At the Baidu WAVE SUMMIT + Deep Learning Developers Conference 2023, Baidu's Chief Technology Officer Wang Haifeng announced that the cumulative number of users of Wenxin Yiyi has exceeded 100 million, and the number of developers on Baidu's PaddlePaddle development platform, deeply integrated with the Wenxin Yiyi technical framework, has reached 10.7 million. On January 5, 2024, the "Smart Win-Win·China AIGC Industry Application Summit and Borderless AI Ecological Partner Conference" will be held in Hangzhou Future Technology City. The conference will focus deeply on AIGC applications, with an expected attendance of 1,000 people. At the same time, Borderless AI will hold the first ecological partner conference at the summit, breaking new ground in the industry.
Image Source: Generated by Borderless AI
Financing
Yuntian Lifei participates in Pre-IPO round financing of AI technology and solution provider Wisdom Interconnection
According to the "Science and Technology Innovation Board Daily" on December 28, the AI technology and solution provider "Wisdom Interconnection," focusing on intelligent road network scenarios, announced the completion of Pre-IPO round financing, with Yuntian Lifei, a listed company on the Science and Technology Innovation Board, as the investor. It is understood that after this round of financing, the company will continue to advance the deep research and development of intelligent hardware, AI algorithm large models, big data, scene innovation, and market promotion, as well as the introduction of high-end AI technology talents.
AI chip unicorn Suiyuan Technology increases capital to 100 million yuan, with Tencent holding 21.37% as the largest shareholder
According to the "Science and Technology Innovation Board Daily," on December 25, Shanghai Suiyuan Technology Co., Ltd. underwent a business change, with its registered capital increased from about 4.43 million yuan to 100 million yuan.
The company was established in March 2018, with Zhao Lidong as the founder, chairman, and CEO, mainly providing AI training and inference products, covering chips, boards, integrated AI computing machines, liquid-cooled computing clusters, and supporting software systems. Shareholder information from Tianyancha shows that the company is jointly held by Tencent Technology (Shanghai) Co., Ltd., Zhao Lidong, Zhang Yalin, and National Integrated Circuit Industry Investment Fund Phase II Co., Ltd. Among them, Tencent Technology (Shanghai) Co., Ltd. holds approximately 21.37% and is the largest shareholder.
Zhi Zhen Cloud completes tens of millions of yuan in Series A financing
According to Zhi Zhen Cloud's official WeChat account, recently, Beijing Zhi Zhen Cloud Intelligent Technology Co., Ltd. completed tens of millions of yuan in Series A financing, led by Qingdao Caitong Asset Management and Huashang Huitong. This round of financing will mainly be used for the research and development and market promotion of the integrated digital audit platform 3.0, and the audit vertical field large model chat-Audit, and other products.
Zhi Zhen Cloud is a leading domestic audit digitalization service provider, integrating business, data, and artificial intelligence technology deeply to provide customers with a more efficient, secure, and user-friendly one-stop solution.
AI innovation company Zhongshu Xinke completes tens of millions in Pre-A round financing
According to the "Science and Technology Innovation Board Daily" on December 27, Zhongshu Xinke recently completed tens of millions in Pre-A round financing, with this round of financing jointly invested by Chengmei Capital and Torch Xingye Innovation Fund.
Zhongshu Xinke is an AI innovation company with "knowledge intelligence" as its core, established by Yuncong Technology, Xiamen Torch Venture Capital, and Minsheng E-commerce, dedicated to creating intelligent application products that integrate knowledge refinement, organization, application, and enhancement, serving clients in industries such as education, elderly care, government affairs, and finance. Its main products are the "Zhongxun" knowledge intelligence platform and related industry scenario applications.
Large Models
ByteDance releases inference large model PixelLM, capable of efficient pixel-level inference
According to PingWest, the ByteDance Intelligent Creation Team recently announced the first efficient pixel-level inference large model PixelLM, jointly proposed with researchers from Beijing Jiaotong University and Beijing University of Science and Technology, which does not rely on SAM.
The research team stated that PixelLM can proficiently handle any number of open-domain targets and diverse complex inference segmentation tasks, while avoiding additional, costly segmentation models, improving efficiency and transferability to different applications. PixelLM outperforms other methods on most indicators of three benchmarks, and due to its independence from SAM, its TFLOPs are much lower than models of the same size.
China Academy of Information and Communications Technology releases "Fangsheng" large model benchmark testing system
According to Sina News, on December 24, 2023, the China Academy of Information and Communications Technology released the "Fangsheng" large model evaluation system, aiming to establish a unified "metric" for industry-wide large model benchmark testing. The "Fangsheng" testing system covers the four key elements of large model benchmark testing: testing indicators, testing methods, testing datasets, and testing tools, and has formed the standard "Overall Technical Requirements for Large-Scale Pre-training Model Benchmark Testing."
