Written by | Zhang Yu Edited by | Yang Bocheng
"Catching up with OpenAI" and "Benchmarking against Open AI has been the goal of Zhipu AI since its establishment" are a few sentences that Zhipu AI CEO Zhang Peng mentioned repeatedly when sharing with the outside world. However, now, the difficulty of Zhipu AI catching up with OpenAI has increased a lot.
After OpenAI released its new flagship language model GPT-4o at its spring conference, Google also launched a series of "AI super buckets" based on Gemini at the I/O Developer Conference. Although Zhipu AI is considered one of the most likely large-scale model companies to become the "Chinese OpenAI", as the "hundred-model war" at home and abroad intensifies, the situation Zhipu AI faces is not encouraging.
Take Zhipu Qingyan, a generative AI assistant under Zhipu AI, as an example. This is a large-model product developed based on the ChatGLM2 artificial intelligence language model. Through pre-training of trillions of characters of text and code, combined with supervised fine-tuning technology, it has the capabilities of general question and answer, multi-round conversations, creative writing, code generation, virtual conversations, AI drawing, document and image interpretation, etc.
However, compared with ChatGPT developed by OpenAI based on the GPT-4 artificial intelligence language model, Zhipu Qingyan is still at a disadvantage. The most direct gap is that ChatGLM2 is inferior to GPT-4 in terms of training data and diversity, cross-language capabilities, model size and parameters, and reasoning capabilities.
In order to make up for the gap with OpenAI, Zhipu AI released a new generation of language model GLM-4 at the first Technology Open Day (Zhipu DevDay) held in January 2024. Although the overall performance of GLM-4 has been greatly improved by 60% compared with the previous generation, it is known as "on par with GPT-4", but in fact it only reached about 90% of the level of GPT-4.
It is worth mentioning that the newly released GPT-4o by OpenAI has increased the processing speed by 200%. Based on GPT-4o, OpenAI has updated and upgraded ChatGPT, adding stronger voice and visual functions, which greatly enhances ChatGPT's perception of reality - GPT-4o can respond to audio input in as little as 0.23 seconds, and the audio response speed has reached a level similar to that of humans.
Obviously, catching up with GPT-4 is already Zhipu AI’s most urgent task at present. Whether it can train a language model that is truly comparable to GPT-4 is particularly critical for commercialization and ecological advancement.
01. There is still a big gap with OpenAI
Zhipu AI was founded in June 2019 and was transformed from the technological achievements of the Knowledge Engineering Laboratory (KEG) of Tsinghua University. It is currently the only fully domestically-funded and fully self-developed large-model enterprise in China. As early as 2020, it began the research and development of the GLM pre-training architecture and trained the 10 billion parameter model GLM-10B; in 2022, it cooperated to develop the 130 billion-level ultra-large-scale pre-trained general model GLM-130B; in 2023, Zhipu AI launched the GLM series of 100 billion open source base dialogue models, and launched GLM-4 in January 2024.
Since Zhipu AI entered the market early and directly benchmarked OpenAI, it has become a darling in the eyes of capital institutions. From July to September 2023, Zhipu AI received five rounds of financing, with a total amount of more than 2.5 billion yuan. The main investment institutions include Zhongguancun Independent Innovation Fund, Meituan Strategic Investment Department, Ant Group, Alibaba, Tencent Investment, Hillhouse Capital, Sequoia Capital, etc. The current valuation has reached 20 billion yuan, making it one of the "unicorn" companies in the domestic AI field.
However, there is still a big gap between Zhipu AI and OpenAI.
From the perspective of technical route, OpenAI pays more attention to versatility, portability and scalability. Its GPT series language models can be applied in multiple scenarios and are highly customizable. In contrast, Zhipu AI's technical route is "big model + small model", which adapts to the needs of different scenarios and tasks through pre-training and fine-tuning of large models. This technical route can improve the generalization ability and application scope of the model, but there are also problems such as high model complexity, large amount of calculation, and long training time.
In terms of model scale, OpenAI's GPT series language models are larger in scale and can process large amounts of natural language data, thereby achieving better model performance, while Zhipu AI's models may be smaller in scale and have limited data processing capabilities, which may affect its model performance and generalization capabilities; in terms of data resources, OpenAI has a large amount of natural language data resources that can be used to train and optimize its models, while Zhipu AI's data resources may be relatively small, resulting in limitations on the effectiveness and performance of its model training.
The direct difference between the two is reflected in the number of users. In November 2022, OpenAI's ChatGPT had more than one million users in just five days. In January 2023, its monthly active users exceeded 100 million, becoming the fastest growing consumer application in history. In contrast, according to institutional estimates, as of November 2023, the daily active user number of Zhipu Qingyan, a subsidiary of Zhipu AI, ranged from 100,000 to 400,000.
