Hangzhou issued implementation opinions on accelerating the innovation and development of artificial intelligence industry. In 2025, the industrial ecology of "high computing power+strong algorithm+big data" will be basically formed, and a virtuous circle will be realized in which big computing power incubates big models, big models drive big industries and big industries promote big development, and Hangzhou will be built into a national computing power cost depression, model output source and data sharing highland. Following last year’s announcement that AI was successfully used to control the nuclear fusion plasma inside the tokamak for the first time, DeepMind achieved a controlled nuclear fusion breakthrough with AI, and the accuracy of plasma shape simulation increased by 65%.
01. Financing
CGI, a Canadian information technology service company, plans to invest about 750 million US dollars to expand the AI field within three years.
According to a report by MarketWatch on July 24th, CGI, a Canadian information technology service company, plans to invest 1 billion Canadian dollars (about 756.3 million US dollars) in three years to cooperate with customers to expand artificial intelligence services and solutions.
Web3 and AI startup ZYBER 365 completed $100 million Series A financing.
According to Financial Express, Web3 and AI startup ZYBER 365 announced the completion of a series of financing of 100 million US dollars, with the company’s valuation reaching 1.2 billion US dollars, and the British SRAM & MRAM Group participated in the investment.
AI security company Protect AI completed $35 million in financing, led by Evolution Equity Partners.
According to TechCrunch’s report on July 26th, AI security company Protect AI announced that it has completed the $35 million Series A financing, with Evolution Equity Partners leading the investment, and Salesforce Ventures, Acrew Capital, boldstart ventures, Knollwood Capital and Pelion Ventures participating. The scale of this round of financing is more than twice that of the Protect AI seed round that ended in December last year, bringing the total financing of the startup company to 48.5 million US dollars.
AI startup Graft received $10 million in financing, led by Radical Ventures.
According to TechCrunch, on July 26th, the AI startup Graft announced that it had received a seed investment of US$ 10 million, with Radical Ventures leading the investment and GV participating. Last year, the company raised a total of $4.5 million before the seed round of financing, and the total financing amount has reached $14.5 million. At the same time, Graft announced that when it starts to open the artificial intelligence development platform to more companies in a controllable way, it will invite companies to apply for access to the platform.
AI protein design company "Liwen Institute" completed tens of millions of angel round financing.
According to the investment community, recently, AI protein design company Hangzhou Liwen Institute Biotechnology Co., Ltd. announced the completion of tens of millions of angel round financing. This round of financing was led by Cailletet Capital, followed by Panlin Capital and Hongshi Capital, and the seed round led the investor Zhenge Fund to make additional investment. The financing funds will be mainly used for the optimization and development of the AI protein design platform, and to promote the incubation and construction of multiple product pipelines.
"Lingao Technology" completed a multi-million-dollar seed round of financing and devoted itself to building large-scale model middleware.
According to 36Kr’s report on July 31st, Lingao Technology, a startup company dedicated to building data pipelines and large-scale middleware, recently officially announced the completion of seed round financing. According to reports, the amount of this round of financing is in the order of millions of dollars, and the investors are Jingya Capital and Plug and Play. At present, the company mainly provides two products-data pipeline (Vanus Connect) and large model middleware (Vanus AI). Among them, Vanus AI, the middleware of AI big model, can help enterprises to link big models and build knowledge base, debug prompt engineering at the same time, and finally build their own AI applications.
The AI-driven enterprise learning platform "Little Knowledge" has completed tens of millions of Pre-A rounds of financing.
According to the report of science and technology innovation board Daily on July 28th, a little knowledge (Beijing) Technology Co., Ltd. (abbreviated as a little knowledge) has completed tens of millions of yuan of Pre-A financing, which was led by Zhencheng Investment, followed by former UFIDA CFO Xu Zhoujin, and joined the company as a co-founder. This round of financing is mainly used for brand building, market expansion and promoting product research and development based on AI large language model.
AutogenAI, an AI generation tool used to write tenders, received an investment of 22.3 million US dollars.
According to TechCrunch’s report on July 27th, London AI startup AutogenAI received an investment of $22.3 million from Blossom Capital. The new funds will be used to hire more talents, expand the platform and expand the customer base. People familiar with the matter said that the investment valued AutogenAI at about "hundreds of millions of dollars". The company had previously raised about $3.5 million.
02. Start a business
Sogou’s former CMO Hong Tao joined the big model company founded by Wang Xiaochuan.
