BestBlogs.dev Highlights Issue #35

The article reports on xAI's latest flagship Large Language Model, Grok-3. The Grok-3 series includes a lightweight version, Grok 3 mini, emphasizing rapid response. Grok-3 significantly outperforms models such as Gemini-2 Pro, DeepSeek-V3, Claude 3.5 Sonnet, and GPT-4o in multiple benchmarks including Math, Science, and Coding, and tops the Chatbot Arena. Grok-3 also features powerful agent capabilities, including DeepSearch comparable to OpenAI's Deep Research. DeepSearch enables in-depth research, brainstorming, data analysis, image generation, and code development. Additionally, the article introduces Grok-3's subscription and pricing information, as well as xAI's open-source principles. Finally, Musk implies that xAI will surpass OpenAI in technological competition, demonstrating confidence in its own technological competitiveness.

This article introduces Google's PaliGemma 2 mix, a new family of fine-tuned vision language models based on PaliGemma 2. These models excel at a variety of tasks, including OCR, captioning, visual question answering, and object detection/segmentation. Available in different sizes and resolutions, PaliGemma 2 mix models are designed to be fine-tuned for specific downstream tasks, showcasing strong performance. The article provides code examples, a demo, and links to further resources for utilizing these powerful, open models.

Stepchaser, partnering with Geely Automobile Group, has released two cutting-edge open-source multimodal large models: Step-Video-T2V and Step-Audio. Step-Video-T2V is the world's largest and best-performing open-source video generation model, using the MIT License and supporting free commercial use. This model uses a Deep Compression Variational Autoencoder (Video-VAE), achieving a 16×16 spatial compression ratio, and a DiT with a 3D Full Attention Mechanism for denoising input noise into potential frames. Step-Audio overcomes the limitations of traditional TTS by generating high-quality synthetic audio data. This enables iterative cycles of synthetic data generation and model training. Both models are now live on the '跃问' App. Step-Video-T2V has powerful capabilities in complex motion, character aesthetics, and visual imagination, with 30 billion parameters, and can directly generate 204 frames of 540P resolution video in a single pass. Step-Audio leads in performance on multiple mainstream public test sets, especially excelling in Chinese, integrating external tools through the ToolCall Mechanism, and possessing high emotional intelligence dialogue and role-playing capabilities. Stepchaser is committed to technology-driven development, adhering to the research and development of pre-training and foundation large models, with the goal of building AGI.

This article introduces VideoLLaMA3, an open-source model from DAMO Academy, which is a new generation of multimodal video-language model designed with an image-centric approach. The model is only 7B in size and performs excellently in three core dimensions: general video understanding, temporal reasoning, and long video understanding. Additionally, the 2B version optimized for edge deployment also performs well in image understanding, outperforming existing models in benchmarks such as InfoVQA and MathVista. The article details VideoLLaMA3's image-centric training paradigm, including four key aspects: visual encoder adaptation, visual-language alignment, multi-task fine-tuning, and video fine-tuning, as well as two major innovative technologies: Arbitrary-Resolution Visual Tokenization (AVT) and Differential Frame Pruner (DiffFP). Furthermore, the article introduces the VL3Syn7M dataset, built to construct high-quality training data, which ensures image quality and text relevance through techniques such as aspect ratio filtering and aesthetic score filtering. Finally, the article mentions the significance and value of the model being open-sourced and provides a link to the research paper and online Demo link.

The article analyzes the core viewpoint of OpenAI's latest paper, which is that training large language models (LLMs) through reinforcement learning can significantly improve their performance in complex programming and reasoning tasks. OpenAI's o3 model achieved gold medal level in the IOI 2024 competition and achieved results comparable to elite humans on CodeForces, ranking in the top 200 globally with a score of 2724 points. The article also cites the blogger's view that this strategy is not only suitable for programming but also a key path to AGI. In addition, the article mentions independent research by DeepSeek R1 and Kimi k1.5 to improve model performance through Chain of Thought (CoT) learning and discusses the application potential of general reinforcement learning in other fields, as well as contrasting perspectives on this approach.

