Broadcom is also reported to want to acquire Intel; OpenAI has launched a huge financing, requiring at least $250 million; Xiaopeng said that the real competition for AI intelligent driving is in the cloud丨AI Intelligence Bureau
Latest update time：2024-09-24
Reads：
Important news
NEWS REMIND
-
Broadcom is also reportedly evaluating the acquisition of Intel
-
OpenAI is reportedly asking investors to contribute at least $250 million in its new round of financing
-
Xiaopeng says the real competition for AI smart driving is in the cloud
-
GPT-4o plays "Black Myth"! Elite monster win rate exceeds human, pure large model solution without reinforcement learning
-
Google announces $120 million fund for global AI education
Today's headlines
HEADLINE NEWS
Broadcom is also reportedly evaluating the acquisition of Intel
Following the news that Qualcomm has made an acquisition offer to Intel, Broadcom is also quietly evaluating potential acquisition plans for processor giant Intel. People familiar with the matter said that Broadcom had previously evaluated potential acquisition offers for chip maker Intel, but has not yet evaluated an acquisition offer for Intel.
In 2018, Broadcom planned to acquire Qualcomm for more than $100 billion, but was stopped by the US government on the grounds of "threat to national security". At the end of last year, Broadcom also successfully acquired cloud computing company VMware for $61 billion. Through a series of large-scale mergers and acquisitions, Broadcom has now become a giant enterprise, with a market value of $799.136 billion (as of the close of US stocks on September 20).
Broadcom's semiconductor business revenue in fiscal 2024 is less than $30 billion, which is lower than Intel's revenue of about $60 billion in recent years. Therefore, it is not ruled out that Broadcom hopes to further strengthen its semiconductor business through the acquisition of Intel. If the merger is successful, the market value of the merged company may exceed $1 trillion. (Bianniushi)
Domestic Information
DOMESTIC NEWS
Xiaopeng says the real competition for AI smart driving is in the cloud
Recently, in Yunqi Town, Hangzhou, Zhejiang, He Xiaopeng, Chairman and CEO of Xpeng Motors, drove the "world's first AI car" P7+ to the 2024 Yunqi Conference. This car is equipped with the industry's leading end-to-end big model. He said at the conference that in the next three years, with the breakthrough of the end-to-end big model, novice drivers will be able to drive as easily as experienced drivers in any city. He emphasized that AI will dominate future industry competition. Li Liyun, head of autonomous driving at Xpeng Motors, further pointed out that end-to-end is only the starting point of the AI competition, and the real competition lies in the deployment of AI big models in the cloud. The cooperation between Xpeng and Alibaba Cloud is laying a solid foundation for this cloud competition. (Bailu Video)
Kai-Fu Lee's Zero One Beijing Company increases capital to 300 million
Recently, Beijing Zero One Everything Information Technology Co., Ltd. underwent industrial and commercial changes, and the company's registered capital increased from RMB 250 million to RMB 300 million.
Beijing 01.AI Information Technology Co., Ltd. was established in August 2023. Its legal representative is Ma Jie and Kai-Fu Lee is its executive director. Its business scope includes computer system services, artificial intelligence application software development, artificial intelligence entrepreneurship and innovation service platform, technology import and export, conference and exhibition services, information technology consulting services, etc. According to shareholder information, the company is wholly owned by 01.AI (HK) Limited. (Sanyan Technology)
GPT-4o plays "Black Myth"! Elite monster win rate exceeds human, pure large model solution without reinforcement learning
Alibaba's research team has developed a new VARP (Visual Action Role Playing) agent framework that can use game screenshots as input and use visual language model reasoning to generate actions in the form of Python code to operate the game.
In the test of Black Myth: Wukong, the research team defined 12 combat-related tasks and constructed a human operation data set containing 1,000 valid data. The agent won 90% of the simple and medium-difficulty combat scenes without the use of reinforcement learning. The agent operated through the action planning system and the human-guided trajectory system, where the action planning system consists of a context library, an action library, and a human-guided library.
