2023 will undoubtedly be the year when AIGC will set off a wave. Since Microsoft-invested OpenAI Research Office released the fairy-level natural language generation AI-chatGPT, the field of artificial intelligence has been completely "killed crazy" .
Some people say that the significance of its birth is no less than the emergence of AI painting generation models such as Stable Diffusion, and some people say that it is equivalent to the "IPhone" in the AI world .
Although the GPT model is currently at the 4.0 stage, it has passed the Turing test that no AI can pass so far, and compared with any AI dialogue robot currently on the market, chatGPT can completely win .
Faced with such a hot and powerful AI, all enterprises naturally see business opportunities, and can't wait to put AI large models, ChatGPT-like products and their applications into practical scenarios.
From Microsoft 's new version of Bing&Edge, "Microsoft 365 Copilot X", to Google 's "Bard" and Meta's "LLaMA", the competition in foreign markets is already fierce.
Turning to the domestic market, many companies also want to seize this trend, keep up with the technology of chatGPT, and take corresponding actions.
1. The layout of the "GPT-like model" of domestic Internet giants
01
Baidu's "One Heart in Classical Chinese"
"All in AI" is the theme and focus of Baidu in recent years.
Baidu's OKR content in the first quarter of 2023 shows that its key task is "leading the intergenerational change of the search experience . "
With the emergence of chatGPT, Baidu is the first to take action in China, announcing the ChatGPT project, the name is determined to be "Wen Xin Yi Yan", and the English name is ERNIE Bot.
Baidu has discussed two solutions, one is to launch an independent portal; the other is to emulate Microsoft Bing's "New Bing" after accessing ChatGPT, that is, to open up with the existing Baidu search portal.
Finally, on March 16, Baidu officially released "Wen Xin Yi Yan". Its related functions have also been launched on Baidu search one after another, including comprehensive capabilities in five usage scenarios: literary creation, commercial copywriting, mathematical calculation, Chinese understanding, and multi-modal generation. The first batch of users can experience the product on the official website of Wenxin Yiyan through the invitation code, and it will be opened to more users in the future.
Source: Baidu Wenyan Yixin official website
Wenxin Yiyan is the first ChatGPT-like product that has attracted a lot of attention in China. It has attracted a large number of users as soon as it was released . At the press conference, Robin Li announced that 650 institutions have announced to join the Wenxin Yiyan ecosystem. According to Baidu's official data, as of 10:00 on March 17, 76,000 companies have applied for Wenxin Yiyan's API call service test .
For this reason, some people said that the golden decade of AIGC may have begun.
In addition, last week, Baidu launched the enterprise-level large-scale model service platform "Wenxin Qianfan", which not only includes Wenxin Yiyan, but also includes Baidu's full set of Wenxin large-scale models and corresponding development tool chains. In the future, Wenxin Qianfan will also support third-party open source large models .
02
Alibaba's "Tongyi Thousand Questions"
In addition to Baidu's positive statement, Dachang Ali will certainly not slack off.
As soon as ChatGPT became popular, Alibaba Dharma Institute has quickly invested in the research and development of [ChatGPT-like] dialogue robots. In March, Ali launched the " Text Generation Video Large Model " on the AI model community "Magic Build" ModelScope, which was first opened to employees in the company for testing.
On April 7, the Ali version of ChatGPT was opened to the public for enterprise invitation testing, called Tongyi Qianwen.
As a large language model, the ability of Tongyi Qianwen mainly focuses on text generation, that is, chatbots like ChatGPT. In addition to dialogue, it can also quickly generate various types of copywriting.
Source: Tongyi Qianwen official website
With continuous internal testing and rapid iteration, Ali soon officially launched the Ali large model at the "2023 Alibaba Cloud Summit" on April 11, and has formed a model-as-a-service (MaaS) architecture .
This architecture includes basic large model/generic large model, enterprise-specific large model, magic building community, API service, etc. Zhang Yong, CEO of Ali Group, said that in the future, all Ali products will be connected to large-scale models for comprehensive upgrades, including Tmall, DingTalk, AutoNavi Maps, Taobao, Youku, Hema, etc.
