BestBlogs.dev Highlights Issue #67

Sam Altman unveils OpenAI's ambition to build an "AI Operating System" at DevDay 2025. The keynote's four core announcements—the Apps SDK for native ChatGPT applications, the Agent Kit for accelerating agent deployment, the official launch of Codex (powered by GPT-5) as a software engineering agent, and the Sora 2 API preview—collectively outline a new paradigm for software development. For developers, this is more than a tooling update; it is a practical demonstration of a future where software creation takes minutes rather than months, making it essential viewing to understand the next generation of the AI ecosystem.

The article introduces Google's Gemini 2.5 Computer Use model, a new specialized AI agent built on Gemini 2.5 Pro's visual understanding and reasoning capabilities. This model allows AI agents to interact with graphical user interfaces (UIs) by mimicking human actions such as clicking, typing, and scrolling, thereby enabling the automation of complex digital tasks like form filling and manipulating interactive elements. The core functionality is exposed via the computer_use tool in the Gemini API, operating in an iterative loop where the model analyzes screenshots and user requests to generate appropriate UI actions. Optimized primarily for web browsers and showing strong potential for mobile UI control, the model demonstrates state-of-the-art performance on multiple web and mobile control benchmarks, offering high accuracy at low latency. Google emphasizes a responsible approach to safety, integrating features directly into the model and providing developers with safety controls like per-step action assessments and user confirmations for high-risk actions. Early testers, including Google teams, have successfully applied the model for UI testing, workflow automation, and personal assistants, reporting significant improvements in efficiency and reliability. The model is now available in public preview through Google AI Studio and Vertex AI.

This article details Ling-1T, a large model launched by the Ling Team. It is a trillion-parameter, open-source, flagship non-deliberative model built upon the Ling 2.0 architecture. Ling-1T achieves state-of-the-art results in complex reasoning, code generation, front-end development, and cross-domain generalization, balancing efficient reasoning with precise output. It supports a context window of up to 128K tokens and enhances reasoning capabilities through a pre-training and post-training Evolutionary Chain-of-Thought (Evo-CoT) approach. During training, Ling-1T, the largest known foundation model trained with FP8 mixed precision, utilizes a heterogeneous pipeline with fine-grained optimization, significantly improving training efficiency and stability. In the post-training phase, an LPO (Linguistics-Unit Policy Optimization) strategy at the sentence level addresses the limitations of traditional reinforcement learning, enhancing training stability and model generalization. The article also highlights Ling-1T's exceptional performance in visualization and front-end development, agent tool calling, while acknowledging limitations like the high inference cost of the GQA architecture and the need for improved agent capabilities and instruction following. Future iteration plans, open-source links, and access to experience pages are also provided.

This article introduces Jina Reranker v3, the third-generation reranker from Jina AI. With only 600 million parameters, it achieves state-of-the-art (SOTA) performance on multiple multilingual retrieval benchmarks, surpassing Qwen3-Reranker-4B with 6x more parameters on the BEIR benchmark. Its core innovation lies in the adoption of Listwise input and a novel "last but not late" interaction mechanism. This mechanism enables deep interaction between queries and all documents within a single context window through causal attention, leveraging global context information between documents to enhance ranking accuracy. The article highlights the model's performance in English (BEIR) and cross-lingual (MIRACL, MKQA) evaluations and its head result stability across various input orders. Jina Reranker v3 also provides GGUF and MLX formats and API interfaces for easy deployment and integration across diverse hardware environments.

The article provides a detailed introduction to the Tencent Hunyuan Image 3.0 model. This model has achieved the top global ranking on the LMArena international text-to-image leaderboard, surpassing models from Google (Nano Banana), ByteDance (Seedream), and OpenAI (gpt-Image). Hunyuan Image 3.0 adopts a native multimodal architecture, enabling it to uniformly process various modalities of input and output such as text, images, video, and audio through a single model, possessing both drawing capabilities and reasoning capabilities grounded in common knowledge. It is based on the Hunyuan-A13B Large Language Model, with a parameter scale of up to 80 billion, making it the industry's first open-source industrial-grade native multimodal image generation model. The article delves into its technical solutions, including a hybrid discrete-continuous modeling strategy combining text autoregression and image diffusion, a generalized causal attention mechanism for processing heterogeneous data, a generalized two-dimensional RoPE compatible with pre-trained LLMs, and a mode that automatically determines image shapes based on context. In terms of data processing, the model employs a three-stage filtering process, a hierarchical Chinese-English description system and the creation of Chain-of-Thought (CoT) based reasoning datasets. The model's training process is divided into four progressive stages, supplemented by post-training optimization techniques such as Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Mix Gradient Ratio Preference Optimization (MixGRPO). Evaluation results show that Hunyuan Image 3.0 performs excellently on both machine metric SSAE and human evaluation GSB, with generation results comparable to or even surpassing top closed-source models in the industry, demonstrating strong technical strength and broad application potential.