Test results show that when fighting against some elite monsters, the AI's winning rate even exceeds that of human players. This research demonstrates the potential of large models to play games without reinforcement learning. (Qubit)
ARM Technology's first self-developed "Linglong" DPU and new generation VPU processor announced, and the shipment volume of self-developed related chips has exceeded 500 million
Arm China Co., Ltd. (hereinafter referred to as "Arm Technology"), an independent entity of the Chinese distributor of chip design listed company Arm, recently launched the first domestically developed "Linglong" D8/D6/D2 display processor, as well as the new generation of "Linglong" V510/V710 video processors.
The new "Linglong" D8/D6/D2 DPU integrates Arm frame buffer compression (AFBC) technology, and has high performance, high image quality, bandwidth saving and power consumption saving functions. The translation buffer unit (TBU) module in its built-in Arm system memory management unit MMU-600/700 can be specially optimized for different display scenarios to further reduce system access latency. It also supports the Arm TrustZone media protection (TZMP) architecture, realizing a hardware and software integrated design.
Zhou Hua, head of multimedia processor R&D at ARM, said that through the release of the "Linglong" D8/D6/D2 DPU and the "Linglong" V510/V710 VPU, ARM hopes to provide better and richer multimedia solutions for the high-definition display and high-definition video encoding and decoding needs of the mainstream market applications, meet the differentiated needs of different customers in terminal, edge and cloud scenarios, and help Chinese local partners seize market opportunities. (Titanium Media AGI)
Huawei Cloud launches AI native cloud infrastructure to provide AI computing power for enterprises
At the 9th HUAWEI CONNECT 2024, Zhang Pingan, Huawei's executive director and CEO of HUAWEI Cloud, delivered a keynote speech on "Migration to the Cloud, AI Reshaping Thousands of Industries".
He emphasized that enterprises must seize opportunities in the intelligent era and use AI to build their own leading advantages. The most important thing is to build AI-native thinking from now on. Enterprises should actively embrace AI, dare to open up scenarios, build their own AI platforms, and let AI serve their core businesses. Today, Huawei Cloud Pangu Big Model has been applied to more than 30 industries and more than 400 application scenarios, accelerating the intelligence of thousands of industries.
At the conference, Huawei Cloud released CloudMatrix, an AI native cloud infrastructure, which interconnects and pools CPU, NPU, DPU, memory and other resources to provide enterprises with powerful AI computing power. In addition, Huawei Cloud has fully upgraded its data governance production line DataArts to provide customers with an AI-oriented data base, including data development governance and knowledge services, which significantly improves resource utilization and data supply efficiency. Huawei Cloud also officially released a host cloud solution to help customers build a new cloud core that is highly available, easy to operate and maintain, and more agile. (Global TMT)
Using AI big models to help the low-carbon transformation of industry, Jifeng Technology received tens of millions of yuan in angel round financing
Recently, Jifeng Technology Co., Ltd. (hereinafter referred to as "Jifeng Technology") completed its first round of angel round financing of tens of millions of yuan, with Qingxin Capital as the sole strategic investor. The funds raised in this round will be mainly used to expand the talent team, purchase computing power and storage resources, and develop industrial control terminals based on general GPU architecture.
Founded in July 2023, Jifeng Technology is a technology-innovative enterprise focusing on the development of large AI models for industrial verticals, providing AI overall solutions for energy, equipment and process scenarios in process industries to increase efficiency and reduce emissions. "We seek to build a light but solid bridge between cutting-edge AI technology and solid industrial production, and try to use the problem-solving ideas of AI large model technology to solve the dilemma of high trial and error costs and difficult benefit quantification faced by high-energy-consuming enterprise users in achieving green dual-carbon transformation." Suo Qi, founder of Jifeng Technology, said.
It is reported that Jifeng Technology has reached long-term strategic cooperation intentions with many leading companies and project owners, including green hydrogen production, environmentally friendly waste incineration, green synthetic chemicals and other industries, and will complete the deployment and trial operation of multiple pilot projects by the end of 2024. (36Kr)
Tencent Yuanqi AI intelligent body supports publishing to official accounts: can create digital avatars, 7*24 hours intelligent customer service
On September 23, according to the official news of Tencent Hunyuan, its AI intelligent product Tencent Yuanqi now supports publishing to WeChat public accounts, bringing many new functions to public account operators.