The picture comes from the Internet
Zhou Jingren, Chief Technology Officer of Alibaba Cloud Intelligence, introduced that "Tongyi Qianwen" supports multiple rounds of dialogue, copywriting, logical reasoning, multi-modal understanding, and multi-language support. And demonstrated various scenarios of using "Tongyi Qianwen" to improve work efficiency, such as calling the travel interface to recommend travel products; video conferences can automatically generate conference summaries; as an intelligent shopping assistant, it will automatically recommend brands, products etc.
Next, Ali is expected to launch an industry application model on April 18, which is very worth looking forward to~
03
Tencent's "Hunyuan Assistant"
For Tencent, it already has relevant technical reserves in the fields of AI large models, machine learning algorithms, and NLP. Therefore, in this "AI large-scale model competition", it is natural not to be afraid of falling behind.
For ChatGPT-like conversational products, Tencent gathered internal heavyweights and established the " HunyuanAide " project team headed by Zhang Zhengyou, specializing in NLP (Natural Language Processing), CV (Computer Vision), multimodal and other basic models and many industry and domain models.
In the end, the team successfully launched the trillion-level Chinese NLP pre-training model——HunYuan-NLP-1T, and implemented the model in Tencent's internal products such as advertising, search, and dialogue, and served external customers through Tencent Cloud.
The picture comes from the Internet
Based on the model, Tencent is already developing a [ChatGPT-like] chatbot. Moreover, once this chatbot is developed, it will be integrated into QQ and WeChat, and will also serve B-side users through Tencent Cloud .
In the field of advertising, the "AIGC Development Trend Report 2023" released by Tencent Research Institute pointed out that its hybrid AI model can support the intelligent production of advertising, that is, using AIGC to automatically generate advertising copy into advertising videos, greatly reducing the cost of advertising video production.
Although compared with giants such as Google and OpenAI, Tencent still has a gap in the technical research and application of large language model AI. However, with its huge user base, diversified business layout and strong technical strength, Tencent still has its unique competitive advantage in the development of large language model AI.
04
Bytedance's "My AI"
Byte has always responded faster to the layout of the AI field.
Feishu, a subsidiary of ByteDance, will recently launch a new intelligent AI assistant - "My AI".
My AI will include automatically summarizing meeting minutes, creating reports, optimizing and continuing to write text content . My AI can also help users automatically create schedules and search the company's internal knowledge base through dialogue.
Source: Feishu official video account
Regarding the large model, ByteDance has also made a layout, focusing on language and image modes.
According to informed sources, the large language model team is led by the byte search department, and currently has a scale of about ten people; the large image model team is led by the intelligent creation team under the product development and engineering architecture department.
In addition, a person familiar with the matter said that the exploration direction of ByteDance's large-scale language model is mainly the combination with downstream businesses such as search and advertising , and also indicated that the large-scale language model team is expected to launch a large-scale model in the middle of this year.
Many people believe that among the top Internet companies, the company that can finally compete with Baidu on the large model will most likely be ByteDance.
On the one hand, senior AI practitioners revealed that Byte does not lack GPUs .
*After the Nvidia A100 graphics card was discontinued in China, the high-performance GPU required for training large models has become a scarce and important resource for many companies.
On the other hand , both the engineering team and the cloud service team ByteDance, which are critical to the development of large models, have certain advantages.
However, what kind of large-scale answer sheet ByteDance can hand over still needs time to verify.
05
Jingdong's "Yanxi"
JD.com, which is also a major manufacturer, has also begun to be questioned whether it is also making relevant actions (creating [ChatGPT-like] products).
Its vice president, He Xiaodong, said—"JD.com has been advancing on the ChatAI and AIGC tracks, but it is more focused on the 2B industry. It will make further breakthroughs in the near future, including launching next-generation products on this track."
Sure enough, at the artificial intelligence large model technology summit forum, JD.com made a formal "announcement", that is, in response to real industrial needs, JD.com plans to release a new generation of 100 billion-level industrial model this year - "Yanxi", " Jingdong version"ChatGPT.
Its pre-training parameters have reached the level of 100 billion , the category covers more than 3,000 types , the pass rate of manual review exceeds 95% , and the text that can be generated is 3 billion+ . It is said that Yanxi, a customer service system, currently interacts with users 10 million times a day .
