This article understands the layout of the "GPT-like model" of domestic Internet giants

On the news, following the release of heavy products such as OepnAI GPT-4, Baidu Wenxin, Microsoft Office Copilot, and Google Bird, the next day domestic Internet giants made new statements in the field of AI.

Comprehensively combing through market data and a number of brokerage research reports, we have compiled the news of major Internet giants in AI large models, ChatGPT-like products and their applications for your reference.

1. Tencent.

On March 22, Tencent revealed at the performance meeting that generative AI can be used to improve the user experience of Tencent's flagship products, and every user may have an artificial intelligence assistant in the future, and if the effect is good, it is possible to incorporate generative AI into WeChat and QQ.

In terms of AI large models, Tencent said that it is investing heavily in artificial intelligence and cloud infrastructure construction, Tencent Mixed-Element AI large models cover NLP (natural language processing), CV (computer vision), multimodal and many industry and domain models, and has also launched trillions Chinese NLP pre-training models.

In terms of specific applications, in early February, Tencent's Tencent Research Institute released the "AIGC Development Trend Report 2023", which pointed out that in the advertising field, Tencent's hybrid AI model can support intelligent production of advertising, that is, the use of AIGC to automatically generate advertising copy into advertising videos, which greatly reduces the cost of advertising video production.

2. Ali.

On March 23, Ali Dharma Academy has launched the "text generated video large model" in the AI model community "Magic Ride" ModelScope.

According to the official introduction, the overall model parameters are about 1.7 billion, and currently only support English input. It is reported that "Magic Ride" is an AI model community launched by Ali Damo Academy and CCF Open Source Development Committee at the 2022 Yunqi Conference, and the first batch of open source models exceeds 300, including vision, speech, natural language processing, multimodal and other major AI directions.

In terms of AI large models, according to information released by Alibaba Research Institute, Alibaba DAMO Academy launched the M6 project of Chinese multimodal pre-training model in early 2020, and launched a basic model with 300 million parameters in June of the same year. In January 2021, the scale of model parameters reached 10 billion; In May 2021, the model with the scale of trillion parameters was officially put into use; In October 2020, the parameter scale of M6 expanded to 10 trillion, becoming the world's largest AI pre-training model at that time.

Alibaba Cloud has said that M6 has been applied in more than 40 scenarios, with hundreds of millions of calls per day. Internally, the application of the M6 model includes but is not limited to the launch of clothing designed for the brand by Rhino Intelligent Manufacturing on Taobao, creating scripts for Tmall virtual anchors, and improving the search and content recognition accuracy of Taobao, Alipay and other platforms, especially good at design, writing, Q&A, and landing in the prospects of e-commerce, manufacturing, literature and art, scientific research and so on.

On February 7, DingTalk publicly claimed that its App can access ChatGPT-like functions in DingTalk robots to realize robot dialogue-related operations. On February 8, according to media reports, Ali's version of the chatbot ChatGPT is under development and is currently in the closed beta stage.

3. Huawei.

Huawei previously responded to "Huawei's layout in the direction of similar ChatGPT" on February 10, saying that the company began to have a layout in large models in 2020, and released the Pengcheng Pangu large model in 2021, which was the industry's first 100-billion-level natural language processing large model for generating and understanding Chinese at that time.

According to the data, at the Huawei Developer Conference 2021 (Cloud) in April 2021, HUAWEI CLOUD released a series of ultra-large-scale pre-training models for Pangu, including a vision (CV) pre-training model with 3 billion parameters and a Chinese language (NLP) pre-training model with 100 billion parameters and 40TB of training data jointly developed with Loop Intelligence and Pengcheng Lab.

The Chief Scientist of HUAWEI CLOUD AI Field said: HUAWEI CLOUD Pangu Big Model can replicate a large AI model in many scenarios in generalization, generalization, and scale, reducing the dependence on data annotation, and using the ModelArts platform to transform AI development from a workshop to a new mode of industrial development. ”

In terms of large-scale model industrialization, Huawei has launched the Intelligent Remote Sensing Open Source Ecosystem Consortium, the Multimodal Artificial Intelligence Industry Consortium, and the Intelligent Fluid Mechanics Industry Consortium.

4. ByteDance.

According to the news of First Finance and Economics on February 24, ByteDance is laying out large models and making efforts in language and image modes. The person in charge of ByteDance's related technologies said that the exploration of the technology middle office in these fields is still in its early stage and has not yet matured.

According to the report, the language big model team is led by the byte search department, and the current scale is about a dozen people; The image model team is led by the intelligent creation team under the product development and engineering architecture department.

People familiar with the matter said that the ByteDance language big model team was formed this year, and the exploration direction is mainly to combine with search, advertising and other downstream businesses, and the language big model team is expected to launch a large model in the middle of this year.

5. JD.com.

According to interface news reports, on February 10, JD Cloud's Yanxi artificial intelligence application platform announced that it will integrate past industry practices and technology accumulation and launch the industrial version of ChatGPT: ChatJD, whose parameters are expected to be 100 billion, and announced the "125" plan of ChatJD's landing application roadmap.

The "125" program consists of a platform, two areas (retail and finance), and five applications (content generation, human-machine dialogue, user intent understanding, information extraction, sentiment classification).

6. NetEase.

According to interface news reports, on February 8, NetEase Youdao said that it will launch ChatGPT homologous technology products in the future, and the application scenarios revolve around online education.

It is reported that NetEase's NetEase Fuxi is also worth paying attention to. Founded in 2017, NetEase Fuxi is a domestic artificial intelligence research and application institution specializing in the game and pan-entertainment industry, with research directions including reinforcement learning, image animation, natural language, virtual humans, user portraits, big data platforms, cloud computing platforms, cloud games and other fields.

At present, NetEase Fuxi has served more than 200 customers, and the average daily application call volume exceeds hundreds of millions. According to the official website, Fuxi's products mainly include the metaverse virtual activity platform "Yaotai", the human-machine collaboration PaaS platform "Youling Robot", the virtual human platform and the intelligent excavator.

7. Three six zero.

On February 7, 360 said on the interactive platform that it is planning to launch a demo (trial version) application of ChatGPT technology as soon as possible.

360 said that at present, 360 search is the top 2 of Chinese search engines, with a market share of 35%, and the company's artificial intelligence research institute has been continuously investing in AIGC technology including ChatGPT-like technology since 2020, but so far it is only used as a productivity tool for internal business use, and the investment scale and technical level are still far behind the current ChatGPT 3, and the technical indicators can only be slightly stronger than ChatGPT2. Due to the training data source and application direction, the actual effect is stronger than that of ChatGPT 2 in the Chinese environment.

In addition, in terms of relevant listed companies, Minsheng Securities has also sorted out.