According to Tencent, using Tencent Yuanqi, public account operators can create digital avatars to interact with fans in real time, and can also provide users with 7*24 hours customer service. Tencent Yuanqi can also provide article insertion services. The intelligent body can insert relevant content into public account articles to enhance the interactivity and information content of the content. It provides readers with a question-and-answer assistant and makes articles more practical.
Tencent Yuanqi is an open platform for intelligent agents launched by Tencent Hunyuan Big Model Team. Developers can quickly and easily create high-quality intelligent agents through plug-ins, knowledge bases, workflows, etc. It supports publishing to platforms such as QQ and WeChat, and also supports API calls. (Pinwan News)
DingTalk launches 365 membership product: providing personalized AI assistant, etc.
On September 23, DingTalk announced the launch of "365 Membership" for individual users, which includes benefits such as AI search, personal AI assistant, AI automatic reply, and automatic speed reading. Non-members can still use DingTalk's existing AI functions such as AI assistant and speed reading.
According to the official introduction, the AI search function in DingTalk 365 membership can help users integrate work and collaborative information. The personal AI assistant can set descriptions and become a hanging window, and supports direct questions to AI, screen recognition questions, photo questions, etc. For parent users, the "growth record" function of new members can intelligently identify children's photos and automatically record the growth trajectory of children. (IT Home)
Alibaba proposes a structure-preserving AI vision algorithm: significantly improving the quality of HDR image conversion to LDR image
It was learned at the 2024 Yunqi Conference that the Computing Technology Laboratory of Alibaba DAMO Academy proposed an AI vision algorithm based on a structure-preserving network, which can automatically convert high dynamic range (HDR) scene images into low dynamic range (LDR) images while maintaining their texture details, improving the image quality on conventional display devices by 7% compared to the industry.
The team innovatively proposed to use a dual-stream controlled diffusion network to maintain the structural details and tonal style of the image during the forward propagation process, which can achieve zero-shot learning. Specifically, this method uses mean subtraction and contrast normalization (MSCN) technology to extract the shared structural features of HDR and LDR images as a "bridge", guiding the diffusion model to learn the HDR image structure and maintain the tonal style of the LDR image, and finally generating an LDR image with rich texture details and balanced colors.
This technology can help ISP chips achieve smarter light adaptability and dynamic range adjustment, and is also expected to be combined with other diffusion models to develop new possibilities for AIGC technology. Facing new computing needs such as graph computing, autonomous driving, multimodal AI, and embodied intelligence, the team will continue to explore full-stack innovations in software and hardware to help the development of the chip industry and the digital economy. (Fast Technology)
Moore Threads: Open source OpenCV-MUSA, supports most data structures and APIs under the cv::cuda namespace
Moore Threads announced that its self-developed unified system architecture MUSA has completed the adaptation with the open source computer vision library OpenCV, and officially released the OpenCV - MUSA open source project. OpenCV is an important open source library in the field of computer vision and is widely used in many fields. Moore Threads has added the MUSA device backend based on the existing OpenCV code, providing acceleration support for multiple algorithm modules and adapting the compilation scripts.
Currently, OpenCV - MUSA already supports multiple modules and most of the data structures and APIs under the cv::cuda namespace. Developers only need to replace cv::cuda with cv::musa to achieve the same functions on MUSA devices. (ITSoul)
AI enables home appliance quality inspection to move from “blind hearing” to “voiceprint recognition”: accuracy exceeds 99%
The World Manufacturing Conference was held in Hefei, Anhui from September 20 to 23. The new generation information technology exhibition area of the conference showcased two major sections: "Advanced Photovoltaics and New Energy Storage" and "New Generation Information Technology (Chip, Screen, Soul and Terminal)", with a total of 263 exhibits.