The article elaborates on the concept of Context Engineering, contrasting it with traditional Prompt Engineering, emphasizing its focus on automated and holistic input management for improved AI Agent performance. It emphasizes that the essence of a language model is text continuation, and to obtain ideal output, optimizing the input (i.e., the context) is crucial in addition to model training. The article breaks down a complete context into seven major components, including the user prompt, system prompt, dialogue history, long-term memory, external data sources (RAG), tool usage, and the model's reasoning process. Subsequently, the article delves into the importance of Context Engineering in the era of AI Agents, revealing challenges such as 'Lost in the Middle' and 'Context Rot' brought about by long context windows. Finally, it proposes three core strategies for Context Engineering: Selection to filter relevant information, Compression to refine historical records, and Multi-Agent division of labor to isolate and manage the context of their respective domains, thereby effectively improving the stability and reliability of AI Agents.

This OpenAI DevDay presentation details the significant advancements and new tools for developers across OpenAI's ecosystem. It traces OpenAI's journey from foundational research in reinforcement and unsupervised learning to the creation of powerful models like GPT-3 and the current GPT-5. Key announcements include the release of GPT-5, optimized for agentic tasks and advanced coding, along with principles for its effective use. The video introduces Sora 2 API for high-quality video generation and smaller, more cost-effective speech and image generation models. A major highlight is GPT-OSS, an open-source initiative aimed at democratizing AI. The presentation extensively covers enhancements to Codex, including GPT-5 Codex for agentic coding, Slack integration, MCP support for tools like Figma and Chrome DevTools, GitHub code review, and the new Codex SDK for embedding coding intelligence into custom workflows and applications. Furthermore, the Agent Kit, built on the Responses API, is introduced as a robust framework for building sophisticated AI agents, showcased through RAMP's procurement agent. Finally, the Apps SDK for ChatGPT is unveiled, allowing developers to create fully interactive, natural language-responsive applications directly within ChatGPT, demonstrated with examples like controlling lights, creating music, and personalized learning experiences. The overall message emphasizes empowering developers to shape the future of AI and software engineering.

Recorded live at OpenAI DevDay, this podcast episode features interviews with representatives from SchoolAI, Jam.dev, Abridge, and Cursor. Each startup discusses how they leverage AI, particularly OpenAI's tools, to innovate within their respective sectors: education, web development, healthcare, and software engineering. They share insights into product development, the transformative impact of new AI tools like the Agent Builder and GPT Builder, and their visions for the future of AI. Key themes include empowering educators and students with AI tutors, enabling non-technical users to fix website issues, alleviating medical documentation burdens, and evolving software engineering with AI-powered coding assistants. The discussions highlight the shift towards more intuitive, collaborative, and self-optimizing AI applications, emphasizing the importance of user experience, trust-building in high-stakes environments, and the "Cambrian explosion" of software creation enabled by accessible AI. The podcast also offers advice for developers and founders, underscoring the early stages of AI adoption and the continuous need for practical, user-centric solutions.

This article, based on an OpenAI presentation, highlights the critical importance of AI model evaluation. It introduces OpenAI's internal 'GDP Eval' framework, designed to assess frontier models' performance on economically valuable, real-world tasks, moving beyond traditional academic benchmarks. GDP Eval employs expert pairwise grading to compare model outputs against human performance across diverse industries and professions, demonstrating significant progress in models like GPT-5. It also serves as a proactive measure to track AI's impact on the workforce and acts as a 'North Star' metric for internal research. However, it acknowledges limitations, primarily measuring performance on clearly defined tasks rather than the full complexity of real-world jobs involving prioritization or iteration. The second segment focuses on OpenAI's 'Evals product,' a suite of tools for developers to rigorously evaluate their AI applications and agents. Key new features include Datasets for building evaluations, Traces for debugging multi-agent systems, Automated Prompt Optimization to accelerate iteration, support for Third-Party Models, and enterprise-grade capabilities. The presentation underscores that robust evaluation is crucial for building high-performing AI applications, particularly in sensitive domains, by addressing challenges such as LLM non-determinism and compounding errors in agent systems. It concludes with best practices for developers, advocating for early and continuous evaluation using real human data and expert-guided automation.

