entrepreneurs' world-class strengths. Addressing the current rampant 'AI Bubble Theory,' he sharply points out that when everyone thinks it is a bubble, it is often not a bubble, and believes that the current AI boom is a new era of technology based on real product and revenue growth. Horowitz also looks forward to the development trend of the AI industry in the next 5 to 10 years, believing that infrastructure, foundation models, and application layers all contain huge opportunities, and stresses the US's crucial role in leading AI innovation. Finally, he shared his charity 'Paid in Full' foundation, as well as insights on trust, culture building, and personal growth, offering listeners valuable, actionable insights.

This article reviews the 30-year development of software engineering in China through a dialogue with senior expert Wu Qiong. Wu Qiong's transformation from an 'introducer' to an 'innovator' is showcased, from introducing RUP and Agile methodologies to recognizing their cultural incompatibility and pioneering the localized Adapt Methodology. He further implemented his ideas through 'Zhiwei,' a software engineering tool platform. The article focuses on the disruptive impact of the AI Era, including challenges in processing private domain knowledge, the divergence of general agents into specialized agents, and the '1+N' organizational model treating agents as managed employees. Wu Qiong stresses that embracing AI transformation requires companies to unify their management information architecture and build flexible tool platforms. He notes AI will shift software from deterministic to probabilistic output, fundamentally altering development goals and processes, presenting both a significant challenge and a golden age for programmers.

Based on Ant Group's '2025 LLM Open Source Development Ecosystem Panorama 2.0' report, this article analyzes the significant transformations in the LLM open source ecosystem. The report highlights a rapid shakeup, with numerous projects exiting and new ones emerging, averaging a lifespan of less than three years. The replacement of TensorFlow by PyTorch exemplifies this drastic ecosystem shift. The classification framework has evolved from a traditional model to focus on three core areas: AI Agent, AI Infra, and AI Data, clearly outlining industry hotspots and technological evolution. The AI Agent layer is the most active, with AI Coding as the most active, highest-frequency, and most in-demand application, evolving from 'code completion' to a 'full lifecycle AI-powered engine'. Model Serving continues to boom, and LLMOps is taking over MLOps as a key enabler for LLM deployment. While relatively stable, the AI Data field is expected to evolve from a 'repository' to a 'center' in the future. The article also explores commercial variations in open source license agreements, revealing the commercial dynamics between openness and control in the LLM era. Finally, it provides additional insights into trends such as the divergence of LLM strategies, the increasing adoption of MoE architecture, the standardization of Reasoning capabilities, the widespread emergence of multimodality, and the diversification of model evaluation.