The "Fangsheng" large model benchmark testing system was jointly released by the China Academy of Information and Communications Technology, Beijing Zhiyuan Research Institute, National Key Laboratory of Cognitive Intelligence, and Tianjin University. In addition, 9 units including State Grid Intelligent Power Grid Research Institute, Capital Window, Tianjin University, China Telecom Research Institute, China Unicom Software Research Institute, Huawei, Jiagu Yi, Haitian Ruishi, and Eastern Fortune have become the first batch of partners for the "Fangsheng" large model benchmark testing.
Shanghai AI Laboratory releases open-source medical large model group "Pu Yi 2.0"
According to AI New Intelligence, on December 26, the official WeChat account of the Shanghai Artificial Intelligence Laboratory stated that at the recent "2023 Healthy China Sinan Summit," the Shanghai AI Laboratory and its cooperative partners, including Shanghai Jiao Tong University School of Medicine Affiliated Ruijin Hospital, jointly released the medical multimodal basic model group "Pu Yi 2.0" (OpenMEDLab2.0), achieving one-stop open-source for "production, study, research, use, and evaluation" of medical large models, aiming to provide support for AI medical applications across disciplines, diseases, and modalities.
It is reported that compared to the first batch of basic model groups released in June this year, the upgraded 2.0 version has added multi-domain models and language parameter increments; added 5 open-source datasets, including the largest medical image segmentation dataset SA-Med2 D-20 M and the pathology dataset SNOW jointly released by the Shanghai AI Laboratory team; added evaluation modules, introducing the benchmark A-Eval for cross-dataset evaluation of abdominal multi-organ segmentation, supporting model performance evaluation with few samples.
Tsinghua KEG Laboratory and Zhipu AI jointly launch image understanding large model CogAgent
According to IT Home on December 27, the Tsinghua KEG Laboratory recently cooperated with Zhipu AI to jointly launch the new generation image understanding large model CogAgent. This model, based on the previously released CogVLM, uses visual GUI Agent to directly perceive GUI interfaces more comprehensively and directly through visual modality (rather than text), in order to make planning and decisions. It is reported that CogAgent can accept high-resolution images of 1120×1120, with various capabilities such as visual question answering, visual localization (Grounding), and GUI Agent, achieving the first place in general ability on 9 classic image understanding lists (including VQAv2, STVQA, DocVQA, TextVQA, MM-VET, POPE, etc.).
Important breakthrough in AI application in the field of agricultural science and technology in China, "Shennong Large Model 1.0" released
According to China News Network, on the morning of December 29, the Agricultural Large Model Project Research and Demonstration Conference and the release ceremony of "Shennong Large Model 1.0" were held at China Agricultural University. At the event, the research team from China Agricultural University publicly released the industry large model "Shennong Large Model 1.0" for application in the agricultural field. The model has multiple functions including agricultural knowledge question answering, agricultural text semantic understanding, text summarization generation, and agricultural production decision reasoning. Experts at the conference stated that this marks a significant advancement of AI applications in China's agricultural field and signifies the addition of a new force in the industry large models developed by universities and research institutes, playing an important foundational and leading role in driving the modernization of agriculture through technology.
Applications
The first AIGC Industrial Application Summit in 2024 is set to open on January 5th!
AI New Intelligence reports that on January 5, 2024, the "Smart Win-Win·China AIGC Industrial Application Summit and Borderless AI Ecological Partner Conference" will open in Hangzhou Future Technology City. The conference is jointly guided by the Hangzhou Future Technology City Management Committee, the Yuhang District Science and Technology Bureau, and the Yuhang District Enterprise (Talent) Comprehensive Service Center, hosted by Shichuo Technology, and media support is provided by AI New Intelligence. The conference will deeply focus on AIGC applications, with an expected attendance of 1,000 people. Borderless AI will also hold the first ecological partner conference at the summit, breaking new ground in the industry.
As the first AIGC "thousand-person conference" of 2024, the summit has invited numerous industry elites and experts from research and development, investment institutions, universities, and AIGC entrepreneurs. Investment veteran Zong Peimin will share investment experience, and AI expert Zhang Wenyu will deliver a speech. The conference will also release the White Paper on the development of China's AIGC text and image industry and the first large model in the domestic silk industry.
Currently, over 300 industry executives and dozens of heavyweight guests have confirmed their attendance at the conference. The number of registered attendees has exceeded 800, with nearly half being founders, partners, directors, and various CXOs. Attendees can now register for free to interact face-to-face with numerous industry elites and experts and share in this intellectual feast.
