BestBlogs.dev Highlights Issue #51

The article reports on the release of OpenAI's latest reasoning model, o3-pro. The model is now available to Pro and Team users. Benchmark test results show significant performance improvements and higher reliability compared to o3 in areas like science, education, and programming, demonstrating particular strength in math, science, and coding tasks. The article details o3-pro's relatively high API pricing and the simultaneously reduced price of the o3 model, now significantly lower. Technical details mentioned include support for text and image input and a 200k context window. Early user feedback on o3-pro is also cited, showing mixed reactions. The main body of the article features a full translation of OpenAI CEO Sam Altman's blog post titled "The Gentle Singularity." In it, he argues that AI has passed a critical turning point and digital superintelligence is progressively being realized. He highlights AI's immense potential for boosting productivity and scientific advancement, while also addressing technical and societal challenges such as coordination problems and achieving widespread availability.

The article reports on a series of significant releases by Volc Engine at the Force 2025 Conference, including the upgraded Doubao Large Model 1.6 (supporting 256K context window, multimodal understanding, GUI operations), the official version of the video generation model Seedance 1.0 Pro (with performance comparable to the industry forefront), and the AI cloud-native platform and infrastructure suites such as AgentKit, TrainingKit, and ServingKit. Through practical tests across multiple dimensions including programming, inference, multimodal understanding, and video generation, the article verifies the significant capability enhancements of the Doubao 1.6 series models and Seedance 1.0 Pro. It also discusses Volc Engine's strategic layout and technical investment in the 'AI Cloud-Native' and 'Agent' directions, concluding that Volc Engine is progressing relatively fast in the practical application of AI.

The article provides an in-depth first-hand review of ByteDance's newly released video generation model, Seedance 1.0 Pro. The author, present at the Volcengine launch event, used the internal beta version of the Jimeng AI Platform to detail Seedance 1.0 Pro's excellent performance across multiple dimensions including multi-shot combination, motion quality, delicate emotional performance, diverse camerawork, realistic physical dynamic effects, and stylization, supported by numerous generated example GIFs. The review results show that the model has reached the industry's first-tier level in aspects such as semantic understanding, motion fluidity, emotional expression, camerawork stability, and simulation of physical laws. Its performance is particularly outstanding in sports, expressions and emotions, and style consistency, earning it the reputation of an "all-rounder" in the current market. The article concludes by mentioning that the model's Enterprise user API is available and it has been launched on the Doubao App, expressing awe at the fierce competition in the AI video field.

Tsinghua University and ModelBest have jointly open-sourced the MiniCPM 4 series models (8B and 0.5B). With only about 22% of the training cost compared to open-source models of similar size, they have achieved optimal performance in multiple benchmarks among models of comparable size, some even surpassing models with a larger parameter count. The article details the four major technical innovations behind this: Efficient Sparse Attention InfLLM v2 , used to solve the problem of excessively high computation and storage overhead in traditional self-attention mechanisms for long-text processing; Edge-side Inference Optimization Framework CPM.cu and Cross-platform Deployment System ArkInfer , aimed at solving the deployment challenges caused by computation and storage limitations on edge devices and chip fragmentation; High Knowledge Density Data Filtering and Training Efficiency Optimization Algorithms like ModelTunnel v2 , significantly enhancing model capability density and reducing training costs; and BitCPM4 Ultra-low Bit Quantization Method , enabling efficient operation of models on resource-constrained devices. Test data shows that MiniCPM4 achieves significant acceleration in long-text processing on typical edge chips like Jetson AGX Orin and RTX 4090, and demonstrates strong competitiveness and practical value in actual applications such as long sequence understanding, survey generation, and tool calling.