According to the science and technology innovation board Journal reported on July 27th, Hong Tao, former CMO of sogou, announced in a circle of friends that he would join Baichuan Intelligent, an AI big model company jointly founded by Wang Xiaochuan, former CEO of sogou, and Ru Liyun, former COO of sogou, to take charge of commercialization. On July 11th, Baichuan Intelligent officially released two quantized versions of the general large language model Baichuan-13 B-Base, dialogue model Baichuan-13 B-Chat and its INT4/INT8 with a parameter of 13 billion.
Wang Xiaochuan AI Company added several shareholders and increased its capital to 6.47 million.
According to Tianyancha App, recently, Beijing Baichuan Intelligent Technology Co., Ltd. has undergone industrial and commercial changes, adding Baichuan Zhongzhi (Beijing) Management Consulting Partnership (Limited Partnership) and a number of natural persons as shareholders, and the registered capital of the company has increased from 5 million RMB to about 6,476,700 RMB.
03, big model
Know Smart Technology Releases 140 Billion Parameter Know Smart KDF Model
According to the report of IT House on July 25th, Beijing Zhiwei Intelligent Technology Co., Ltd. recently released a large model of intelligent KDF in Shanghai, and a series of products developed based on this model, including financial industry tools such as "KDF Zhixun", "KDF Never" and "KDF Zhongshu".
It is reported that the training data of the KDF model is mainly in Chinese, and it contains a lot of financial data to improve the problem-solving ability of the model in the commercial and financial fields. In addition, some English and code data are combined in the training data to adapt to the general ability of the model. In the training process, the intelligent KDF model treats a single Chinese character as an independent Token. The model parameters amount to 140 billion, and the number of training Token reaches 400 billion.
TransGPT?, an open source transportation model of Beijing Jiaotong University, is far-reaching.
On July 19th, Beijing Jiaotong University, together with china computer federation Smart Transportation Branch and Zuzhi Multimode Company, officially released and opened the first comprehensive transportation model in China-TRANSGPT Zhiyuan, which was independently developed by the research team of the Key Laboratory of Transportation Big Data and Artificial Intelligence of the Ministry of Education of Beijing Jiaotong University.
"Zhiyuan" takes the meaning of "going far and being far-reaching", and adopts the large text model, multi-modal model and real-time scene data calling mode based on Transformer architecture, forming an overall framework based on comprehensive traffic model and supplemented by traffic segmentation industry applications. Some of its main functions and features include: intelligent traffic planning, intelligent traffic management, intelligent travel assistant, traffic policy research and traffic safety education. At present, some data, codes and models of Zhiyuan Big Model have been opened in GitHub.
Yongyou released YonGPT, a big model of enterprise service.
According to a report by Cailian on July 27th, UFIDA released YonGPT, a large enterprise service model. According to the official introduction, YonGPT is an important component of UFIDA BIP, which can adapt to the mainstream general language model in the industry, and realize enterprise intelligence in four aspects: business operation, human-computer interaction, knowledge generation and application generation by providing general capability services, domain services and industry services.
Netease has officially released the "Zi Yue" big model.
According to the report of science and technology innovation board Daily on July 26th, Netease Youdao officially released the "Zi Yue" big model this evening. According to Zhou Feng, CEO of Netease Youdao, the "Zi Yue" big model has been applied to the software and hardware in Youdao Ecology, and it has been commercialized, including big model translation, AI composition correction, English grammar intensive speaking, virtual human oral coach, AIBox, document question and answer and so on.
Xinyi Technology released the first vertical video model in China.
According to 36Kr’s report on July 27th, Xinyi Technology held the "New Vision Xinyi Video Model Release Conference" in Beijing. At the meeting, Xinyi Technology released the first large-scale model focusing on video in China, and the AIGC video generation platform "Sechuang Video Creation Platform", the first commercial digital platform "Xinyi Digital People Platform". In addition, it also released the first domestic star digital person blessing video production product "Secondwish".
Google’s team of experts released the medical model Med-PaLM M, which can encode and interpret biomedical data.
According to the report of Webmaster’s House on July 31st, a team of experts composed of several research institutes under Google recently published a paper and released a multi-modal model named Med-PaLM M.