In the exhibition area of the National Intelligent Voice Innovation Center, staff demonstrated the home appliance AI quality inspection device and automated AI detection system. Zhao Yixiao, a staff member of the National Intelligent Voice Innovation Center, said: "The first application of air conditioner indoor unit quality inspection has achieved the upgrade from 'blind human hearing identification' to 'automatic machine voiceprint recognition'." The exhibition area simulated a real assembly line scene, combined with artificial intelligence technologies such as voiceprint, machine vision and cognition, and realized AI quality inspection through algorithms.
Zhao Yixiao said that AI quality inspection products have been implemented in home appliance companies such as Haier and Gree, covering air conditioner indoor and outdoor units, washing machines, water heaters and other categories. The number of quality inspection products has reached 2 million, with an accuracy of over 99%. In addition to the home appliance industry, this AI quality inspection product can also be used in manufacturing industries such as automobiles, photovoltaic semiconductors, etc. (IT Home)
Meitu Qixiang's large model upgrades its video generation capabilities and is gradually being implemented in Meitu Xiuxiu, Wink, MOKI and other products
On September 23, Meitu announced that the video generation capability of the Meitu MiracleVision model has been fully upgraded. Based on the triple advancement in generation capability, generation efficiency, and model performance, combined with Meitu's many self-developed technology advantages in the field of computer vision, the video generation duration, image quality, fluency, authenticity, and credibility have been significantly improved.
The single-text-generated video length and single-image-generated video length of MiracleVision are both 5 seconds. It supports the generation of ultra-long videos of 1 minute, 24FPS frame rate, and 1080P resolution, and can be output in any video size.
Currently, the new video generation capability of Meitu MiracleVision has been applied to some of Meitu's products. With its highly reusable modular capabilities, the upgrade advantages will gradually cover Meitu Xiuxiu, Beauty Camera, Wink, Kaipai, Meitu Design Studio, WHEE, MOKI and other products, further deepening the depth and breadth of Meitu's application in the video tool track. (Pinwan News)
International News
FOREIGN NEWS
OpenAI is reportedly asking investors to contribute at least $250 million in its new round of financing
Recently, OpenAI is in late-stage negotiations for a huge round of financing that could value it at $150 billion. According to people familiar with the matter, the company has asked investors to contribute at least $250 million. Earlier news said that OpenAI is close to reaching a new round of financing of $5 to $7 billion, led by Thrive Capital, which has pledged to invest $1 billion. Microsoft, Nvidia and Apple are in talks to invest. One of the people familiar with the matter said that the technology companies' share of this round of financing is expected to reach $2 to $3 billion, but it is not clear whether they will invest in cash or a mix of cash and other resources (such as computing power). Tiger Global and Abu Dhabi artificial intelligence investment company MGX are also discussing participating in this round of investment. (Interface)
Google announces $120 million fund for global AI education
On September 21, Google CEO Sundar Pichai spoke at the United Nations Future Summit, describing AI as the most transformative technology to date and announcing a new fund for AI education and training around the world.
Pichai pointed out four major opportunities between AI and sustainable development: helping people get information in their own language, accelerating scientific discovery, providing climate disaster alerts and tracking, and driving economic progress. Pichai acknowledged that there are risks with AI, such as deep fakes, but he did not mention the impact of AI on the climate. He added that he hopes to avoid a global AI divide, and Google is creating a $120 million Global AI Opportunity Fund, through which it will work with local nonprofits and NGOs to provide AI education and training in communities around the world.
On the same theme, Pichai called for regulations on smart products that mitigate harm and resist national protectionist impulses — otherwise, he predicted, regulation could widen the AI divide and limit the benefits of AI.
The U.S. Department of Commerce will announce measures to ban the use of Chinese AI autonomous driving software and hardware
The U.S. Department of Commerce will propose to ban the use of Chinese software and hardware in smart connected cars and cars with AI self-driving technology in the United States for "national security" reasons. The report said that this move is a major escalation of the United States' ongoing restrictions on Chinese cars, software and parts.
This move has previously caused concerns among stakeholders, including US allies and automotive industry groups, who believe that the software and hardware of automotive systems usually need to undergo a series of pre-production tests and cannot be easily replaced. Others said that it is difficult to completely avoid Chinese technology and parts given the wide range of smart connected vehicles, urging the US government to carefully define and give the industry sufficient time to check and reorganize the supply chain.