Source: Tencent News Website
At the same time, JD.com also announced the roadmap for the application of the large-scale industry model of "Yanxi" - the "125" plan. The plan includes a platform, two fields (retail and finance), and five applications (content generation, human-computer dialogue, user intent understanding, information extraction, and sentiment classification).
It can be seen that as the hottest artificial intelligence research and development direction at present, the industrial model has become one of the follow-up development plans of many technology companies.
06
Huawei's "Pangu"
Huawei also made a positive response to the news about the layout of the [ChatGPT-like] direction.
As early as September 2020, Huawei started the layout of the large model, and released the Pengcheng Pangu large model at the Huawei Developer Conference in April 2021 - a large model for generating and understanding Chinese natural language processing at a level of 100 billion. Since then, it has been in the intensive development and testing phase.
In September 2021, Huawei Cloud released the Pangu drug analysis model;
In June 2022, Huawei Cloud released a large model of the Pangu Mine;
In November 2022, the Pangu weather large model, the Pangu ocean wave large model, and the Pangu financial OCR large model will be unveiled at the same time.
Until recently, with the frenzy of AIGC set off by chatGPT, Huawei has made new progress——
The picture comes from the Internet
Huawei's Pangu large model demonstrated its progress and application at the "Artificial Intelligence Large Model Technology Summit Forum" on April 8. The "Pangu Series AI Large Models" include NLP (Natural Language Processing) large models, CV large models, scientific computing large models (meteorological large models), multimodal large models, and speech large models.
But throughout the narrative, it can be seen that the Pangu model is different from language models such as "Wen Xin Yi Yan". Its application on the consumer side is not particularly prominent, but the B-side application is more eye-catching . Moreover, Huawei does not focus much on the large-scale NLP language model, but focuses more on the application of all aspects of the industrial end .
07
360 GPT - "Red Boy"
As a top 2 search engine company in China, 360 Search has been continuously investing in AIGC technology including [ChatGPT-like] technology since 2020, but the relevant results are only used as productivity tools for internal business use.
In terms of investment scale and technical level, it is naturally unable to match ChatGPT 4, and various technical indicators can only be slightly stronger than ChatGPT2.
But in order to keep up with the times, 360 Search is also trying to keep up with the latest technology.
Soon 360 Search developed an artificial intelligence product matrix based on the 360GPT large model , and announced that it will open internal testing for enterprise users.
On March 29, at the Digital Security and Development Summit Forum, Zhou Hongyi officially released the 360 version of "ChatGPT".
The picture comes from the Internet
Judging from the live demonstration, 360GPT does have a certain reasoning ability, but when answering questions, it still needs to ask and remind many times before it can answer a satisfactory answer. Therefore, 360GPT was ridiculed as "Red Boy" by netizens during the live broadcast.
In addition, the founder also said that 360 Search will launch smart office products soon, and with the addition of ChatGPT-like capabilities , it can't help but remind people of the "365 days, every day office" approach of Microsoft.
However, in the field of browsers in the domestic market, 360 Search has successfully completed the first landing of the search scene.
08
NetEase's "Zi Yue"
In order to keep up with the wave of localization of ChatGPT, Netease Youdao also claimed recently that it will be the first to launch a [ChatGPT-like] model for self-developed education scenarios in China, and the name of the model has been determined as "Ziyue".
It is reported that the AI oral teacher and Chinese composition correction DEMO developed based on "Ziyue" have been completed. This will also be the first "ChatGPT-like" model launched by a domestic company in an educational scenario .
Previously, NetEase Youdao had opened up the industrial chain by establishing a "content + hardware + AI" ecological closed loop, so as to facilitate its rapid commercialization.
On the one hand, Youdao covers lifelong learning courses from children to adults, and has accumulated a large amount of teaching content and user data, providing an essential infrastructure for model development.
Source: Netease Youdao official website
On the other hand, Youdao has explosive hardware products such as Youdao dictionary pen and AI learning machine. At present, its smart hardware business has become a strong growth point of the company, providing an important support for the implementation of AI technology on the end.
