360+ Wisdom AI, the Chinese version of 'Microsoft + OpenAI” is here

In the research and development of core technologies for large models, 360 adopts independent research and development and cooperates to develop "dual-engine" drives to enhance the competitiveness of the core "engines".

Author/Luo Zhi Nanming

produced/newly excerpted from business reviews

The success of ChatGPT is by no means due to its own efforts. Behind this fastest-growing application in history, the collaborator Microsoft plays a vital role. It is the cooperation between industry and research that has contributed to the subversiveness of large language model technology. change. China has set off a big model war. However, after several months of hustle and bustle, no company has embarked on this proven path. Whether the Chinese version of "Microsoft OpenAI" will appear is the focus of the industry.

Today, 360 Company and Zhipu AI announced a strategic cooperation and finally announced the cooperation plan. The "360GLM" large-scale model jointly developed by the two parties has reached the level of a new generation of cognitive intelligence general model.

1. The era of large models in China is coming

The release of ChatGPT triggered a big change. Microsoft announced that ChatGPT was connected to the Office family bucket, which brought disruptive innovation in applications and shocked the whole world.

Looking back at China, after the release and application of ChatGPT, major manufacturers quickly followed up. Baidu’s Wen Xinyiyan, Alibaba’s Tongyi Qianwen, iFlytek’s Xinghuo, etc. all continued to have success in the field of large models. New breakthrough.

At the same time, major manufacturers have also recognized the urgency of integrating large-scale model technology into practical applications. Alibaba Cloud announced that all Alibaba products will be fully upgraded by accessing the "Tongyi Qianwen" large model in the future; iFlytek's Spark large model has core capabilities such as text generation, knowledge question and answer, and mathematical capabilities.

The reality is that although many companies have launched large models, the capabilities of each large model are almost the same. No one can be far ahead and dominate the world. But what is certain is that when everyone is "painting the cake", Landing is the key.

There is a consensus in the industry that the decisive factor in winning is the "scenario". Only companies with scenarios can occupy a leading position in the future large-model industry competition.

Large model technology itself is no longer a threshold. How to better combine industry and research and how to implement it better are issues that all players entering the large model industry need to think about further.

Obviously, it is easier to achieve success by following a proven path.

The cooperation between 360 and Zhipu AI can be regarded as the Chinese version of "Microsoft OpenAI".

If there are only industrial companies, there will be a lack of long-term ideals of Open AI scientists, and no one may think of going further on the road of encoding all human knowledge in large models.

The same is true in reverse. If there is only Open AI, there is no point in publishing papers continuously. After all, if papers and technologies cannot be combined with users' scenarios, they will not be successful.

Microsoft and Open AI have a very good division of labor: Microsoft bundles Open AI services into its products, allowing Open AI services to truly be engineered, scenario-based, product-based, and even commercialized.

Open AI is not necessarily good at commercialization, nor is it necessarily good at product and scenario development, but it is constantly moving from GPT3.5 to GPT4.0 in terms of core technology. Maybe GPT5 is already on the way.

Abroad, a successful combination of industry and research has been pioneered, and 360 has no reason to refuse to imitate this successful path.

2. “China GPT Combination” of technical scenarios

Zhipu AI is a company transformed from the technical achievements of the Department of Computer Science of Tsinghua University. It was rated as the company with the "most OpenAI temperament and level" in China. In November last year, the Stanford University Large Model Center conducted a survey of 30 mainstream large-scale enterprises around the world. The model was evaluated in an all-round way, and GLM-130B, a bilingual 100-billion-level ultra-large-scale pre-training model developed by Zhipu AI, was the only large model selected in Asia.

Professional evaluation results show that the key indicators such as the accuracy of GLM-130B are close to or the same as large models from companies such as OpenAI, Google Brain, Microsoft and Nvidia. More than 1,000 institutions in 70 countries around the world have applied for use.

As a leading large-model technology company, Zhipu AI is in the leading position in catching up with OpenAI. However, simply being technically competent is not enough. The more important achievement lies in its application in real-life scenarios.

This cooperation between 360 and Zhipu AI is an important practice in combining large models with scenes. Zhou Hongyi, founder of 360, has always emphasized the importance of scenarios in the development of large models. He once said that if technology cannot be combined with users' scenarios, it will not be successful.

As ChatGPT’s impact on mankind continues to deepen, more and more domestic companies are breaking into the field of large models. In this regard, Zhou Hongyi said: The most important thing is not that you have large model technology, but how to learn from Microsoft and OpenAI's cooperation model explains how to get better data for training, how to implement better engineering training methods, and how to find better user scenarios to use your large model.