In response, Chinese Foreign Ministry spokesman Mao Ning repeatedly said that the relevant remarks are not only false narratives, but also a typical manifestation of the politicization and security of economic and trade issues. "According to this logic, shouldn't China be more worried that Washington can allow hundreds of millions of Chinese users' Apple phones to transmit information to the United States, or even turn off the screen at the same time?" (Titanium Media AGI)
AI company Perplexity plans to launch a new advertising model: brands can bid for sponsorship questions
Artificial intelligence startup Perplexity AI is in talks with several well-known brands, including Nike and Marriott International, to launch a "sponsored Q&A" advertising system by the end of the year. The system aims to challenge Google's dominance in the $300 billion digital advertising market.
Perplexity AI is developing a new advertising model that allows brands to bid for "sponsored questions" and get AI-generated answers that have been reviewed by advertisers. This means that when users search using Perplexity, they will see brand-related questions and their answers, thereby increasing brand exposure. Perplexity CEO Aravind Srinivas said that the effectiveness of ads lies in their relevance and the quality of the brand. He hopes that this new advertising system will become a "money printing machine" to create more value for brands. (AIbase)
TSMC and Samsung consider building chip factories in UAE
According to foreign media reports, TSMC and Samsung Electronics, the world's two largest chip manufacturers, are discussing the construction of potential chip projects in the United Arab Emirates in the coming years, which could be worth more than $100 billion.
TSMC executives recently visited the UAE and discussed building a factory complex comparable to some of the company's largest and most advanced facilities in Taiwan, according to reports. South Korea's Samsung Electronics is also considering a major new chip manufacturing operation in the country. "We are always open to constructive discussions to promote the semiconductor industry, but we remain focused on current global expansion projects and have no new investment plans to disclose at this time," TSMC said in a statement on September 23. Samsung declined to comment on the report.
Under the terms of the preliminary discussions, the projects would be funded by the UAE, with Abu Dhabi's sovereign development agency Mubadala playing a central role. The report added that the broader goal is to increase global chip production and help lower prices without hurting chipmakers' profitability. (Saasverse)
G42 announces collaboration with NVIDIA to advance AI-driven climate solutions using Earth-2
According to foreign media reports, AI company G42 announced a partnership with NVIDIA. The two parties will use AI solutions to promote the development of climate technology and improve the accuracy of global weather forecasts. This cooperation is based on NVIDIA's Earth-2, an open platform that accelerates climate and weather forecasts through interactive, AI-enhanced, high-resolution simulations. G42 and NVIDIA will initially focus on weather forecast models with square kilometer resolution to improve the accuracy of meteorological forecasts.
In addition to promoting innovation in climate technology, the collaboration will also focus on building a powerful framework that combines enhanced weather prediction capabilities with comprehensive data indicators and visualizations. This will help global organizations achieve their sustainable development goals through well-informed, data-driven environmental strategies. (Pinwan News)
OpenAI o1 is now open to enterprise and educational users, with a maximum of 50 uses per week
OpenAI's first models with "reasoning" capabilities, o1 and o1-mini, are now accessible to enterprise and education users. Users can input up to 50 messages to the o1 model per week, and the o1-mini model can be used up to 50 times per day. The o1 series models have a higher ability to understand, and will think deeply before answering questions to generate a chain of thoughts and then give answers. They can handle more complex tasks and are good at science, programming, and math problems.
Although GPT-4o is still the best choice for most tasks, the o1 series performs better in handling complex, problem-solving oriented tasks. Because the o1 model consumes more computing resources, it is initially only open to Plus/Team/Enterprise/Edu users in limited quantities. In terms of API, the price of access to related model access is relatively high. The input word cost of o1 is US$15 (about RMB 107) per million, and the output word cost is US$60 (about RMB 427) per million. The o1 model can be called up to 500 times per minute, and the o1-mini model can be called 1,000 times per minute. (ITSoul)
Intel considers $5 billion investment proposal from private equity firm Apollo
Private equity firm Apollo Global Management Inc. is reportedly planning a multibillion-dollar investment in Intel Corp. The proposed investment comes at a critical time for the chipmaker, once the world's most influential, which is now struggling and has seen its stock price fall nearly 60% since the beginning of the year.
