'Zhipu AI launches the Chinese version of GPTs' new generation large-scale model GLM-4, triggering ecological changes?'

Image source: Zhipu AI public account

Zhipu AI demonstrated its new generation base model GLM-4 and Wenshengtu product CogView3 at the first Technology Open Day as a demonstration of its strength in catching up with Chat GPT. In addition, Zhipu AI has also launched GLMs personalized agent customization capabilities and an agent center. These new functions are aligned with GPTs and its sharing platform GPT Store launched by the American artificial intelligence company OpenAI last year. Although the latter was delayed due to high-level turmoil, the new functions of Zhipu AI have been officially launched on January 10, 2024 local time. These new features will provide users with a more convenient and personalized experience in using intelligent agents, helping Zhipu AI gain a greater competitive advantage in the future artificial intelligence market.

In terms of base large model products, Zhang Peng, CEO of Zhipu AI, said: "In terms of standard large model evaluation, our GLM-4 is close to the GPT-4 level." Compared with the previous generation, The performance of GLM-4 has been improved by nearly 60%, and it has longer context processing capabilities, stronger multi-modal support, and faster inference speed. ”

Specifically, GLM-4 can support a text length of 128k for a single prompt word, which is equivalent to processing 300 pages of text. Even in the face of such a huge amount of text, GLM-4 can achieve almost 100% precision recall rate in the Neddle test (needle in a haystack) test. In addition, Zhipu AI has also launched the third generation of CogView, an image generation AI product with powerful multi-modal generation capabilities. The performance of this product is close to that of DALL·E 3 released by OpenAI.

Zhang Peng compared GLM-4 and GPT models and found that GLM-4 performed better than GPT-4 in many capabilities. According to the official data of Zhipu AI, GLM-4 scores higher than GPT-4 in Chinese alignment ability, and also shows higher scores in Chinese professional ability, Chinese understanding ability and role playing. In addition, GLM-4 performed better than GPT-3.5 in benchmark tests such as multi-task language understanding, prompt word following, and instruction following, as well as in tests using multiple data sets such as GSM8K and Math. These results deserve our attention.

Image source: Zhipu AI’s first technology open day live broadcast

In addition to the improvement in single performance, GLM-4 also has the All Tools capability-a single command can mobilize various tools. Thanks to the powerful Agent capabilities of the GLM model, GLM-4 has the ability to automatically understand and plan complex instructions based on user intentions, and freely call WebGLM search, Code Interpreter code interpreter and multi-modal formation capabilities. GLM-4's All Tools capability can handle various types of tasks, including data analysis, chart drawing, PPT generation and other common file processing, aiming to solve the pain point of complex prompt words in an automated way. For example, let GLM-4 present global GDP data for the past ten years. The large model can be networked to retrieve multiple data sources and visualize the found data into a simple chart. This process covers the search extraction code interpreter and Drawing multiple tools.

At present, the AI ​​large model competition has crossed over from the "PGC" era to the "UGC" era. Zhipu AI will continue to develop a new generation of base large models and is committed to establishing a large model ecosystem to provide developers with better services and technical support.

This ambition has been implemented into GLMs' personalized intelligent agent customization capabilities. GLMs benchmarks OpenAI's GPTs launched in November last year, allowing users to create their own GLM intelligent dialogue robots with simple prompt words. At the same time, Zhipu AI also launched the Agent Center, allowing users to share the agent GLM models they created. This move is similar to OpenAI’s GPT Store, which will be launched on January 10, 2024 local time.

It has been more than three years since Zhipu AI was registered and established. Behind this company, which is known as the “most OpenAI” company in China, are major investment institutions and technology giants. Tianyancha APP shows that Zhipu AI has currently received 8 rounds of financing exceeding 2.5 billion yuan, of which 5 rounds of financing occurred in the second half of 2023. On September 19, 2023, Tencent, Alibaba and Hillhouse Capital became the Series B investors of Zhipu AI. This 1.2 billion yuan transaction was the largest financing round for Zhipu AI. On August 1, 2023, Ant Group and Capital Today invested RMB 440 million in Zhipu AI. On July 17, 2023, Meituan invested 300 million yuan in Zhipu AI.

Zhipu AI has also achieved certain results in commercialization. Zhipu AI has proposed the commercial implementation path of MaaS (Model as a Service) in the large model market, and provides a variety of business solutions such as cloud API, cloud privatization and local privatization according to different customer group types and needs. At present, Zhipu AI has more than 2,000 partners and conducts in-depth model co-creation and co-construction with more than 200 units, covering finance, Internet, energy manufacturing and other industries.

The above is the detailed content of 'Zhipu AI launches the Chinese version of GPTs' new generation large-scale model GLM-4, triggering ecological changes?'. For more information, please follow other related articles on the PHP Chinese website!