This article details the technological innovations within the DeepSeek series models, including the fine-grained expert division and shared expert separation of the DeepSeek MoE architecture, alongside load balancing strategies designed to address load imbalance issues. It offers a deep dive into the GRPO algorithm's improvements over PPO, which reduce computational resource consumption by minimizing the Value model. It elaborates on how MLA reduces KV Cache through low-rank decomposition, thereby lowering inference costs. Furthermore, it introduces how MTP enhances training efficiency and inference speed by predicting multiple tokens at once. The article particularly emphasizes the groundbreaking significance of R1-Zero, trained entirely via reinforcement learning to create reasoning models, and the advantages of DeepSeek V3 in terms of training efficiency and cost control. It also highlights how R1, building upon R1-Zero, resolves readability and language mixing issues through multi-stage training strategies, facilitating real-world deployment. Overall, the article offers a thorough and insightful analysis of DeepSeek's core technologies.

This article provides an in-depth analysis of the AI Coding track, outlining numerous AI Coding companies at home and abroad through an industry landscape overview, and selecting 12 well-known products for a side-by-side comparison, including GitHub Copilot, Cursor, and Windsurf, across IDE plugins, web-based IDEs, AI-native IDEs, and pure models. The article designed two test cases: a Snake game and an automatically generated resume website, comparing and analyzing them based on accuracy in fulfilling requirements, design diversity, error handling capabilities, and context understanding. The evaluation results show that Windsurf excels in fulfilling user requirements and design diversity, with Cursor and o3-mini-high also performing well. AI Coding products assist developers by automatically generating, reviewing, and optimizing code, improving software development efficiency, reducing human errors, and ensuring code consistency and quality. The article also discusses the potential changes in production relationships brought about by AI Coding, as well as new opportunities and models that may emerge in content distribution, foreshadowing a shift in software development models.

This article introduces the LangMem SDK, a library designed to help AI agents learn and improve through long-term memory. It provides tools for extracting information from conversations, optimizing agent behavior via prompt updates, and maintaining long-term memory about behaviors, facts, and events. The SDK supports various memory types (semantic, procedural, episodic) and integrates with LangGraph, enabling developers to build more personalized and adaptive AI experiences. They are also launching a related managed service.

This article introduces the innovative project DeepClaude, which is developed in Rust and integrates seamlessly the CoT logical reasoning ability of DeepSeek R1 with the text generation ability of Claude, forming a unified LLM reasoning API. Experimental results show that DeepClaude performs excellently in code editing benchmarks, even surpassing the separate use of o1-high and R1 models. The project is 100% free and open source, and has received 3k stars on GitHub, especially suitable for code security checks. The authors of DeepClaude believe that this AI agent combination represents a "digital world first" paradigm shift, transforming intelligent systems into proactive collaborators. Some netizens have further developed a triple-combination approach, combining the thinking results of DeepSeek-R1 and Gemini 2.0 Flash with Claude Sonnet, achieving better results in specific tests.

This article details the advantages of dedicated vector databases over general-purpose databases for vector search tasks. It highlights the challenges posed by high-dimensional vector data and storage requirements. The article emphasizes the characteristics of vectors as data transformations and the resulting data update and index maintenance issues. It also discusses the architectural trade-offs in vector databases, such as the choice between ACID and BASE principles, and how a BASE architecture enables high availability and scalability. Furthermore, the article explores the complexity of vector indexes, especially the implementation and optimization of the HNSW index, and how segmentation and filterable indexes enhance performance. Finally, it underscores the potential of vector search in areas like discovery and recommendations, extending beyond traditional RAG applications.