Med-PaLM M is a large-scale multi-modal generation model, which can flexibly encode and interpret biomedical data. Compared with the existing model, Med-PaLM M is competitive in many tasks, and even performs better in some tasks. The researchers also demonstrated the transfer learning and zero sample reasoning ability of Med-Palm M. However, in practical application, further work is needed. This research has a positive prospect for future medical solutions based on artificial intelligence.
MosaicML, an AI company, has launched a 7 billion parameter model MPT-7 B-8 K, which is said to be "commercially available for processing 8,000 words of text at a time".
According to the news of IT House on July 25th, AI startup MosaicML recently released its 7 billion parameter model MPT-7 B-8 K. It is reported that the model can handle 8000 words of text at a time, and it is quite good at dealing with key abstracts and questions and answers of long articles, and can further fine-tune relevant configurations according to specific tasks on the MosaicML platform.
It is reported that the series model uses 1.5 million Token, and it takes 3 days to complete the model training with 256 H100 GPU. MosaicML released three versions of the model this time, including MPT-7 B-8 k, MPT-7 B-8 k-Instruct and MPT-7 B-8 k-Chat.
Foreign media: OpenAI is developing a new open source model G3 PO to fight against Microsoft Llama 2.
According to IT House, citing The Information, a foreign technology media, OpenAI is currently developing a new open source model codenamed "G3 PO" in order to counter the open source model Llama 2 jointly developed by Microsoft and Meta, and it is not clear when it will be released.
Reddit community users believe that OpenAI has delayed the release of the open source model for two reasons: the OpenAI team is relatively small, focusing on promoting the application mall and providing customers with a market for selling customized AI models. This helps OpenAI to build a moat; OpenAI also wants to create a personalized ChatGPT assistant. According to sources, launching a real "derivative version" directly competes with Microsoft, and this effort "may take several years."
Shanghai "Fengwu" big model accurately predicts "Du Surui", and its accuracy is better than that of Europe and America.
According to the news of China Meteorological Bureau, the No.5 typhoon Du Surui made landfall on the coast of Jinjiang, Fujian at 9: 55 on July 28th, and the maximum wind force near the center was 15. Shanghai Artificial Intelligence Laboratory (Shanghai AI Laboratory) predicted the typhoon path based on the original research and development of the Fengwu meteorological model (hereinafter referred to as "Fengwu"), and achieved accurate prediction. Based on the comparison between the forecast path and the actual path of typhoon from July 21 to July 27, the 24-hour advance error of Fengwu is 38.7 kilometers, which is better than the 54.11 kilometers of ECMWF, the international authority of meteorological forecasting, and the 54.98 kilometers of NCEP. The numerical comparison of forecast 120 hours in advance shows that the error value of wind and black is 121.4 km, which is obviously better than that of ECMWF (293.8 km) and NCEP (826.5 km).
04. Application
Xiaohongshu online "at the moment" pure text note-taking function, using AI technology to generate pictures.
According to 36Kr’s report on July 25th, Little Red Book recently launched the "Now" function. With the "Now" function, the user can directly input text, and the algorithm will generate a picture for the user based on the input text content and complete the publishing. In addition, users can also doodle on the drawing board and use AI to beautify the doodle before publishing it. The landing of this function benefits from AIGC technology. At present, this function has been fully launched in the latest version.
BMW is using AI technology to reduce the production cost of automobile factories, saving $1 million a year.
According to CNBC reported by IT House, BMW recently deployed artificial intelligence technology in its factory in Fort Sparthan, which is said to save $1 million (about 7.15 million RMB) every year. BMW did not disclose the name of the artificial intelligence tool, but simply stated that the tool is "mainly aimed at the company’s production workshop" and can play the role of "optimizing the manufacturing process, improving efficiency, reducing waste and predicting equipment failures".
Runway, the behind-the-scenes special effects company of "Instant Universe", opened the AI video generation tool, and registered an account for free trial.
According to a report by IT House on July 26th, Runway, the behind-the-scenes technology company of Instant Universe, was previously invested by Google, and its AI-enabled video editing tool Gen-2 was made public. Users can directly use text prompts to generate "realistic video content" and "automatically edit videos".
Photoshop beta introduces the AI function of "generative expansion", which can easily expand images.
According to the report of IT House on July 27th, Adobe has launched an artificial intelligence function called "Generative Expand", which has been provided to Photoshop beta users. This function allows users to easily expand the image and create a seamless scene effect. Adobe introduced this function in its blog post, and showed its usage and application scenarios.