It is understood that the average MAU of 360 Browser is 416 million, and the average daily request volume of 360 Search exceeds 1 billion, which lays the foundation for the subsequent traffic operation, user feedback, and model optimization of the large model.

360’s full-end application that is comparable to Microsoft’s “Family Bucket” provides technology opportunities for Zhipu AI. At the same time, through the data of the full-end application, it can also help large model technology to provide data support. For large models, more data support and scene information feedback are also an important step in assisting their improvement. Technology and data work together to create a flywheel effect.

To a certain extent, the "volume" of large models is not just technology, but also the combination of technology and application.

The cooperation model between Microsoft and OpenAI has proven to be feasible, proving this point. The implementation of domestic large-scale model technology requires a more comprehensive consideration of its application and should not be limited to a certain field or scenario.

With user scenarios that rival Microsoft, the cooperation between 360 and Zhipu AI will become the best example of the Chinese version of the "Microsoft OpenAI" model, occupying a certain advantage in the competition in the large model industry. Both parties The strong alliance in technology and the complementarity of technology and scenarios can elevate the competition in the domestic large model industry to a whole new dimension.

3. AI era, two lines of 360

360’s artificial intelligence strategy is to “fly on both sides”. On the one hand, it develops core technologies, and on the other hand, it seizes user scenarios.

In order to improve the competitiveness of the core "engine", 360 has adopted a "dual-engine" driving strategy of independent research and development and cooperative research and development.

360 has always been known for its security leadership, and has been in the first echelon of domestic leading artificial intelligence technology, with continuous accumulation. As early as 2015, 360 established an Artificial Intelligence Research Institute to apply AI cutting-edge technologies such as CV and NLP to IoT, security big data and other fields. At the same time, it undertook the Safe Brain-National Artificial Intelligence Open Platform Project.

Thanks to its domestically advanced multi-modal R&D team, 360 has achieved industry leadership in natural language understanding, machine vision and motion, and speech semantic interaction. Its core members and team have won multiple AI-related competition championships/nomination awards. .

In terms of large model-related technologies, 360 solves cross-modal retrieval through multi-modal learning and improves search efficiency. Then, it uses transfer learning technology to quickly implement text classification, information retrieval, human-computer dialogue, NER, Summary and other capabilities. Conducted Chinese language model pre-training research.

Later, 360 began to study the cross-modality of images and texts, and selected 23 million high-quality Chinese images and texts from 5 billion images and texts for open source cross-modal data sets. Subsequent research on CV, AIGC and MLLM gave 360 ​​a more solid foundation for its early preparations on large models.

On CV, OVD opened the world target detection large model Zero/Few Shot; on AIGC, the Chinese Stable Diffusion model tried to solve 4 billion image-text pairs to solve the problem of Chinese understanding; on MLLM, it has native vision and language The multi-modal GPT of capabilities realizes cross-modal knowledge transfer through domain alignment.

Taken together, 360’s independent research and development has gone through processes such as natural language understanding, machine vision and motion, and multi-modal large language models, and has become increasingly sophisticated.

Based on AI core technology, 360 leverages its own advantageous scenarios to implement "four-way concurrency" large-scale models, providing large-scale model-related products and services to consumers, small and medium-sized enterprises, industries, government enterprises and cities.

The product matrix of 360GPT is 360 Intelligent Brain. The 360 ​​Intelligent Brain panorama includes 360CV large model, 360GPT large model, 3 60GLM large model, and 360 multi-modal large model.

Zhou Hongyi has a clear plan for the implementation of his own large model. After "360 Intelligent Brain" takes the lead in the search scene, the smart browser, AI mapping tool and enterprise intelligent marketing cloud combined with GPT capabilities will soon be available to B-end users. Open for testing. Obviously in terms of strategy, 360 is not in a hurry for success, but focuses on the rigid needs of users, matures a scenario, and opens up a scenario.

Zhou Hongyi has emphasized many times that China's large-scale models have lagged behind GPT for about two years in terms of development. To promote their development, users' support, understanding and tolerance are needed. In terms of some issues and skills, Chinese players may catch up from behind. Some abilities, such as strong reasoning ability, chain of thought, and emergent abilities, require a certain process to be developed.

It is true that the development of GPT in China is not mature enough, but who knows how the final winner will emerge, and will this player be 360 ​​who copied the successful model of "Microsoft OpenAI"? We'll see.

The above is the detailed content of 360+ Wisdom AI, the Chinese version of 'Microsoft + OpenAI” is here. For more information, please follow other related articles on the PHP Chinese website!