Group 1 - Keda Xunfei's Chairman Liu Qingfeng stated that the domestic AI infrastructure has taken initial shape, with domestic large models matching international standards despite having half the parameters [2] - Michael Burry warned that the era of tech giants earning huge profits with minimal investment is ending, primarily due to AI, and investors should focus on Return on Invested Capital (ROIC) rather than revenue growth [3] - A BlackRock survey revealed that while investors are optimistic about AI, they are shifting their focus to energy and infrastructure suppliers, with only one-fifth considering large US tech companies as attractive investment opportunities [4] Group 2 - Demis Hassabis, CEO of Google DeepMind, indicated that Chinese AI models are only a few months behind those in the US and have made significant advancements in the past year [5] - DeepSeek released a new paper on conditional memory, significantly improving model performance in various tasks, and has open-sourced a related memory module [6] - Baichuan Intelligence's CEO Wang Xiaochuan mentioned that the company has 3 billion yuan on hand and may initiate an IPO plan in 2027 [7] Group 3 - Zhiyun and Huawei have open-sourced a new image generation model, GLM-Image, which is the first SOTA multimodal model trained entirely on domestic chips [8] - Kuaishou Technology announced that Keling AI's revenue exceeded $20 million in December 2025, with an annual recurring revenue (ARR) of $240 million [9] - Yongyou Network expects a net loss of 1.3 to 1.39 billion yuan for 2025, a reduction in loss compared to the previous year [10] Group 4 - JD.com and Lenovo deepened their "hybrid AI" cooperation, launching new products at CES 2026, with a focus on strategic collaboration around smart devices and services [11] - Alibaba's Qianwen app has integrated with various Alibaba ecosystem services, allowing users to perform complex tasks like ordering food and booking flights [12] - Alipay and partners launched China's first AI commercial agreement, ACT, designed to facilitate collaboration between AI and e-commerce platforms [13] Group 5 - Yunhai Medical released the "YunJian AI Spirit," a product that reduces long-term costs for users of infrared medical technology [14] - Zhiyuan purchased thousands of hours of robot training data for various tasks [15] - Meituan launched the open-source "ReThink" model, which significantly reduces training costs for new tools in real-world scenarios [16] Group 6 - Teslin introduced the upgraded T-Cluster 512 super node architecture, achieving over 500 PFlops of total computing power [17] - Keda Xunfei launched a marketing AI platform based on the "SuperAgent" framework to enhance efficiency in marketing strategies [18] - The first domestically trained text-to-image model, GLM-Image, was released by Zhiyun and Huawei [19] Group 7 - Tencent Cloud ADP launched the first "AI native Widget," enhancing task delivery experiences through natural language interaction [20][21] - Anthropic implemented stricter measures that disrupted several AI programming tools, affecting developers' projects [22] - Google announced a partnership with Walmart to expand AI model shopping capabilities, allowing direct transactions through its AI assistant [23] Group 8 - Mark Zuckerberg announced the "Meta Compute" project, aiming to build substantial AI infrastructure by 2030, while also planning layoffs in the Reality Labs department [24][29] - Alphabet's market value surpassed $4 trillion, joining a select group of companies [25] - Google and Apple finalized a multi-year AI collaboration agreement to support Siri with Google's AI technology [26] Group 9 - Nvidia and Eli Lilly plan to invest $1 billion in an AI drug lab over the next five years [27] - The US relaxed export controls on Nvidia's H200 chips to China [28] - Microsoft announced a plan to limit the impact of data center energy costs and water usage [30] Group 10 - Gemini launched personal smart products that allow users to personalize their experience through connected applications [31] - Microsoft is expected to spend $500 million on AI initiatives, including partnerships with Anthropic [32] - OpenAI is seeking US hardware suppliers for its planned consumer devices and robotics expansion [33] Group 11 - Elon Musk's lawsuit against OpenAI is set to go to trial in late April [34][35] - OpenAI and Cerebras announced a partnership worth over $10 billion to deploy a large-scale AI inference platform [36] - Zhi Variable Robotics completed a 1 billion yuan A++ round of financing, attracting investments from major firms [37] Group 12 - Qiangna Technology submitted a confidential IPO application in Hong Kong [38] - OpenAI acquired the AI health application Torch for approximately $100 million [39] - K2 Lab, founded by a former DingTalk executive, secured seed funding for its AI-driven content e-commerce platform [40] Group 13 - Alibaba Cloud completed a strategic investment in ZStack, achieving a controlling stake [41] - Skild AI raised nearly $1.4 billion, reaching a valuation of over $14 billion [42] - WeLab completed a $220 million D-round financing, marking its largest single round to date [43] Group 14 - Merge Labs, a brain-machine interface startup, raised $252 million in seed funding, with OpenAI as a major investor [44] Group 15 - A report indicated that by 2026, the Chinese tech giants index is expected to surpass the US tech giants in profitability growth for the first time since 2022 [45] - China is accelerating the establishment of a data property registration system [46] - A report predicted a significant increase in storage prices due to rising demand from AI and server capacity [47] Group 16 - A new AI model developed by US researchers can predict the risk of about 130 diseases based on sleep data [48] - Foreign investment firms are increasingly incorporating AI into their research processes in China [49] - UBS stated that the probability of an AI bubble in China is low, with monetization relying on cloud and advertising [50] Group 17 - The number of AI companies in China has exceeded 6,200, with applications expanding across various industries [51]

Core Insights - DeepSeek has released a significant technological achievement by open-sourcing a new paper and module called Engram, which introduces a "lookup-computation separation" mechanism to enhance the performance of large language models in various tasks [2][5]. Summary by Sections Introduction of Engram - Engram is a scalable, lookup-based memory module designed to improve the efficiency of language models by separating memory retrieval from computational tasks [10][18]. Need for Engram - Traditional large language models rely on Transformer and Mixture-of-Experts (MoE) architectures, which combine memory and computation in a way that can lead to inefficiencies. Engram aims to address this by allowing models to handle factual memory and logical reasoning separately [8][9]. Core Technology of Engram - Engram utilizes modernized hashed N-gram embeddings, allowing for O(1) time complexity in memory retrieval, which significantly reduces computational costs while maintaining high retrieval speed [11][13]. Relationship with MoE - Engram provides a new axis of sparsity that complements MoE by offering static memory retrieval capabilities, thus optimizing parameter efficiency. In a 27 billion parameter model, Engram can utilize a large number of parameters for memory while consuming minimal computational resources during inference [15][16]. Performance Metrics - Engram has shown improved performance metrics across various benchmarks, such as achieving a loss of 1.950 on the Pile dataset and an accuracy of 60.4% on MMLU with 5-shot learning, outperforming both Dense and MoE models [17]. Community Reception - The Engram technology has received positive feedback from the community, with users highlighting its potential to separate memory pattern retrieval from neural computation, marking a new direction in model architecture design [18][19][21]. Future Implications - Observers speculate that Engram will be a core component of DeepSeek's upcoming V4 model, indicating a significant architectural advancement in memory and reasoning collaboration [22][23].