AI & Tech

DeepSeek Applications and Strategic Thinking

• DeepSeek is now being smoothly applied in cross-border e-commerce.
• The success of "Nezha 2" and DeepSeek shares a venture capital-style mindset. In the face of complex and rapidly changing environments, enterprises and individuals should learn from the venture capital industry, actively embrace uncertainty, dare to take risks, not be afraid of failure, and be brave to innovate.

DeepSeek's Open-Source Matrix Multiplication Library: DeepGEMM

• DeepSeek has open-sourced DeepGEMM, a matrix multiplication library optimized for the Hopper architecture GPU, supporting standard matrix calculations and Mixture of Experts (MoE) calculations.
• This library provides strong support for the training and inference of DeepSeek-V3/R1.
• The core code is only about 300 lines and outperforms existing solutions in most matrix sizes.
• DeepGEMM utilizes HopperGPU's tensor cores and TMA technology to optimize FP8 precision matrix multiplication.
• In standard matrix multiplication, DeepGEMM is 1.0 to 2.7 times faster than optimized implementations based on CUTLASS3.6. In Mixture of Experts model calculations, continuous arrangement is accelerated by about 1.1 to 1.2 times, and mask arrangement is also accelerated by 1.1 to 1.2 times.

Tech Industry Updates

New Deep Learning Models Released

Several tech giants have recently launched new deep reasoning models. This surge in model development is indicative of the growing competition in the artificial intelligence sector, where firms are investing heavily in advanced technologies to enhance their offerings.

Exponential Financing and Valuation Increases

In conjunction with these new model releases, companies are experiencing excess financing, leading to a remarkable valuation increase of up to two times. The influx of capital reflects investors' confidence in the potential of AI and its applications across various industries.

Built-in JIT and Clean Implementation

The latest models feature built-in Just-In-Time (JIT) compilation capabilities, showcasing a commitment to efficiency and performance. The implementation is noted for its clean design, making it easier for developers to integrate and utilize these new technologies.

AI

AI Model Applications

The focus is on AI large model applications.

Finance

Biren Technology IPO

Biren Technology is considering a Hong Kong IPO and plans to raise $300 million.

AI & Machine Learning

Open Source Trends in AI

The trend of AI open source is emerging, with inference models becoming mainstream.

DeepSeek's Impact and Future AI Development

Redpoint Ventures partner Jacob Effron and David Luan discussed DeepSeek's implications for the large model space, suggesting its success indicates a shift towards improving model efficiency. Luan anticipates that large "teacher models" will be trained internally and then compressed for client use. The increasing complexity of AI applications, from simple chat to drug discovery, requires more intelligent models, with reinforcement learning playing a key role in enhancing agent intelligence.

AI Search and Advertising

Controversy arose around AI search tool DeepSeek due to suspected advertising within its answers, showing bias in recommending purchasing channels. This has led to discussions about the possibility of advertising integration in AI search. While technically feasible and seen as valuable by advertisers, user acceptance, content accuracy, and privacy issues remain challenges. Domestic AI search platforms are cautious about incorporating ads.

AI Research Assistants

OpenAI Expands Deep Research AI Agent

OpenAI has broadened access to its powerful Deep Research AI agent, now available to ChatGPT Plus, Team, Education, and Enterprise users. This expansion is expected to heighten competition with other players like DeepSeek and Anthropic within the swiftly advancing AI research assistant market.

DeepSeek's Open Source Success and Commercialization Challenges

DeepSeek, having achieved a successful 1.0 phase as an open-source software, has attracted major tech companies such as Tencent, Baidu, and Alibaba. However, it is facing challenges in its commercialization strategy. Despite having substantial funding, the need for a solid business model is crucial for leveraging its technological advancements effectively. This analysis explores three business models for open-source technology:

1. Google's approach of “establishing rules” to control core services like GMS.
2. MySQL's model of offering an enterprise version alongside a partially open-source product.
3. The “open source + product/service” model exemplified by Chrome and RedHat, which may serve as a primary model for DeepSeek: generating revenue through services rather than direct model sales. The focus is on converting user numbers into value through branding and service offerings.

DeepSeek Launches Open Source Week

DeepSeek has kicked off its “Open Source Week,” which highlights the importance of open-source development and its implications. This initiative could hold significant relevance for developers and organizations engaged in open-source projects, fostering collaboration and innovation within the tech community.

Artificial Intelligence

OpenAI DeepResearch Expansion

OpenAI has broadened access to its DeepResearch AI agent, now available to ChatGPT Plus, Team, Edu, and Enterprise users. Pro users receive 120 deep research queries monthly, while other users get 10. DeepResearch, leveraging a version of the o3 model, conducts multi-step research by collecting web information and synthesizing it into detailed reports. It can search and interpret texts and images, generating reports in 5-30 minutes, complete with citations and summaries of its reasoning, a process OpenAI claims would take humans hours.