This article details how the author uses AI tools, especially the Dify platform, to build a multi-round review and refinement Workflow, aimed at solving the issue of stilted or unnatural translation commonly encountered when translating English technical articles into Chinese. The Workflow includes key stages such as content crawling, preliminary rewriting, multi-round review focusing on language fluency, content accuracy, and style consistency, comprehensive improvement, and final refinement. The article delves into the design ideas of Prompt Engineering and shares experiences in model selection, such as using Google Gemini 2.0 Flash for main rewriting and refinement tasks, and Qwen-max-latest and OpenAI o3-mini for language and content review, respectively. Experimental results show that the Workflow can effectively improve translation quality, resulting in translations that are fluent, natural, and aligned with Chinese reading habits. In addition, the author also shares the subsequent typesetting, cover generation, and publishing process, emphasizing the great potential of AI in the field of content creation and the effectiveness of combining different AI tools in a workflow.

The article delves into OpenAI's newly launched Deep Research feature. This feature uses AI to conduct multi-step research on the internet, aiming to reduce information integration costs and quickly generate research reports. The author's use cases demonstrate Deep Research's value in areas like interview preparation and medical information search. It also points out its limitations, including its inability to access non-public information, the potential for misleading information, and its reliance on information quality. The author believes that Deep Research has advantages in information integration speed and cost but relies on public information. The article further analyzes the impact of Deep Research on knowledge value, information confidentiality, and the future information ecology, arguing that AI tools may exacerbate information overload while also becoming a key means of filtering effective information. The article predicts that information confidentiality will be more valuable in the future.

Perplexity introduces Deep Research, a newly launched free feature. It provides users with in-depth research reports by searching massive amounts of data and offering expert-level analysis. Deep Research performed well in the Humanity's Last Exam and SimpleQA benchmarks, and it is also fast, completing research tasks in an average of 3 minutes. Perplexity CEO publicly thanked DeepSeek for being open source, cost-effective, and fast. The article also showcases Deep Research's application cases in finance, marketing, technology, and other fields, and compares its differences with ordinary search functions. The article also mentions netizens' questions regarding the naming of Perplexity's new feature, "Deep Research." It further covers the CEO's response to Perplexity's advantages and previews of future features.

The article delves into the significant impact of the DeepSeek R1 model in the AI field after it was open-sourced. The model, with its excellent performance, especially in writing style and logical reasoning, and its open-source strategy, challenged the industry's traditional perceptions. Experts analyzed the technological innovations of DeepSeek R1, such as the application of the GRPO algorithm in synthesizing high-quality data, and its impact on reducing the threshold of AI applications and accelerating the design of AI-native products. In addition, the article also discussed the role of the open-source model in promoting the AI ecosystem, as well as domestic teams' strategies to optimize software for training efficiency amid computing power constraints. The article also looks forward to the application prospects of AI in the entertainment field, especially in video generation, and explores the future evolution direction of product forms such as chatbots, as well as the importance of AI-native product managers.

This article interviews Wang Xingxing, founder of Unitree Robotics, detailing his transformation from a Tech Geek to an industry rising star. Wang Xingxing, not a Top Student traditionally, overcame resource and education limitations through passion and focus, developing low-cost, high-performance Robot Dogs and Humanoid Robots. He shares Unitree Robotics' self-developed experience with Core Components like Motors and 3D LiDAR, plus strategies in Technology Selection, Product Iteration, and Cost Control, such as reusing Robot Dog tech and hardware on Humanoid Robots for rapid iteration and low-cost development. He emphasizes focus, rapid iteration, and continuous learning as keys to Unitree Robotics' lead. Additionally, he shares his unique learning methods and business understanding, advocating rational business practices: maximize profits even in a Niche Market and leverage prevailing trends.