The article delves into LangChain's strategic decision not to develop a visual workflow builder, contrasting their approach with recent industry moves like OpenAI's AgentKit. It highlights that the primary motivation for such builders is to empower non-technical users to create agents due to engineering resource constraints and domain knowledge. A key distinction is made between 'workflows' (predictability over autonomy) and 'agents' (autonomy over predictability), emphasizing the pursuit of 'reliably good' outcomes. The author argues that visual workflow builders are not truly low-barrier-to-entry and become unmanageably complex for intricate tasks. Instead, the article proposes that future solutions will gravitate towards simple no-code agents for low-complexity problems and code-based workflows (like LangGraph) for high-complexity scenarios, especially as code generation improves. The core 'interesting problems' lie in making reliably good no-code agents easier to create and enhancing code generation models for LLM-powered workflows/agents.

This is a deep practical guide to building an enterprise-grade Intelligent Incident Review Agent. The article addresses the pain points of "blame avoidance" and "shallow attribution" in SRE post-mortems, detailing how to utilize a Multi-Agent architecture to transform passive response into active defense. Key highlights include engineering details: it proposes a "Denoise-Summarize-Preserve" Memory management strategy to solve Token overflow issues and demonstrates a four-stage Prompt evolution logic from "generalized tagging" to "fact-based questioning." Furthermore, the article abandons traditional text similarity metrics (like BLEU/ROUGE) in favor of an LLM-as-Judge evaluation system based on business value.

The article introduces 'vibe engineering' as a disciplined, accountable approach for seasoned software engineers, distinguishing it from 'vibe coding,' which denotes a fast, irresponsible use of AI. This new paradigm, significantly enabled by the recent rise of coding agents (like Claude Code, Codex CLI, and Gemini CLI) that can iterate, test, and modify code, allows experienced professionals to accelerate their work with LLMs while maintaining full accountability for production-quality software. The author emphasizes that effective LLM integration for non-toy projects is challenging and highlights how LLMs amplify existing top-tier software engineering practices. These include automated testing, thorough planning, comprehensive documentation, strong version control habits, effective automation, a culture of code review, a unique form of management, robust manual QA, strong research skills, the ability to ship to preview environments, an instinct for what to outsource, and an updated sense of estimation. The article argues that these tools empower senior engineers, amplifying their expertise, and justifies the potentially controversial name 'vibe engineering' for its clear distinction and memorable nature.

Albert Cheng, a growth leader from Duolingo, Grammarly, and Chess.com, shares his unique "Explore & Exploit" framework for identifying and scaling growth opportunities. He emphasizes rapid experimentation, a deep understanding of user psychology, and the critical importance of user retention in consumer subscription products. Key insights include Grammarly's successful strategy of exposing free users to premium features to double conversion, the significance of 'reviving' churned users, and the transformative application of AI to accelerate growth experiments (e.g., text-to-SQL bots, AI prototyping). The discussion also covers retention benchmarks for consumer apps, the nuances of freemium vs. trial models, successful gamification strategies (core loop, metagame, profile), and how AI is influencing both product functionality (like Chess.com's coaching) and the evolving role of a growth expert. Cheng highlights the value of 'high agency' individuals in team building and the importance of fostering an experimentation-driven company culture, drawing lessons from his diverse experiences at highly successful companies.

This session deeply deconstructs how top creator Dan Koe leverages AI (primarily Claude and ChatGPT) to build a high-efficiency content ecosystem. Unlike the common "auto-generate" approach, Dan proposes a "reverse engineering" methodology: asking AI to analyze the deep structure and psychological patterns of viral content, converting them into SOPs, and then using interactive Meta-Prompts to guide AI in generating on-brand drafts. The video details a complete workflow using Twitter as a creative testing ground before scaling validated ideas into Newsletters and YouTube videos. It is a must-watch for creators and product managers looking to use LLMs as a force multiplier for thought leadership rather than a replacement for thinking.

The article presents 'Intent Prototyping,' a disciplined methodology leveraging AI to bridge the gap between design intent (UI sketches, conceptual models, user flows) and a functional, live prototype. It addresses the 'lopsided horse' problem of mockup-centric design and the ambiguity of 'vibe coding' by emphasizing clear, unambiguous specifications. The workflow involves four steps: expressing intent (sketches, conceptual model via multimodal LLMs like Gemini 2.5 Pro), preparing technical specifications and plans (AI-generated), executing the plan (agentic AI like Gemini CLI building DAL and UI), and continuous learning and iteration through user testing. This method is particularly suited for complex enterprise applications, allowing early testing of underlying logic and preventing design debt. The author contrasts Intent Prototyping with other design tools like Figma and Axure, highlighting its unique strength in mitigating architectural flaws. Ultimately, it shifts the design focus from creating 'pictures of a product' to architecting 'blueprints for a system' using AI as a powerful enabler.