This article highlights the importance of Text Embedding and Reranking in NLP (Natural Language Processing) / IR (Information Retrieval). It also addresses the challenges of existing methods in scalability, contextual understanding, and task alignment. The Qwen3 series models are built on the powerful Qwen3 base model and adopt an innovative multi-stage Model Training process, especially utilizing the Qwen3 LLM to synthesize Large-scale high-quality training data, and combining Model Merging strategies to improve robustness. Experimental results show that the Qwen3 Embedding series achieves SOTA (State-of-the-art) performance in MTEB (Massive Text Embedding Benchmark) multilingual and code Benchmark tests, surpassing Gemini-Embedding by X%. The Qwen3 Reranker model also demonstrates excellent Reranking capabilities. The article details its innovations, comparison with previous work, and experimental results, and emphasizes the contribution of Open-source Model to the community.

Meta released its 1.2 billion parameter world model, V-JEPA 2, trained on video. Based on the JEPA architecture, the model learns to understand and predict the behavior of the physical world through over 1 million hours of self-supervised pre-training on video and images. Subsequently, it was trained with a small amount of robot data using action-conditional training, enabling it to perform Zero-Shot planning and robot control in new environments. The article introduces V-JEPA 2's leading performance on tasks such as action prediction and video Q&A, and demonstrates its application in short-term and long-term robotics tasks like grasping and object placement. In addition, Meta simultaneously released three new physical understanding benchmark tests: IntPhys 2, MVPBench, and CausalVQA, aiming to more rigorously evaluate the model's ability to understand the causality of the physical world, and pointing out that current top models still have a significant gap compared to human performance. The paper, code, and models have been simultaneously open-sourced.

This article is an interview with Xin Huajian, the lead author of DeepSeek-Prover, delving into the breakthroughs achieved by AI, particularly Large Language Models (LLMs), in the field of formal mathematical proof. Xin Huajian elaborates on the significance and challenges of formal mathematics, and its connections to program verification and broader reasoning tasks. He emphasizes that AI's progress on complex mathematical proofs, such as the DeepSeek-Prover series, demonstrates the role of Reinforcement Learning (RL) and Chain-of-Thought (CoT) in enhancing AI's reasoning capabilities. A core viewpoint of the interview is that formal mathematics is an ideal environment for exploring and developing AI Agents and AGI, and introduces the concepts of Proof Engineering Agents and Certified AI. The article also discusses the importance of data synthesis, Test Time Scaling, and Evaluation Benchmarks, positing that high-quality evaluation standards are the crucial 'nails' driving model capability growth.

This article, based on a podcast interview with Anthropic's Emmanuel Amiesen, delves into the technical field of Mechanical Interpretability (Mech Interp) for large language models (LLMs), also touching upon personal career journeys into research and the history of the field. It highlights the challenges unique to LLMs compared to vision models, particularly the concept of superposition where models pack more features than dimensions. The discussion covers key Mech Interp concepts like sparse autoencoders, which aim to automatically extract independent features represented as directions in the model's residual stream. Amiesen introduces Anthropic's 'Circuit Tracing' work and associated open-source tooling (with Neuronpedia) as practical methods for revealing computational graphs and understanding specific model behaviors and reasoning pathways. The piece emphasizes the growing accessibility of this research field.

The article deeply analyzes the engineering challenges and capability requirements for building Agent systems in the Large Model era. The author first maps the technological layers of the Large Model era to the Java era, redefining Agent, Tool, Prompt, data, fine-tuned models, and evaluation sets as core business assets. Following this, drawing lessons from microservice architecture evolution, the article predicts that Agent systems will also evolve from monolithic to multi-agent collaboration, and discusses task allocation, collaboration patterns, and conflict resolution mechanisms. Regarding Agent collaboration and Tool calling, the author analyzes the pros and cons of the MCP protocol in detail and suggests that engineering aspects need to supplement capabilities such as public/private network access, user-level access control, fast tool integration, and optimization for long tool lists. Finally, the article points out that Agent frameworks require supporting capabilities like problem understanding, memory recall, knowledge bases, and evaluation, and based on this, envisions the core modules of an Agent platform, offering directions for thinking about the deep integration of the Agent paradigm.

