Baidu Wenxin said A

On March 20, 2023, Baidu officially launched Baidu's version of ChatGPT - Wenxin Yiyan. Its functions mainly include: (1) literary creation; (2) Business copywriting; (3) mathematical and logical deduction; (4) Chinese comprehension; (5) Multimodal generation. At the press conference, Robin Li proposed three major industrial opportunities in the AI era, including: emerging cloud computing - MaaS model as a service; Fine-tuning of industry models – industry, finance, transportation, energy, media, etc.; Application development – text, image, audio and video generation, digital human, 3D generation, etc.

Wenxin Yiyan has six core technologies: supervised fine-tuning, RLHF, prompt construction, knowledge enhancement, retrieval enhancement and dialogue enhancement. The first three are very similar to ChatGPT's technology, and knowledge enhancement includes knowledge internalization ("infiltration" knowledge into model parameters) and knowledge externalization (meaning that the model can directly use external knowledge); Retrieval enhancement refers to the search of content based on the Baidu search engine, and then the useful part of the integrated output is screened; Dialogue augmentation refers to techniques such as memory mechanisms, contextual understanding, and conversation planning.

Ali Tongyi model

On September 2, 2022, Ali DAMO Academy released the Tongyi large model series. The model creates the first unified AI base in China, and builds a hierarchical artificial intelligence system with the collaboration of general and professional models, and realizes the unification of modal representation, task representation, and model structure for the first time. Through this unified learning paradigm, the single M6-OFA model in the Tongyi unified base can simultaneously process more than 10 unimodal and cross-modal tasks such as image description, visual positioning, literal diagram, visual entailment, and document summarization without introducing any new structure, and has reached the international leading level.

In March 2023, Ali Damo Academy has launched the "Text Generation Video Large Model" on ModelScope, an AI model community. At present, the large text generation video model is composed of three sub-networks: text feature extraction, text feature to video hidden space diffusion model, and video hidden space to video visual space, with about 1.7 billion overall model parameters, and currently only supports English input. The diffusion model adopts the Unet3D structure, and realizes the function of video generation by iteratively denoising the process from the pure Gaussian noise video.

腾讯混元大模型

In April 2022, Tencent disclosed the hybrid model for the first time, covering NLP models, CV models, multimodal models and tasks in many fields. The model has advantages and characteristics in advertising content understanding, industry feature mining, copywriting creative generation, etc.

Huawei Pangu model

In April 2021, Huawei released the Pangu Model, which has developed a mature system that includes three stages: the basic model (L0), the industry model (L1), and the industry segmentation scenario model (L2).

The model is based on Pengcheng Cloud Brain II. The automatic hybrid parallel mode with the full-scene AI computing framework MindSpore realizes large-scale distributed training on 2048-card computing power clusters, which is the first time that a domestic full-stack AI infrastructure supports 200 billion-level ultra-large-scale language model training, and achieves performance indicators better than industry SOTA models in 16 downstream tasks.

The Dark Side of the Moon: Tsinghua Department

Moonshot Since 2020, Yang Zhilin, the founder of AI, has been firmly believing that "big models are the future", and in the second half of 2020, he did core technology research and development for Huawei's Pangu large model, and in 2021, he participated in the research and development of KLCII's Wudao large model.

At the end of 2022, Yang Zhilin went to the United States for a two-month investigation, and the American people's enthusiasm for GPT made him even more convinced that this was the right time to set up a large model company, so he completed the first round of financing in the next 1 month, got $60 million in start-up funds from Sequoia and other leading institutions, and set up a technology research and development team of about 40 people within 3 months.

Kimi, the dark side of the moon Chat large model

Kimi, an intelligent assistant based on a 100 billion model Chat, the core competency is the long text ability (Long Context): October 9, 2023, Moonshot AI launched its first product, with up to 200,000 Chinese characters for contextual processing and rapid analysis and summary. In late March 2024, Kimi further increased its contextual processing capabilities to 2 million kanji.

On the C-side, the company is committed to building Kimi into a super application and becoming the entrance to AI-native interaction; On the B side, via Moonshot The AI open platform provides APIs compatible with OpenAI, and during the internal testing period, applications in the fields of law, game reading, and other fields have been tested, and the feedback has been good. Kimi has shown obvious advantages in the field of Chinese against international large models such as GPT-4 and Claude.

Zhipu: Tsinghua KEG was hatched

Zhipu AI was incubated by Tsinghua KEG (Knowledge Engineering Laboratory) led by Professor Tang Jie of the Department of Computer Science of Tsinghua University, and was established in 2019. The vision is to "make machines think like humans in the future", and hopes to benchmark OpenAI; In August 2022, Zhipu AI launched the self-developed large model GLM-130B, which combines GPT's one-way backward prediction model framework with BERT's two-way prediction model framework. The company focuses on ToB and ToG services, and its customers are mainly enterprises and government agencies. In 2023, Zhipu AI will receive financing of 2.5 billion yuan, with a valuation of more than 10 billion yuan.

At present, the cooperation cases of Zhipu include: 1) G-side, cooperating with the Beijing government to provide intelligent customer service products; Cooperate with Beijing Hospital of Traditional Chinese Medicine to build a model for online consultation of traditional Chinese medicine; 2) B-side, cooperate with PetroChina ERP to transform the enterprise workflow; Cooperate with Meituan to dialogue, advertising, search and other functions; 3) On the C side, it provides the entrance of ChatGLM general large model.

Zhipu large model: The products cover the AI open platform, Zhipu Qingyan and a series of AIGC products

Zhipu AI cooperated in the research and development of the large model GLM-130B, and based on this, launched the dialogue model ChatGLM, the open-source single-card version of the model ChatGLM-6B, and launched the AI efficiency assistant Zhipu Qingyan (ChatGLM).

The AIGC model and its product matrix include the high-efficiency code model, CodeGeeX, the multimodal understanding model, the CogVLM model, and the Wensheng graph model, CogView, etc. Establish a large-scale model MaaS open platform and devote itself to the construction of AI ecosystem.

Baichuan Intelligence: Founded by top AI talents in a well-known technology company

Baichuan Intelligence was founded on April 10, 2023 by Wang Xiaochuan, the former CEO of Sogou, and Ru Liyun, the former COO of Sogou, and the core members of the team are mostly top AI talents from well-known technology companies such as Sogou, Baidu, Huawei, Microsoft, Byte, Tencent, etc., as of October 2023, the team size is more than 170 people, of which nearly 70% are employees with master's degree or above, and more than 80% are R&D personnel.

In terms of financing, the start-up capital of 50 million US dollars came from the personal support of founder Wang Xiaochuan and his friends in the industry, and in October 2023, the A1 round of strategic financing was completed, and the financing amount of 300 million US dollars was received from technology giants and top investment institutions such as Alibaba and Tencent, setting a record for the fastest promotion of domestic large-scale model start-ups to unicorns.

In terms of model research and development, 8 self-developed large models have been released in 8 months since its establishment, and four open-source models of Baichuan-7B/13B, Baichuan2-7B/13B and four closed-source models of Baichuan-53B, Baichuan2-53B, Baichuan2-192K and Baichuan2-Turbo have been released successively. And with the national scientific research institution Pengcheng Laboratory, the longest context window model "Pengcheng-Baichuan Mind 33B" based on domestic computing power training was jointly released.