This AMA features founders of Decagon (AI customer support agents) and Clay (AI-driven GTM platform) alongside an Andreessen Horowitz investor. They delve into critical aspects of scaling AI applications for enterprise, including methodologies for evaluating novel AI models and ensuring infrastructure flexibility in a rapidly evolving market. The discussion also covers strategies for balancing AI experimentation with essential enterprise safety guardrails, overcoming common deployment failures by focusing on quantifiable ROI and iterative launches, and achieving product differentiation in a crowded AI landscape through unique market philosophies and empowering non-technical users. Finally, they offer advice on resource prioritization and key considerations for new enterprise AI ventures, emphasizing self-awareness and following genuine curiosity.

The article presents the transformation of serial entrepreneur Liu Ye from the founder of the education unicorn 'Homework Box' (Zuoyebang) to the AI language learning product 'Talkit' through an in-depth interview. After the 'Double Reduction Policy' severely impacted the education industry, Liu Ye experienced three years of confusion and exploration, during which he studied various fields such as healthcare and coffee chain, and finally re-invested in the AI + education track. He believes AI revolutionizes oral English learning. Based on Task-Based Language Teaching (TBLT), he created Talkit, an AI x 3D virtual world for oral practice. Talkit aims to immerse users in a language learning experience akin to immigrating to an English-speaking country. The article details Talkit's three major modules: 'Course,' 'Rehearsal,' and 'Social,' as well as its core technological concept of generating virtual characters, scenes, and tasks through the 'Gen World Engine.' The interview also delves into Liu Ye's understanding of competitors such as Duolingo, as well as his mindset adjustments regarding entrepreneurial philosophy, coping with major industry changes, and advice for the new generation of AI entrepreneurs. He emphasizes that entrepreneurship should pursue things that are 'unquestionably valuable' and 'sufficiently difficult,' and uphold the belief of 'Vision > Circumstances > Skills'.

This podcast features a conversation with Chen Mian, the founder of Lovart, providing an in-depth analysis of the profound impact of OpenAI's Sora release on the AI industry, particularly on AI To Consumer applications and the social sector. Chen Mian shares his experience using Sora, emphasizing its innovation in video generation quality, cinematography, and social features such as Remix Co-creation. He predicts that Sora could become a virtual social super-application with billions of users. The podcast also reviews Lovart's rapid growth, achieving 200,000 daily active users (DAU) and a $30 million annual recurring revenue (ARR) prediction within four months. Chen Mian discusses Lovart's product vision of democratizing creation, highlighting its role as an AI-powered design agent that empowers everyone to create. Furthermore, the conversation covers globalization strategies, differences in perceptions of the AI market between the US and China, how AI application companies can achieve growth through predictive model evolution, and the importance of entrepreneurs embracing a sense of urgency and building a highly iterative team in the context of accelerated technological iteration. The podcast concludes by highlighting the vast potential of the AI era To Consumer Market, along with the challenges and opportunities for startups.

This conversation between Sam Altman of OpenAI and Jony Ive of LoveFrom delves into their partnership aimed at creating new AI-powered devices. Jony Ive recounts how ChatGPT clarified his team's mission to build exceptional creative teams, leading to their collaboration with OpenAI. They explore the iterative design process, highlighting the importance of deep motivation and 'craft and care'—a commitment to unseen details driven by a belief in humanity's deserving of better tools. Both emphasize the need to move beyond existing device paradigms (like the smartphone) to truly harness AI's capabilities, envisioning interfaces that evoke delight and reduce anxiety, and fundamentally rethinking the nature of operating systems and user interfaces. They acknowledge the challenge posed by AI's rapid development, which generates a multitude of compelling product ideas, making focus difficult. The discussion concludes with a shared hope that AI tools will ultimately lead to more fulfilling, peaceful, and less alienating human experiences, rejecting the notion that current tech interactions are the immutable norm.