The article is an interview with Academician Ya-Qin Zhang, Dean of the Institute for AI Industry Research of Tsinghua University (Tsinghua AIR), mainly discussing his vision for the future development path of AGI. Academician Ya-Qin Zhang believes that the realization of AGI will be carried out in stages, in the order of AI-powered Information Processing, AI-enhanced Robotics, and AI-driven Bioengineering. He predicts that AI-powered Information Processing will reach AGI level within 5 years, AI-enhanced Robotics such as humanoid robots will take about 10 years, and AI-driven Bioengineering will take 15 to 20 years. In the interview, Academician Ya-Qin Zhang also shared his views on fields such as Autonomous Driving and Embodied Intelligence, emphasizing the breakthrough role of Large Models in solving data, generalization, and model integration problems, thereby accelerating the development of technologies such as Autonomous Driving and Embodied Intelligence. In addition, he boldly predicts that AGI will profoundly affect human society, expand the human brain, extend lifespan, and may even create new species, and looks forward to possible changes in employment structure, education models, and other aspects.

This article is a speech draft by Yu Yi, Tencent Qing Teng AI & Globalization Project Manager, at the Tencent Tech for Social Good Innovation Festival. She shared her experience of 2000 hours of AI collaboration, demonstrating the application of AI in emotional support, decision assistance, and work efficiency improvement through multiple practical cases. She emphasized that the key to AI collaboration lies in seeing AI not just as a tool or software, but as an intelligent partner capable of understanding and simulating human behavior. At the same time, she also proposed new perspectives such as 'Embracing Imperfection' and 'AI and Human Collaboration', believing that the existing Large Language Model capabilities still have huge potential to be explored. Finally, she shared seven strategies for AI collaboration, encouraging everyone to adjust their perspective, boldly explore new models for AI collaboration, and redefine companionship and leadership.

This article delves into the concept of "Deep Research" agents, which leverage Large Language Models (LLMs) to automate complex research tasks by gathering information from multiple sources and generating in-depth reports. It analyzes Deep Research products from companies like OpenAI and Google, discussing their technical implementations, including custom-tuned models (o3, Gemini 1.5 Flash) and tool calls. The article also explores user experiences, highlighting the challenges of asynchronous UX and the benefits of editable chain of thought. Furthermore, it touches on the challenges of evaluating the quality of these Deep Research agents and their potential to accelerate knowledge work and discover new insights.

Taking a historical perspective on AI's evolution, the author explores the technical foundations and practical applications of LLMs. They express optimism about the technology's future and offer advice for practitioners navigating the AI landscape. The article elaborates on the AI wave triggered by LLMs from global, local, and personal perspectives. Subsequently, it introduces the basic knowledge of LLMs, including their definition, the difference between LLMs and small models, the scaling laws and emergent properties of LLMs, and the relationship between LLMs and AI. At the same time, it analyzes the current situation and reasons for the landscape of numerous competing models and compares the development levels of LLMs at home and abroad. In addition, the article also delves into the theoretical knowledge of LLMs, including the development history of language models, the pre-training technology of general-purpose LLMs and mainstream LLM structures, the efficient fine-tuning technology of domain-specific LLMs, the choice between RAG and Fine-tuning, RLHF for human alignment, prompt learning, and model compression. Finally, the author combines their experience of attending the AICON conference to introduce practical cases of LLMs in fields such as Search, Advertising, and Recommendation, and predicts that Generative AI and interest-based recommendation will be important breakthroughs in this field. Overall, this article is comprehensive, easy to understand, and suitable for readers interested in LLMs.

Qiji LLM Daily summarizes key AI developments on February 20, 2025. Microsoft's topological qubit quantum processor marks a quantum computing breakthrough with potential impacts on medicine and materials science. AgentLand Festival 2025 showcased AI's innovative applications in game development, offering a more engaging and interactive experience. DeepSeek significantly improved model reasoning with large-scale reinforcement learning, offering a new direction for LLM development. xAI's Grok 3 model excelled in benchmarks, sparking discussions on AI benchmark reliability. Also featured: AI CUDA Engineer for efficient CUDA kernel generation, Lingo.dev AI automating translation into pull requests, and Hyperlume's $12.5 million seed round. Recommended reading: The theory of LLMs｜Zezeyuan Zhu ICML Speech Summary.