It is reported that "Generative Expand" is part of the Firefly artificial intelligence project under Adobe. The project initially focused on image generation and text effects, and later added vector recolor and Generative Fill functions. The Generative Fill function allows users to fill the blank areas in the image with one click, and it has been added to the Photoshop beta in May this year. The "Generative Expand" function allows users to expand the canvas of the image and create content matching the original image without giving any hints.
Trailer: Genesis, the first sci-fi movie trailer made entirely by AI, was released online.
According to Machine Heart, a sci-fi movie trailer "Trailer: Genesis", which was all done by AI, became popular on social media. The trailer lasted less than a minute, and everything from image and video synthesis to music editing was done by AI. Nicolas Neubert, the producer, listed the corresponding AI tools, namely, Midjourney to process images, Runway to process videos, Pixabay to process music, and CapCut to edit videos. It is understood that Neubert spent 7 hours, Midjourney generated 316 tips and enlarged 128 images, Runway generated 310 videos, and there was a video with text, and 44 videos were used in the trailer.
Stack Overflow officially launched its OverflowAI code tool.
According to the news of IT House on July 28th, Stack Overflow, a question-and-answer platform for programmers, has quietly "officially announced" its code tool OverflowAI. It is reported that OverflowAI is not a "single product", but a collection of multiple AI tools, and users can use these tools to achieve related applications in a targeted manner.
DeepMind uses AI to achieve controlled nuclear fusion breakthrough: the accuracy of plasma shape simulation is improved by 65%.
According to the science and technology innovation board Journal reported on July 28th, after DeepMind announced last year that it had successfully controlled the nuclear fusion plasma inside the tokamak for the first time, the company announced the latest experimental simulation results, which further improved the plasma shape accuracy by 65%. Related research was submitted to the preprint website ArXiv on July 21st, and the paper was titled "Practical Reinforcement Learning for Magnetic Control of Tokamak".
05. Policy
Hangzhou will accelerate the innovation and development of artificial intelligence industry, and has defined four tasks: computing power support and model ecology.
On July 27th, the General Office of Hangzhou Municipal Government issued the Implementation Opinions of the General Office of Hangzhou Municipal People’s Government on Accelerating the Innovation and Development of Artificial Intelligence Industry. The "Implementation Opinions" pointed out that Hangzhou will thoroughly study and implement the spirit of the important speech of the Supreme Leader General Secretary on artificial intelligence series, take the deep integration of artificial intelligence and the real economy as the main line, and make every effort to build an innovation system from algorithm model innovation breakthrough to industry transformation and application, so as to realize a virtuous circle in which large computing power incubates big models, big models drive big industries and big industries promote big development, and provide strong support for reshaping the first city of the national digital economy at a high level and striving to promote the "two firsts". By 2025, the industrial ecology of "high computing power+strong algorithm+big data" will be basically formed, and Hangzhou will be built into a national computing power cost depression, model output source and data sharing highland.
06. Views
Open AI CEO: AI can improve employee productivity, but it will also create more unemployment.
According to the news of webmaster’s home on July 28th, Sam Altman, CEO of OpenAI, said that generative artificial intelligence technologies such as ChatGPT may improve the productivity of many employees in the next few years, but it may also lead to some people losing their jobs. In an interview with Atlantic Monthly, Altman refuted the view that the AI craze will only have a positive impact on employees. He said, "Many people who are engaged in AI pretend that it will only benefit; It will only be a supplement; No one will be replaced. " But the widespread application of AI may also create new jobs. Altman expects to create better jobs in addition to any disrupted jobs, although he is not sure whether the replaced employees can find these new positions.
Musk: The programmer’s favorite Stack Overflow Q&A platform will be replaced by AI model in the future.
According to a report by IT House on July 26th, the number of users of Stack Overflow, a well-known software developer forum operated by Stack Exchange, is gradually declining. A user named mediator shared an article entitled "The Fall of Stack Overflow" with a chart showing that "the traffic of Stack Overflow has decreased by about 50% in the past year and a half, and the number of questions/answers appearing on the website has also decreased by about 50%".
Musk made a brief comment on this article, arguing that such question-and-answer platforms will gradually be replaced by AI models. He said that the Stack Overflow platform was "killed by LLM (Large Language Model)". Musk also said: "LLM will lead to more and more common (question and answer platform) deaths."
Musk: Within two years, it will change from "lack of silicon" to "lack of electricity" and hinder the development of AI.