This Lightcone episode delves into Hamilton Helmer's 'Seven Powers' framework, adapting it for the contemporary AI startup environment. It addresses the increasing concern among founders about building 'moats'—defensive strategies—against competition, especially given the perception of AI applications as easily replicable 'ChatGPT wrappers.' The discussion emphasizes that early-stage startups should prioritize speed as their initial moat and focus on solving real customer problems, as other moats are only relevant once a valuable product exists. The article then reinterprets Helmer's seven powers, including Process Power (complex, mission-critical AI agents requiring extensive real-world refinement), Cornered Resources (proprietary data, fine-tuned models, or strategic government/regulatory integrations), Switching Costs (deep workflow customization of agent logic), Counter-Positioning (innovative outcome-based pricing models and agile product development to disrupt incumbents), Network Economies (data-driven model improvement), and Scale Economies (foundational model infrastructure). It also highlights Brand as a significant moat, especially in consumer AI, and discusses the impact of AI on labor replacement, concluding with advice for founders to focus on acute pain points before overthinking long-term defensibility.

Jordan Fisher, co-founder of Standard AI and AI alignment researcher at Anthropic, presents a series of critical questions for AI founders to ponder in an era potentially years away from Artificial General Intelligence (AGI). He emphasizes the current state of confusion as a fertile ground for innovation and stresses that founders must plan for AGI's impact on strategy, product, and team building, looking beyond the next 6 months to a 2-year horizon. Key themes include the potential commoditization of software, the shift towards AI-native teams, and the paramount importance of trust, security, and alignment in a world increasingly reliant on AI agents. Fisher also explores the concept of 'defensibility' for startups against future powerful models and large enterprises, and the ethical dilemma of pursuing world-changing impact versus merely making money. The talk encourages deep, critical thinking to navigate the unprecedented changes brought by AI.

This podcast features a conversation with Wu Minghui, the founder of MiningLamp Technology, who provides a detailed review of the company's 19-year long entrepreneurial history, from the initial AdMaster to MiningLamp Technology, which is preparing for its IPO. The interview delves into AI technology, especially the application prospects and challenges of Agentic Models in enterprise-level services, emphasizing the importance of establishing a data defensibility strategy with proprietary data. Wu Minghui shares his experiences in multiple transformations, M&A decisions, financing difficulties, and his growth from a technical idealist to an astute business leader during the entrepreneurial process. The discussion also covers the restructuring of operational dynamics in the AI era, the future model of human-machine collaboration, and how to use AI to improve efficiency and create value in a complex business environment. The overall content demonstrates the broad potential of AI in enterprise-level application fields and the deep thinking involved in its actual implementation.

The author shares their experience during the National Day holiday of discovering the most efficient way to learn in the age of AI through in-depth debates with AI (Gemini). The author first presented their view on 'filter failure' caused by the information explosion in the AI era and, based on this, engaged in a 'relentless debate aimed at rigorously self-critiquing' with AI. AI sharply refuted the author's three inferences (imbalance in information production efficiency, constant total attention span, and excessively high cost of content discernment), pointing out that AI itself is also a powerful information filtering tool that can improve attention efficiency and may replace interpersonal trust through technology. The author then countered AI's refutation, raising new issues such as 'the cost of selecting filtering tools' and 'the selection cost brought about by AI summaries,' and emphasized the importance of source reputation in high-risk areas. Ultimately, the debate between the two parties was elevated to a philosophical level of 'recommendation engines based on taste profiles,' 'decentralized trust,' and 'Bards.' The author concludes that the win, loss, and conclusion of the debate are not the most important aspects, but rather the pure, emotionally unbiased intellectual exchange with AI, directly confronting one's own ignorance and prejudice, thereby refining thoughts, reshaping cognition, and achieving deeper self-learning and growth.

This article provides a comprehensive digest of recent AI developments, starting with Andrew Ng's new 'Agentic AI' course, which highlights critical design patterns and best practices for building effective, evaluable AI agents. It then details Anthropic's Claude Sonnet 4.5, showcasing its advanced performance in coding and reasoning, and an enhanced Claude Code SDK designed to boost workplace productivity and address business AI ROI concerns. The article also covers OpenAI and Meta's strategic diversification into consumer-facing AI products, including social video apps (Sora 2, Vibes), personalized briefings (ChatGPT Pulse), and in-chat shopping (Instant Checkout), signaling a move towards broader engagement and new revenue models. Alibaba's expanded Qwen3 family is introduced, featuring the large Qwen3-Max and powerful open-weights multimodal models (Qwen3-VL, Qwen3-Omni) that excel in understanding various media, making advanced agentic applications more accessible to developers. Finally, Sakana AI's innovative Text-to-LoRA research is presented, offering a streamlined, cost-effective method for generating task-specific LoRA adapters from natural language descriptions, enhancing model adaptability for evolving requirements.