Baidu CTO Wang Haifeng: Wenxin Yiyi users exceed 100 million
According to Sina Technology on December 28, at the Baidu WAVE SUMMIT + Deep Learning Developers Conference 2023, Baidu's Chief Technology Officer Wang Haifeng announced that the cumulative number of users of Wenxin Yiyi has exceeded 100 million. Behind these 100 million users, the daily number of user queries is rapidly increasing. At the same time, Wang Haifeng also revealed that the number of developers on Baidu's PaddlePaddle development platform, deeply integrated with the Wenxin Yiyi technical framework, has exceeded 10 million, reaching 10.7 million.
Quark App launches "Quark Health Assistant" large model application
According to Webmaster's Home on December 25, Quark App announced a comprehensive upgrade of its health search function, introducing a large model application called "Quark Health Assistant" and incorporating a new content interaction method in some search results and functional sections.
Quark has established partnerships with over 200 authoritative medical experts, more than 60 well-known public tertiary hospitals nationwide, and over 40 medical institutions. These partners provide Quark with rich health content resources, including information on the causes of illnesses, medication advice, and medical guidance.
Lebird X2 launches self-developed large model voice assistant RayNeo AI
According to 36Kr, the AR brand Lebird has launched the self-developed large model voice assistant Rayneo AI for internal testing on Lebird X2. Rayneo AI, in addition to traditional voice assistant functions, further incorporates capabilities such as multi-round natural language dialogue, itinerary planning, convenient encyclopedia Q&A, and brainstorming.
OPPO Find X7 series phones debut with 70 billion parameter AndesGPT edge-side large model, Xiao Bu assistant gets a new upgrade
According to IT Home on December 27, at today's OPPO Find X7 series product technology communication meeting, OPPO announced the new upgrade of the AndesGPT large model. AndesGPT large model is divided into three sizes: 180 billion parameters, 70 billion parameters, and 7 billion parameters, supporting edge-cloud collaborative deployment and intelligent scheduling according to user needs in different scenarios. OPPO stated that AndesGPT large model, through its powerful knowledge, memory, and tool capabilities, brings a broad knowledge base to Xiao Bu assistant, as well as personalized intelligent service capabilities. Through the innovative SwappedAttention technology, AndesGPT also solves the problem of delayed reasoning caused by long-term memory, bringing ultra-fast initial response speed. In the scenario of text content summarization, it only takes 0.2 seconds to generate the first 200 characters, leading competitors by 20 times; and 2.9 seconds to generate the first 2000 characters, leading competitors by 2.5 times. The maximum summary length can reach 14,000 characters, easily accommodating a common journal article, surpassing competitors' model capabilities by 3.5 times.
Major Companies
Google launches Google AI SDK to simplify integration of high-performance model Gemini Pro in Android applications
According to IT Home on December 26, Google recently launched the new Google AI SDK, aiming to simplify the integration of its high-performance Gemini Pro model in Android applications. With this SDK, developers no longer need to build and manage their own backend infrastructure.
For Android applications, Google provides the Google AI client SDK for Android, which encapsulates the Gemini REST API as a Kotlin native API. With this, developers no longer need to interact directly with the REST API or implement server-side services for accessing Gemini models in Android applications.
Microsoft AI chatbot Copilot available on iOS: Completely free, supports GPT-4
According to Fast Technology, Microsoft recently launched the Copilot App on the iOS platform, which is a more powerful AI chatbot application than the free version ChatGPT.
Unlike the free version ChatGPT using GPT-3.5, Copilot allows users to access OpenAI's new large language model GPT-4 and is completely free. Copilot also integrates the text-to-image generator DALL-E3, which can generate the desired images based on keywords or phrases, while ChatGPT only supports text chat.
Perspectives
Zhou Hongyi: Next year, large models will pursue "big" on one hand and also pursue "small"
According to IT Home on December 31, Zhou Hongyi, the founder and chairman of 360 Group, announced the official launch of the digital spokesperson "Digital Lao Zhou" and released an annual report. Zhou Hongyi made several predictions about the development trends of artificial intelligence in 2024:
Large models will not monopolize, unlike operating systems where there are only a few worldwide; their development path is more like PCs and will be ubiquitous in the future.
Next year, large models will pursue "big" on one hand and also pursue "small." This is a trend that has already been observed. Small-scale large models will soon be deployed on mobile phones and various Internet of Things devices, not just working in the cloud. Moreover, more large models may be deployed in cars next year.
Multimodal will become a standard feature of domestic large models. Currently, whether it's GPT-4 V or Gemini, they can understand images, videos, and audio. Currently, domestic large models are mainly text-based, but next year, they will also focus on multimodal capabilities.
Domestic will see many vertical large models, entering various industries and moving towards industrialization. The most important thing in 2024 is to prove that large models can truly solve problems.