The article progressively elaborates on the evolution path of AI application architecture, from initial direct user interaction with Large Language Models to the gradual introduction of key enhancement layers. First, it emphasizes the importance of context augmentation (such as RAG) to address limitations related to the timeliness of model knowledge and domain limitations. Next, it discusses the importance of input/output guardrails for user privacy and system security, and lists common types of prompt attacks and defense strategies. The article further introduces the design of intent routing and model gateways to support multi-functional applications and unify the management of heterogeneous underlying models. Subsequently, it explores the role of caching mechanisms in improving performance and reducing costs. Finally, the architecture evolves to the Agent pattern, which has planning and external interaction capabilities. The article also analyzes AI application observability metrics and methods to optimize inference performance through batching, parallelism, and other techniques.

This article reviews the evolution process of LLM applications from pure conversation to Workflow orchestration and then to Agents from an engineer's perspective. It focuses on explaining the three core components of an Agent: Memory, Planning, and Tools. It details the two planning paradigms of Agents (Decomposition-first and Interleaved Decomposition) and the classification of memory (short-term and long-term). Taking the browser-use project as an example, the author dissects its engineering architecture, including components like Agent Core, MessageManager, Memory, LLM Interface, Controller, and BrowserContext, and their interaction flow. The article specifically highlights the role of SystemPrompt, AgentMessagePrompt, PlannerPrompt, and toolPrompt in Agent execution and analyzes how browser-use ensures structured output through SystemPrompt, example guiding, and Pydantic. Finally, the article discusses the memory management implementation in browser-use and provides suggestions for persistent storage in a production environment.

The article explores how Browserbase became a key AI tool as 'the web browser infrastructure built for AI,' based on an interview with founder Paul Klein IV. It addresses the challenges of deploying, stabilizing, and maintaining headless browsers at scale, allowing AI Agents to interact with web pages like humans. Browserbase provides easy-to-use APIs and SDKs, supports multiple languages and mainstream frameworks, and innovatively launched the Stagehand Framework, enabling LLMs to control browsers using natural language. The article details the differences between Browserbase and competitors such as Browser Use, Zeta Labs, Induced AI, Island, and Arc, emphasizing its advantages as a universal, reliable underlying platform for developers. Typical use cases include data scraping, UI testing, RPA, and AI Agents.

This article benchmarks three common multi-agent architectures (Single Agent, Swarm, Supervisor) using a modified Tau-bench dataset with added distractor domains. It explores motivations for multi-agent systems, including scalability, modularity, and integrating agents from different teams. Experiments using gpt-4o reveal that Single Agent performance degrades significantly with increasing context (distractors), while Swarm and Supervisor are more stable. The Swarm architecture generally outperforms the Supervisor in score and token cost due to the Supervisor's 'translation' layer. However, the authors detail specific improvements to their LangGraph supervisor implementation, such as removing handoff messages and forwarding messages, that dramatically increase its performance, narrowing the gap with Swarm. The article concludes that generic multi-agent architectures will become more prevalent and that careful design, especially in handling inter-agent communication, is key to improving performance and scalability.

This article is a record of Sequoia Capital's interview with the OpenAI Codex team, delving into the newly launched Codex Agent. This Agent is distinguished from code completion tools, aiming to become an AI assistant capable of accepting task delegation and autonomously completing the programming process asynchronously. The interview reveals Codex's technical concept, particularly how reinforcement learning is used to better align the model with the 'tastes and preferences' of professional software engineers, thereby generating directly mergable code. The article emphasizes that effectively using the Codex Agent requires developers to adopt an 'abundance mindset' and attempt parallel delegation of multiple tasks. The article posits that AI programming will not reduce the number of developers but rather act as an efficiency multiplier, shifting developers from coding to focusing more on high-level planning, design, review, and verification. The interview also discusses the challenges of creating realistic training environments for Agents, how to handle long-duration tasks, and OpenAI's vision for future general-purpose Agents and human-computer interaction methods (combining synchronous and asynchronous, even a feed pattern similar to TikTok). Finally, it underscores the importance of writing 'Agent-friendly' code, such as using typed languages, small modules, good tests, and clear naming.