AI

OpenAI DeepResearch Expansion and Features

• OpenAI is expanding its DeepResearch AI agent, initially for ChatGPT Pro, to Plus, Team, Edu, and Enterprise users. This tool enables ChatGPT to conduct in-depth research and synthesize information into comprehensive reports within 5-30 minutes, citing sources. Plus users receive 10 queries monthly, while Pro users get 120. The tool now includes embedded images with citations and improved document analysis.

OpenAI's Decision to Withhold DeepResearch API

• OpenAI will not release the DeepResearch AI model via its developer API due to concerns about AI's persuasive capabilities and potential for spreading misinformation. The company is revising its methods for assessing “real-world persuasion risks.” Though deemed unsuitable for mass disinformation campaigns due to cost and speed, OpenAI will investigate factors like personalized persuasive content, as tests showed the model could write persuasive arguments, but not better than humans. DeepResearch is powered by a version of the o3 model.

AI Development & Applications

AI Model Updates and Competition

Chinese AI startup DeepSeek has reopened its API after a three-week pause due to capacity issues. Their R1 "reasoning" model rivals OpenAI's. Meanwhile, Alibaba launched a preview of its QwQ-Max reasoning AI model, planning to open-source it, indicating growing competition in the Chinese AI market.

AI Tools in Research

The author shares their evolving perspective on AI tools, embracing them for research purposes. They use Ollama with the Llama3.2 LLM for quick answers and DeepSeekR1 with the Mysty GUI for in-depth research, significantly enhancing their research process.

Workplace Dynamics

The Importance of Workplace Relationships

Employees who feel connected and can collaborate are more likely to succeed and be happy. Social connections increase life satisfaction and well-being, while isolation can lead to burnout. Hybrid work models pose challenges to maintaining meaningful workplace relationships, requiring future solutions.

AI Model Updates

Anthropic's Claude 3.7 Sonnet

Anthropic has launched its first "hybrid model," Claude 3.7 Sonnet, integrating real-time responses with deep-thinking capabilities, allowing users to obtain various types of answers without switching models. The model excels in following instructions, general reasoning, multi-modal capabilities, and autonomous coding, with significant improvements in mathematics and science. Its coding abilities surpass DeepSeek R1 and OpenAI's o1 and o3 models.

Agent Claude Code

Anthropic has also released Agent Claude Code, designed specifically for coding tasks. It can run directly in the terminal, assisting developers with programming tasks.

Claude 3.7 Sonnet Availability and Pricing

Claude 3.7 Sonnet is now fully available, although the extended thinking mode is not accessible for free users, and its pricing is higher than competitors' pure reasoning models.

Anthropic's Funding

Anthropic is reportedly nearing the completion of a new $3.5 billion funding round, potentially valuing the company at $61.5 billion.

AI & Technology

AI in Education

DeepSeek is emerging, but ChatGPT maintains its leading position in American universities.

The Impact of AI Companions

The rise of AI girlfriends may affect the birth rate.

Value Reconfiguration

Artificial Intelligence

AI Development & Open Source Initiatives

DeepSeek has open-sourced DeepEP, the world's first full-stack communication library for MoE models, aiming to address AI computing power issues. DeepEP optimizes NVLink technology to increase data transfer speeds between GPUs and utilizes RDMA technology to reduce data transfer latency. It also features intelligent sorting and FP8 data compression to further improve data processing efficiency. DeepSeek also released FlashMLA code during "Open Source Week" to reduce costs in large model training. Through these open-source technologies, DeepSeek is helping to reduce costs for the entire industry chain.

US Restrictions on China's AI

The impact of the White House ban on the rise of Chinese AI is a key consideration.

Finance

DeepSeek Large Model Application in Banking

Banks are increasingly adopting the DeepSeek large model, but challenges remain, particularly in addressing technical hallucinations.

AI and Investment

Anthropic's Valuation Soars

Anthropic's valuation has climbed to $61.5 billion.

Claude 3.7 Sonnet Release

Anthropic has released Claude 3.7 Sonnet, featuring a hybrid reasoning mode to enhance thinking control and computational capabilities.

Reasoning Mode

The new model allows switching between two thinking modes and offers precise control over thinking time.

AI and Technology Updates

DeepSeek's DeepEP Communication Library

DeepSeek has launched DeepEP, a communication library tailored for Mixture of Experts (MoE) and Expert Parallelism (EP) systems. This library aims to boost the efficiency of large-scale AI training and inference by optimizing GPU kernels, particularly in data routing and output integration within MoE models.

Key features of DeepEP include native support for FP8 intelligent compression transmission and flexible GPU resource management, making it suitable for resource-constrained or real-time applications. By minimizing data transfer wait times and enhancing GPU utilization, DeepEP aims to reduce costs and increase efficiency in areas like natural language processing, code generation, and recommendation systems.