In addition, NetEase Fuxi, a subsidiary of NetEase, is also worth looking forward to.
NetEase Fuxi was established in 2017. It is a department within NetEase that specializes in the research and application of artificial intelligence in games and entertainment. Its research directions include reinforcement learning, image animation, natural language, virtual human, user portrait, big data platform, cloud computing platform, Cloud games and other fields.
With the continuous maturity and application of ChatGPT technology, NetEase Fuxi has also begun to pay attention to the potential of this technology. Internally, designers conducted in-depth research and exploration on ChatGPT technology , and tried to apply this technology to the design work of multiple businesses in Fuxi's department.
After learning and mastering the application technology of ChatGPT, designers can conduct user needs analysis and interaction design more efficiently, and can improve work efficiency and quality.
2. Other "outstanding players" in AI technology layout
01
HKUST Xunfei's "1+N Cognitive Intelligence Model"
As a leading AI company, iFLYTEK was not absent from this competition, and announced that it will release the product-level model of iFLYTEK's "1+N cognitive intelligence model" and generative pre-training model on May 6. application.
Among them, "1" represents the base platform for the research and development of general cognitive intelligent large model algorithms and efficient training, "N" is a dedicated large model version used in multiple industries, and the demonstration application products of "N" scenarios are also will be presented accordingly.
The iFLYTEK AI Learning Machine, iFLYTEK Hearing and other products will be equipped with chatGPT technology and become the first products to be launched.
Source: HKUST Xunfei official website
The reason why HKUST Xunfei has full confidence to say the exact product line and clear time point is also inseparable from its years of accumulation in the three major aspects of algorithms, data and computing resources.
In the field of algorithms , iFLYTEK has remarkable achievements in NLP technology and research and development experience in pre-trained language models;
In terms of data , HKUST Xunfei has accumulated large-scale professional data in education, medical care and other fields, and has a large amount of data brought by Xunfei Input Method, Xunfei Open Platform, and consumer office products;
In terms of computing resources , as an AI player for decades, HKUST Xunfei has naturally accumulated a lot of computing power and funds.
Although the effects and functions of iFLYTEK's new products are still unknown, it is foreseeable that a new route of intelligent educational hardware will eventually be opened.
02
"Consultation" of SenseTime
On April 10, SenseTime, known as the "leader in the AI industry", launched the first large-scale model system in China - the "SenseNova" large-scale model system, and released the first self-developed Chinese language large-scale model Application platform - "Consult SenseChat".
The feature of "Consultation" is that its parameters have reached the level of 100 billion , and it can realize text generation, image generation, multi-modal content generation and other capabilities and scene applications.
The release of Shangtang is also different from the previous domestic ChatGPT players. It also released other generative AI models and applications in one breath:
-
Miahua : AI content creation community platform.
-
Ronin : AI digital human video generation platform.
-
Qiongyu, Gewu : 3D content generation platform.
The picture comes from the Internet
SenseTime's GPT large-scale model products will be applied in medical, short video, education, marketing, and development industries.
In addition, SenseTime's "SenseNova" large-scale model system has fully supported business sectors such as smart cars, smart life, smart commerce, and smart cities , and SenseTime will provide customers with image generation, natural language dialogue, visual reasoning and Annotation service and other API interfaces.
Compared with other large-scale model products, Wang Xiaogang, co-founder of Shangtang Technology, said that the GPT version of Shangtang is mainly for business-end (B2B) business. Domestically speaking, SenseTime has successfully stood in the first echelon dedicated to To B (enterprise end) business .
03
Kunlun Wanwei's "Tiangong"
Kunlun Wanwei has also always maintained enthusiasm and belief in AGI. Its R&D team has also been working from a 100 million-level model to a billion-level model as early as 2020, and then to a 10-billion-level model... So far, it has been working for three years.
Recently, Kunlun Wanwei finally announced that "Tiangong" 3.5, a domestic large language model jointly developed by Kunlun Wanwei and Singularity Zhiyuan, will be released soon, and invite testing will start on April 17.
Source: Kunlun Wanwei official website
It is said that this will be the first GPT-like large model that truly realizes the emergence of intelligence .