Addressing configuration and getting started issues they might encounter when initially using the VS Code-based Cursor IDE, this article provides a detailed guide for Java developers. Beginning with Cursor's download and installation, it recommends a list of necessary plugins for Java development. Next, it elaborates on methods for optimizing core configuration files like settings.json and launch.json, aiming to solve common pain points such as slow startup and improve debugging efficiency. Furthermore, the article covers practical tips including common keyboard shortcuts, Git operations, and cache clearing, and introduces advanced capabilities like Cursor AI rule customization and the MCP toolchain. Finally, it emphasizes Cursor's potential as a development tool for the AI era in assisting with business analysis, integration and innovation, and development efficiency improvement.

Thanks to the rapid development of Generative AI (GenAI), the field of software development has seen the emergence of the GenAI Application Engineer role. Although the specific responsibilities are still being clarified, they can build powerful applications at unprecedented speed. This article delves into the two core capabilities of this role: flexibly applying various AI building blocks (such as prompt engineering, RAG, vector databases, model fine-tuning, etc.) to quickly build applications, and skillfully using AI-assisted coding tools to improve development efficiency. The article emphasizes that excellent product thinking and design intuition are significant advantages. Furthermore, the author shares methods for identifying excellent GenAI engineers during interviews, specifically highlighting the criticality of continuous learning ability, suggesting assessment methods such as staying updated via professional subscriptions, gaining hands-on experience, and engaging in community interaction. The article provides clear guidance for developers to understand role requirements and for recruiters to evaluate candidates.

This article provides a detailed review of JiMeng Intelligent Reference 3.0's new feature of generating posters based on reference images. By uploading images, highly consistent AI poster recreation and design can be achieved. The article introduces how to use the feature, showcases a large number of poster cases in different styles such as commercial promotions and event exhibitions, highlights its outstanding performance in maintaining consistency in composition, style, and subject, and shares prompt tips and local re-drawing methods to improve generation effects. A structural formula is also included. The author also points out the minor issues of the current version in English details and local fine-tuning. Finally, the article discusses the impact and opportunities of AI technology on the design industry, encouraging designers to actively embrace and learn AI tools to improve work efficiency and creative space.

The article focuses on introducing Google's AI experimental platform, Google Labs, which serves as Google's proving ground for exploring and incubating frontier AI applications. The author selected and thoroughly experienced over ten experimental tools based on generative AI, including Image Generation (Whisk), Art and Music Integration (National Gallery Mixtape), Creative Recipe Customization (Food Mood), Personalized Chess Generation (Gen Chess), Font Design (Gen Type), Virtual Tour Guide (Talking Tours), Career Path Exploration (Career Dreamer), Conversational Learning Assistant (Learn About), Text-to-Podcast (Illuminate), and UI Interface Design (Stitch). Through rich visual and animated examples, the article vividly showcases the unique features and potential value of these applications, reflecting Google's broad deployment and innovative attempts in the AI application layer, and providing technical professionals with a window to understand and experience the latest AI experimental tools.

This article provides an in-depth analysis of Flux Kontext, the latest AI image generation model launched by Black Forest Labs. It introduces its innovative architecture based on Flow Matching and focuses on practical tests of the model's powerful editing capabilities in areas such as character consistency, local refinement, style transfer, and image text recognition/replacement, verifying its breakthroughs in handling the key challenges of AI-generated images. At the same time, the article provides detailed tutorials and pricing strategies for using Flux Kontext's online platform (Flux Playground, Fal, Lib AI (哩布AI)) and local deployment (based on ComfyUI API calls). Finally, through commercial application cases of IP Design and e-commerce product images, the practical value and efficiency of the model are demonstrated. The article points out its stable performance in prompt following but also mentions potential limitations in processing Asian portraits and Chinese watermarks. Overall, Flux Kontext is a powerful and practical image editing tool.