In addition, Zhou Hongyi believes that next year, there will be a breakthrough in the multimodal capabilities of large models in the AIGC field, such as text and video. As for what kind of killer C-end applications will emerge next year, it remains to be seen.
Researchers: Global AI water consumption may reach 6.6 billion cubic meters by 2027
According to CCTV Finance, training and running AI large models typically require tens of thousands of servers to provide computing resources, storage resources, and low-latency network connections. These servers are deployed in clusters in physical locations called "data centers." When they work together, they generate highly concentrated heat in a short period of time, requiring a large amount of water resources for cooling.
Data shows that the computational power required for AI is expected to double every 100 days, and may increase by over a million times in the next five years. Researchers estimate that by 2027, global AI demand may consume 6.6 billion cubic meters of water, almost equivalent to the annual water consumption of the state of Washington in the United States.
Early OpenAI supporter predicts "huge deflationary" impact on the economy from AI in the next 25 years
According to Business Insider, billionaire and early OpenAI supporter Vinod Khosla posted on Monday that artificial intelligence will have a "huge deflationary" impact on the economy within 25 years. He stated, "Capital should be scarce for a while, current GDP and economic measures will no longer be as important, but goods and services should be very abundant." "The key question is what are the right measures and the right questions."
Khosla is one of the early supporters of OpenAI. His venture capital firm invested $50 million in OpenAI in 2019, the largest investment in the company's 15-year history. On December 12, Khosla stated at the Fortune conference that artificial intelligence is not the biggest threat in the world.
Lee Kai-fu: China's large model companies are in the stage of technical verification, and the large model is about to "reshuffle"
According to Wall Street News, Lee Kai-fu, Chairman and CEO of Innovation Works and CEO of Zero One Million, stated that competition in China's artificial intelligence field is still in the preliminary stage. After experiencing a major "reshuffle" in the industry, there will ultimately be only a few big winners, and some companies may gracefully exit, but most companies will either give up halfway or turn to more practical goals, such as building applications and solutions for specific industries.
Lee Kai-fu stated that China's large model companies are in the stage of technical verification, and they need to prove their ability to develop high-quality models. Those companies that pass the test will move on to the next stage, which is how to increase revenue and achieve profitability. When discussing the prospects of Chinese AI companies, Lee Kai-fu stated that he is more inclined to compete globally and has the opportunity to build different special models for different countries in the future.
Zhipu AI Zhang Fan: Rationally view large models, the biggest challenge next year is to commercialize and land them successfully
According to Time Weekly, Zhang Fan, Chief Operating Officer of Zhipu AI, stated in a recent interview that everyone's focus on AI is changing. In 2023, everyone was more focused on models, parameters, and metrics, forming the current situation of "models are king." In 2024, everyone will be more focused on the landing of AI and whether the technology can truly bring commercial value and user value. "I think the change in everyone's focus is our biggest challenge next year, and in the future, we will focus on the transition of this identity or scenario and accelerate the progress of commercialization and landing." Zhang Fan added.
Research Report
SuperCLUE releases the 2023 Annual Report on Chinese Large Model Benchmark Evaluation: The gap between domestic and foreign models is still significant
According to Beijing Business Daily, on December 28, the domestic Chinese model evaluation agency SuperCLUE released the 2023 Annual Report on Chinese Large Model Benchmark Evaluation. According to the report, in the past six months, leading domestic large model companies have achieved a miraculous intergenerational catch-up of large models. From a 20-point difference with GPT3.5 in July, there has been a stable and significant improvement every month, and by the November evaluation, they had surpassed GPT3.5 in total score.
The evaluation data selected the results of SuperCLUE's December evaluation, selecting 26 representative large models from domestic and foreign sources in December. The results show that the gap between domestic and foreign models is still significant. GPT4-Turbo leads by a wide margin with a total score of 90.63, higher than other domestic and foreign large models. The best domestic model, Wenxin Yiyi 4.0 (API), scored 79.02, with a gap of 11.61 from GPT4-Turbo and 4.9 from GPT4 (web).
However, domestic large models have made significant progress in the past year. There are 11 models with comprehensive capabilities surpassing GPT3.5 and Gemini-Pro, such as Baidu's Wenxin Yiyi 4.0, Alibaba Cloud's Tongyi Qianwen 2.0 and Qwen-72B-Chat, OPPO's AndesGPT, Tsinghua & Zhipu AI's Zhipu Qingyan, and ByteDance's Yunque large model, all performing well. In addition, domestic open-source models perform better in Chinese than foreign open-source models, such as Baichuan2-13B-Chat from Baichuan Intelligence, Qwen-72B and Yi-34B-Chat from Alibaba Cloud, all superior to Llama2-13B-Chat.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