And Kunlun Wanwei said that the "Tiangong" large model is very close to the intelligence level of OpenAI's ChatGPT. Since ChatGPT is based on the GPT3.5 large model, Kunlun Wanwei named this version "Tiangong" 3.5. In the future, "Tiangong" 4 and "Tiangong" 5 will also be born one after another.
04
Maimai launches "Almighty Maiyou"
Maimai is a job hunting and professional community software. After the ChatGPT became popular, its first action was to quickly prepare for access to the powerful question-and-answer capabilities of AI, and launched a question-answer robot with ChatGPT-like functions - "Almighty Maiyou".
The picture comes from the Internet
The robot can make real-time answers to community questions, but whether this feature can really help people find jobs and gain workplace knowledge remains to be verified.
Lin Fan, the founder and CEO of Maimai, said that the functions of ChatGPT will definitely replace some of the existing work content . Human beings will gradually develop and optimize in the fields they are best at during the entire process of machine and artificial intelligence development. It is a general direction and trend we have seen that the work content with relatively poor quality of life is handed over to computers and machines.
05
The virtual digital human of Tianyu Mathematics
The "MetaSurfing-Yuanxiang Intelligent Cloud Platform" launched by Tianyu Digital has incorporated AIGC function modules. Its virtual digital human beings "Tianyu" and "Yan Zhichu" are even connected to ChatGPT and other models.
The picture comes from the Internet
At present, Tianyu Technology is applying for the official access of GPT-4, which is mainly used to continuously optimize content production efficiency and creativity, and accelerate digital-real integration and industrial upgrading.
It is reported that the virtual characters "Tianyu" and "Yan Zhichu" will join hands with Baidu's Jidu Automobile to embark on a trip to the moon.
06
Cross-border communication: the company has connected to openAI
According to the news on March 30, Kuajietong stated on the interactive platform that the company has already connected to openAI.
At the same time, the company also stated that chatGPT has a large number of application scenarios in the fields of product details writing, keyword extraction, intelligent customer service, and small language translation.
At the same time, the company is also actively exploring the application of stable diffusion and other image generation models in the fields of AI model image generation and clothing pattern generation .
07
Sogou's "Baichuan Smart"
Wang Xiaochuan, the founder of Sogou, has also started a large-scale model business. He and former Sogou COO Ru Liyun jointly founded an artificial intelligence company called "Baichuan Intelligence", aiming to create a Chinese version of OpenAI's basic large-scale model and disruptive upper-level applications .
Wang Xiaochuan stated in the open letter that the new company has initially formed a team of 50 people and received 50 million US dollars of start-up capital, and strives to release the best large-scale models and disruptive products in China within this year.
In recent interviews, it was also known that Sogou will make a large model to benchmark GPT 3.5 by the end of the year.
According to incomplete statistics, in addition to the above-mentioned companies, New Oriental, Douyin, Kuaishou, Xiaomi, Waterdrop, Wondershare Technology, etc. have all announced that they have or are about to get involved in ChatGPT-related businesses.
write at the end
The appearance of the AI technology chatGPT is obviously high-profile, and many people have exaggerated its amazing performance. What's more, it is constantly outputting signals of anxiety and stress to the outside world.
But in fact, chatGPT is not a leading technology . Large companies almost all started to study large language models at the same time. It just seized the right moment and "pre-emptive strikes."
In addition, compared with some robots that were ridiculed as "artificial mental retardation" on the market before, chatGPT is obviously much more intelligent and humane, so it is a matter of course for it to go out of the circle.
As for the domestic chatGPT technology layout battle, although it is eye-catching, the ultimate test is the strength of each company.
You know, to reach the level of GPT4, domestic hardware resources, AI training, and data corpus are obviously not enough, let alone issues such as industrial technical barriers.
But the IT Bear Bureau still wants to wait and see, hoping that domestic companies can realize a real technological revolution.
*Part of the reference content in this article comes from the public domain of the Internet, and the copyright belongs to the original author
END
PS. If you are also an IT person, or a small partner who wants to enter the IT industry, you are welcome to join our public account: [IT Insider] online community , where you can share the experience and knowledge of technical experts, Exchange and share resources with peers, and even collide with wonderful sparks of cooperation~