The article provides a detailed review of 360's newly released Nano AI Super Search Agent. It emphasizes that this product has surpassed traditional search engines, becoming an AI Agent focused on task completion and result delivery. Through practical cases (such as shopping price comparison, CSL public opinion analysis, Gaokao college application), the article demonstrates Nano AI's powerful capabilities in obtaining and integrating web-wide information (including social media, e-commerce platform reviews, etc.), dynamically planning search tasks, performing multimodal output (such as visual reports, interactive webpages), and directly executing specific actions (such as adding to cart). The article posits that the combination of this strong search capability with Agent significantly enhances AI's practicality and "deliverability," signaling the deconstruction of the traditional search paradigm and a shift in AI interaction models, which can effectively address issues of information asymmetry and inefficiency.

This article, based on insights from three partners at the Sequoia US AI Ascent Summit, explores the market structure, product evolution, technology path, and long-term cognitive shifts in the AI era. Key takeaways include: the AI market is far larger than anticipated, reshaping both software and services; the application layer is crucial for building high-value enterprises, requiring user-centric barriers; Agents are evolving from standalone tools to collaborative Agent economies, with technical challenges in persistent identity, communication protocols, and trust; AI is driving labor costs down, making taste and value propositions more critical than mere functionality; the AI era demands a shift from deterministic to probabilistic thinking, managing intelligent Agents in a high-leverage, high-uncertainty environment. The article emphasizes the imperative for data flywheels, deep moats, and the urgency for speed and proactive engagement.

This article is compiled from a speech given by Meng Xu, Head of Hardware Products at NetEase Youdao, at the AICon conference, exploring the transformative impact of AI large language models (LLMs) on intelligent learning hardware. The core view is that the evolution of learning hardware is the result of a virtuous cycle involving user needs, hardware innovation, and AI technology, with particular emphasis on the importance of software-hardware integration in the era of large language models to solve real user problems. Taking the Youdao AI Answer Pen as an example, the article elaborates on the breakthroughs in LLM applications in language interaction (specifically mentioning the industry's first implementation of an on-device offline large language model to solve problems in no-network scenarios), multi-subject Q&A tutoring, and multimodal interaction, enabling it to provide personalized, detailed explanations akin to a real teacher. It also looks ahead to the future of deep integration between AI Agents and the education ecosystem, believing this will truly make learning hardware an exclusive AI partner for children, achieving the ultimate goal of personalized learning.

This article introduces CAIR (Confidence in AI Results), a psychological metric crucial for AI product adoption. It argues that unlike model accuracy, CAIR is primarily controlled by product design decisions and can be measured as Value / (Risk * Correction). The authors analyze successful AI products like Cursor and contrast them with scenarios like Monday.com's AI, demonstrating how design choices impact CAIR. High-stakes domains like finance and healthcare highlight the need to design around AI's inherent limitations, particularly with numerical reasoning, rather than waiting for perfect models. Five principles for optimizing CAIR are presented: strategic human-in-the-loop, reversibility, consequence isolation, transparency, and control gradients. The article concludes by reframing AI readiness assessments to include CAIR, emphasizing that designing for user confidence is key to widespread adoption.

This episode of 'Late Post Chat' features Xuzhang Xie, co-founder of iVerse Technology, providing an in-depth look at the entrepreneurial journey, product strategy, and growth secrets behind their AI video generation company. Founded in April 2023, iVerse Technology specializes in AI video generation technology. Its core product, PixVerse, has achieved significant user adoption through continuous model iteration (now in its seventh generation, V4.5) and innovative product features, particularly template-based creation designed for regular users. Xie Xuzhang discusses his transition from investor, forming a synergistic team with technical partners, and the experience of training models with a resource-efficient strategy under relatively limited resources. PixVerse rapidly accumulated a significant overseas user base through features such as the 'Venom Template,' with monthly active users (MAU) approaching 20 million and monthly subscription revenue exceeding 10 million RMB. The podcast also explores key aspects of the AI video landscape: technical challenges like multi-modal fusion, the competitive environment with both domestic and international players, business models encompassing subscription and API strategies, and future goals of reaching hundreds of millions of users and achieving scaled profitability. Finally, Xie Xuzhang announces the upcoming official launch of the domestic version 'PaiWo AI' and shares his perspectives on AI startup trends and talent.