According to reports from Wall Street, Musk predicted at the energy conference held in Austin last month that with the increase of electric vehicle consumption, the power demand in 2045 will triple compared with the current one. Musk predicts that by 2030, half of the new cars sold worldwide will be electric vehicles. Recently, Musk once again warned that power shortage may have serious consequences, just as the recent lack of core hinders the technology and automobile industries. This time, the lack of electricity may hinder the development of AI. "We will go from … today’s extreme shortage of silicon to … power shortage in two years." He said.
According to the latest report released by Digital Information World, the energy consumption generated by data centers for training AI models will be three times that of conventional cloud work. It is estimated that by 2030, the power demand of data centers in the United States will increase at an annual rate of about 10%.
Viewpoint: Most large models are running with each other. If GPT3.5/4 is open source, it will face a devastating blow.
According to articles published by Wall Street, most large models are running with them. If GPT3.5/4 is open source, it will face a devastating blow. The article believes that, first of all, the gap between major models and Open AI is still obvious, and the market space originally reserved for other major model manufacturers is already very small. The emergence of open source promotes the big model to enter the hundred-model war, but the problem is that the product homogeneity is serious and there is no ability comparable to GPT4, which will lead to the very low discourse power of the big model and the pricing of the big model will only become more and more difficult; Secondly, in order to optimize the output results, large model enterprises need to iterate the model continuously, which requires a large number of professional data feeding, so data enterprise providers will also become an important role restricting the development of large models. Finally, the training cost of the big model is very high. If you want to make a super big model, you have to work hard, that is, to see who can grab more NVIDIA A/H100 chips; Generally speaking, most large model enterprises are faced with the dilemma that it is difficult to collect fees while desperately spending money. Presumably, many large model enterprises will give up this crowded and unprofitable track in the future.
07, research report
Mizuho: NVIDIA’s AI-related revenue is expected to reach 300 billion US dollars in 2027, and its AI server chip market share will be 75% by then.
According to the science and technology innovation board Journal reported on July 25th, Mizuho analyst Vijay Rakesh recently released a research report that NVIDIA’s AI-related revenue is expected to reach 300 billion US dollars in 2027, when its AI server chip market will still account for 75%. According to the report, with the increasing demand for generative AI, there will be a lot of opportunities for hardware vendors supporting the high computing demand of large language models (LLMs), especially in NVIDIA. However, this does not mean that the market closes the door to AMD. Rakesh predicts that by 2027, AMD’s share of the AI server chip market will be about 5%.
IDC: It is estimated that by 2026, 50% of terminal equipment processors in China will have AI engine function.
According to the science and technology innovation board Journal reported on July 25th, IDC announced a forecast that by 2026, nearly 50% of the terminal devices in China market will have AI engine technology. IDC believes that in the future, AI terminals will be widely developed in the fields of creation and creative future, games and virtual worlds, speech synthesis and conversion, vision and image processing, chat robots, personalized recommendation, medical assistance and so on.
Huatai Securities: It is estimated that by 2030, the regulatory market scale of domestic AI big model will reach 50-100 billion yuan.
According to the science and technology innovation board Journal reported on July 25th, Huatai Securities released a research report saying that the scale of China’s artificial intelligence core industry will exceed 1 trillion yuan in 2030. In general, the proportion of safety in information investment is at least 5%-10%, and due to the particularity of AI big model, AI safety will become a problem that all participants must consider in the future, which runs through the whole process from data labeling, model training and development, content generation and application development, and the investment is no less than that of traditional safety investment. Therefore, assuming that the proportion of AI supervision in the whole industrial chain is 5%-10%, it is expected to reach 2000.
ABI Research: In the next seven years, generative AI will bring in more than $450 billion in revenue.
According to a report by the webmaster’s house on July 27th, ABI Research, a global technical intelligence company, predicts that in the next seven years, generative AI will bring more than $450 billion to the enterprise market in 12 different vertical fields.
Although there are hundreds of application cases of generative AI in enterprise vertical field, its adoption will be divided into three stages. The current stage is internal focus, which increases the productivity of employees by providing generating tools. The second stage will have a greater impact on external services, and enterprises will be able to start building products or services around generative AI. The third stage will be the most valuable creator, and the manufacturing and logistics industries will use generative AI to automate and optimize processes. At present, the enterprise market is still in its infancy, and most markets are still exploring use cases and deployment options.
1. Bill of Quantities for Prevention and Control of Epidemic Normalization in COVID-19. pdf