Some media reported that Intel executives are considering Apollo's proposal. Discussions about the deal are still in the early stages and details have not been finalized. Sources told Bloomberg that the investment size may change and nothing may be achieved in the end.
Earlier this year, Apollo offered Intel $11 billion for a 49% stake in the company's Fab 34 chip manufacturing facility in Leixlip, Ireland, a deal designed to help with the continued development of the facility. (Saasverse)
CAIS issues a $500,000 reward: Bringing together top experts to solve difficult problems
Scale AI and CAIS (Center for AI Safety) plan to jointly launch Humanity's Last Exam, claiming to build the most difficult open source benchmark for large language models; the MMLU previously released by CAIS is still the most downloaded benchmark dataset in the Huggingface community.
The two organizations raised a total of $500,000 to collect the "hardest and most extensive" questions ever, and to collect those sharp questions from all human experts to push the artificial intelligence model to its limit: the first 50 selected questions will receive $5,000 each; the next 500 questions will receive $500 each. After the data set is released, the submitters of the selected questions will also become co-authors of the paper. The more questions are accepted, the higher the position will appear. Currently, researchers from MIT, UC Berkeley, Stanford University and other universities have participated in the prize competition. (New Intelligence)
Sharp launches AI Smart Link wearable device: an AI assistant hanging around the neck, supporting natural language communication
At the Sharp Technology Day event held on September 17-18, Sharp announced that it had collaborated with Kyoto University of Arts in Japan to jointly develop a wearable device that can enable natural communication with generative AI - AI Smart Link.
AI Smart Link is equipped with Sharp's edge AI technology "CE-LLM (Communication Edge-LLM)". This technology can instantly judge and process user inquiries, support fast-response end-side AI, and cloud AI (such as ChatGPT) that can obtain rich information. In addition, the device adopts a lightweight neck-hanging design of about 100 grams, which can be worn for a long time.
According to Sharp officials, Sharp will conduct empirical experiments with Kyoto University of Arts starting this month, and will also start negotiations with Fairy Devices and NEC to jointly develop the next generation of hands-free AI devices including AI Smart Link, with the goal of commercializing the product in fiscal 2025. (IT Home)
Google CEO Pichai: AI will not replace programmers, but will make more people become programmers
Sundar Pichai, CEO of Google and its parent company Alphabet, talked about the impact of AI on programming jobs in a speech at Carnegie Mellon University. He said that AI is most likely to help people rather than replace them. It can help existing programmers focus on more advanced tasks rather than repeatedly solving the same problems. In addition, Pichai believes that AI will lower the threshold for programming and enable more people to program through natural language, just like using a creative tool.
Pichai also described AI's ability to help many professionals in their daily activities as the "right perspective" to look at artificial intelligence. He also said that the term "artificial intelligence" is not appropriate, which easily leads to the idea of comparison, but it is not necessary. He believes that a better term should be "empowered intelligence" and he "bets that more people will be programming in the future." (IT Home)
DeepMind's latest 50-page paper proposes the AligNet framework: "Aligning" humans with hierarchical visual concepts
DeepMind's latest research proposes the AligNet framework, which aims to improve the human-likeness of artificial intelligence systems by simulating the human visual concept hierarchy. The study points out that human visual perception is robust and generalizable to different environments, while the performance of existing deep learning models will deteriorate when the distribution of training and test data is inconsistent.
AligNet trains a teacher model to mimic human judgment and transfers this human-like structure to a pre-trained visual base model, thereby improving the model's performance on multiple tasks and enhancing generalization and robustness. Experimental results show that this framework can significantly improve the consistency of the model with human behavior, especially when dealing with more abstract category levels. In addition, AligNet also provides a set of open source visual models, enabling science, medicine, and industry to use more human-like and robust visual models for downstream applications. (New Intelligence)
Today's Hot Videos
//
Recent Hot Articles
京公网安备 11010802033920号