This interview records a deep conversation between Lex Fridman and Google CEO Sundar Pichai. Pichai reviewed how Google overcame doubts and caught up in the AI race over the past year, emphasizing AI's position as the company's core strategy, and key decisions like merging DeepMind and Google Brain. The interview detailed how Google Search will deeply integrate AI in the future (AI mode, AI Overviews), providing rich context while still guiding users to discover valuable web content. Furthermore, Pichai discussed the effectiveness of Scaling Law, the limitations of compute on model deployment, and the significant improvement in Google's internal engineering efficiency (about 10%) due to AI. He also proposed that AR is the next important human-computer interaction paradigm after command line, GUI, touch, and voice, considering AI essential for achieving a seamless AR experience. Regarding AGI, Pichai believes the current stage is 'Asymmetrical AI' (AJI), predicting that full AGI will still be difficult to achieve by 2030, but emphasized that AI, as a self-improving technology, will have a long-term impact far exceeding historical inventions like electricity, and expressed optimism about humanity's ability to manage potential risks.

This article is the transcript and summary of the latest interview between the YC President and Cursor CEO Michael Truell. Truell elaborates on Cursor's vision to go beyond existing AI coding assistants, ultimately replacing traditional coding methods entirely to achieve intent-based software construction. He emphasizes that as AI takes over implementation details, the core value of future software engineers will lie in high-level 'taste'—the judgment regarding product logic and direction. The article deeply analyzes the key strategic significance of Cursor's early decision to build a standalone editor instead of a VS Code extension, believing it was necessary to fully control the user interface to adapt to future interaction paradigms. Furthermore, Truell points out that in the AI era, the true moat lies in acquiring data through large-scale user acquisition and forming a 'data flywheel' that continuously optimizes products and models, while also emphasizing strategic planning that aligns with the trend of continuously strengthening AI model capabilities. The interview also recounts Cursor's entrepreneurial journey from a CAD project to the AI programming field.

Sam Altman's article elaborates on his unique perspective on the technological singularity, believing it is not a sudden transformation depicted in science fiction, but is occurring quietly through continuous technological advancement and everyday integration. He points out that AI has already surpassed humans in multiple areas, and this process is accelerating, with core driving forces including using AI to create stronger AI, capital investment attracted by the immense economic value generated by AI, and the future physical flywheel of robots building robots. He looks forward to the 2030s, believing that although human nature will remain unchanged, the liberation of intelligence and energy will greatly enhance individual capabilities and redefine work. The article emphasizes that to ensure a bright future, the AI alignment problem must be solved, ensuring its goals are aligned with human well-being, and promoting the widespread accessibility and benefit of superintelligence, avoiding monopolization. Ultimately, he believes this will be an era of 'visionaries,' with lower technical barriers, and good ideas will be the key.

This article provides an in-depth interpretation of Apple's WWDC25 developer conference. The core of the conference revolved around the new Liquid Glass design language and deeply integrated Apple Intelligence, alongside updates to the full range of operating systems including iOS, macOS, iPadOS, watchOS, tvOS, and visionOS. The author believes Apple abandoned ambitious large model OS plans in favor of a pragmatic strategy, integrating AI features into the details of daily applications, such as on-device real-time translation, smart call identification, and enhanced visual intelligence. The article details the new features of each system, especially the significant AI upgrades to macOS Shortcuts and Spotlight, seen as the prototype of tool-based Agents. Simultaneously, the article also discusses the purpose of the Liquid Glass design (to divert attention) and features like Workout Buddy in watchOS which are seen as features for AI's sake. Overall, the conference is evaluated as lacking revolutionary breakthroughs but highlights pragmatic incremental innovation which is significant for user experience and AI implementation, and affirms Apple's open and cooperative stance in the AI field.

This episode features an in-depth interview with Chen Mian, AI application entrepreneur and founder of Lovart. He reflects on his years of experience at leading mobile internet companies (Tencent, 360, Baidu, Didi, Mobike, Meituan, ByteDance), as well as his reflections on business models, product development methodologies, and career decisions. Chen Mian believes AI is a more profound revolution than the mobile internet, bringing tremendous opportunities for application entrepreneurship. He details Lovart's entrepreneurial journey in the design vertical domain, including how to choose a multimodal approach to avoid the large model main track, build differentiated capabilities, and the struggles and responses when facing challenges such as subsidy wars, product removal, and broken capital chains in the early stages. The podcast discusses the company's process from near-death to recovery through fundraising and product iteration (from professional tools to more inclusive AI agents). Chen Mian shares lessons learned in customer acquisition, cash flow management, and team building in a competitive market, and offers unique insights into the future of the intelligent agent ecosystem, the relationship between general AI and vertical AI agents, the evolution of business models, and changes in team organizational structure. He emphasizes that entrepreneurship requires a high degree of vigilance to maintain sharp decision-making. The sense of accomplishment far outweighs monetary rewards and job titles.

The article features an in-depth interview with WaveSpeedAI founder Cheng Zetong, exploring his journey and practices from being a key technical member at a large company to becoming an AI Infra entrepreneur. After encountering a growth bottleneck at a large company, Cheng Zetong validated his technical value through open source and discovered the general undervaluation of AI Infra in the domestic market. He believes that inference acceleration is key to determining AI application performance and commercialization. WaveSpeedAI adopts a 'light company, heavy system' model, building a small and efficient remote team focused on providing stable, efficient, and low-cost inference infrastructure services for global AI content platforms. By collaborating with compute providers and model teams, and employing a strategy of integrating closely with client systems, WaveSpeedAI quickly achieved revenue after product launch and became profitable within months, proving that AI Infra is a viable business. The article highlights WaveSpeedAI's cost optimization solutions and success cases (like Freepik) in the AI video generation domain, showcasing its vision of empowering global creators through technical advantages.

This episode of Silicon Valley 101 features seven experienced users of AI Agents from diverse fields, analyzing them from user, builder, business, and philosophical perspectives. Users shared their experiences with Agents improving efficiency and aiding creativity, alongside concerns about instruction failures and impersonal experiences. Builders addressed technical difficulties such as understanding complex instructions, processing unstructured information, and multi-agent collaboration challenges, emphasizing the importance of user feedback and scenario focus. On the business front, they discussed how startups can leverage novel data and build user insights to compete with large language model companies, highlighting the value of vertical applications and customized solutions. Finally, the guests explored the potential impact of AI Agents on future human-machine relationships, the role of human values, social structures, and the development of an AI-friendly environment. The program offers a comprehensive and insightful view of AI Agents.

The article deeply explores the evolution of productivity units in the age of Artificial General Intelligence. From Horsepower in the Industrial Revolution and Person-day in the Knowledge Economy, it progresses to Token and MTH (Megatoken-hour) in the AI era. It elucidates the importance of Token as the fundamental AI resource unit and proposes three key metrics for measuring an intelligent society: Capacity (MT/GT/TT), Speed (T/s, GT/s, TT/s), and Price/Energy Consumption (¥/MT, kWh/MT). The article points out that current AI compute infrastructure is still in its '2G era' concerning speed and accessibility, and achieving high-speed Token flow requires massive hardware and energy investment. Ultimately, the article foresees the reconfiguration of human labor value and the establishment of a new economic system and governance framework based on Tokens at the individual, enterprise, market, and even national levels, emphasizing that standardized units of measurement are crucial for driving the leap in intelligent civilization.