It is worth mentioning that Zhipu Qingyan is not even a rival of Wenxin Yiyan, a subsidiary of Baidu. As of November 2023, the number of daily active users of Wenxin Yiyan was about 800,000, and as of April 2024, the number of users of Wenxin Yiyan had exceeded 200 million.
Zhang Peng also admitted that compared with foreign large models, the development of domestic large models started a little late. Coupled with the limitations of high-performance computing power and the gap in data quality, domestic large models have a certain gap with the world's advanced level in terms of scale and core capabilities. This gap is about one year.
02. Zhipu AI has a long way to go in commercialization
How to commercialize big models is a huge challenge facing all big model companies, including Zhipu AI.
Zhipu AI is one of the earliest large-model companies in China to promote commercialization, and was the first to propose the implementation path of "Model as a Service (MaaS)". At present, Zhipu AI has explored four business models: one is lightweight, which encapsulates large models into an open platform and provides APIs (application programming interfaces) to developers, enterprises, etc. for calling, and pays according to the number of calls. This method is very simple and mature, and is not much different from that abroad; the second is to meet the needs of some medium and large enterprises for data security protection. Zhipu AI provides cloud-based private deployment solutions, and helps users open up special model areas based on cloud computing power; the third is a completely private deployment, which provides the base capabilities of large models on the company's own hardware and computing power platform to meet the needs of enterprises to develop related applications and business development; the fourth is a combination of software and hardware solution, which adapts and binds large models to domestic information and innovation hardware, which can eliminate the process of deployment and debugging in the customer environment, and sell and deploy through software and hardware integration.
Zhang Peng believes that the willingness to pay on the B-side is much better than that on the C-side, especially for the leading companies in the industry, which generally invest more and act faster in large models. Therefore, Zhipu AI has been targeting the B-side from the beginning. As of 2023, Zhipu AI has more than 2,000 ecological partners and more than 1,000 large-scale models. In addition, more than 200 companies have conducted in-depth co-creation with Zhipu AI, covering multiple leading companies in multiple fields such as media, consulting, consumption, finance, new energy, Internet, and smart office. Zhang Peng revealed that Zhipu AI will have about hundreds of signed customers in the second half of 2023, and the overall signed orders in 2023 will be in the billions.
However, before Zhipu AI was commercialized, the large model industry started a "price war".
On May 15, ByteDance announced that the price of Doubao's main model (Doubao Universal Model Pro) in the enterprise market is 0.0008 yuan/thousand tokens, while the price of the same specification model on the market is generally 0.12 yuan/thousand tokens, which is 150 times the price of the Doubao model. On May 14, OpenAI announced at its spring conference that the price of GPT-4o's API has dropped by 50%.
Prior to this, French AI company Mistral AI released its latest large model Mistral Large and launched Le Chat, the first chatbot benchmarked against ChatGPT. In terms of pricing, the input and output prices of Mistral Large are about 20% cheaper than GPT-4 Turbo; the API pricing of the new second-generation MoE large model DeepSeek-V2 released by DeepSeek, an AI company under Magic Square Quant, is 1 yuan per million tokens input and 2 yuan per output (32K context), which is nearly 1% of the price of GPT-4 Turbo.
Faced with the "price war", Zhipu AI also announced that the calling price of its entry-level product GLM-3 Turbo model will be reduced from 5 yuan/million tokens to 1 yuan/million tokens, a drop of 80%.
It is worth mentioning that the continued decline in pricing of large models is expected to bring about faster commercialization, but at the same time, "price wars" often mean that companies need to make concessions on prices. For Zhipu AI, its own profitability is limited. If a "price war" is carried out again, it may lead to further declines in profits, and achieving profitability will become more difficult.
Zhang Peng also admitted that the challenges faced by Zhipu AI in 2024 are very daunting. On the one hand, OpenAI will achieve new breakthroughs in super cognition and super alignment technologies in 2024, which requires Zhipu AI to continuously iterate its technology and keep up with the world's leading pace; on the other hand, large models will usher in a wave of commercialization in 2024, and Zhipu AI's commercial competition pressure will also increase.
For Zhipu AI, the current commercialization path is already relatively clear, but whether it can successfully commercialize large models depends not only on the exploration and attempt of business models, but also on solving the underlying problems of large model development.
03. Keep each other warm
Open source is also a major feature of Zhipu AI. Zhipu AI hopes to create a prosperous community and ecosystem to further promote the development of the big model open source community.
Zhang Peng once said that Zhipu AI is one of the few leading technology companies in China that can compete with OpenAI. The company hopes to work together with all participants in the AI field, including upstream and downstream partners in the industry chain, developer communities, and academia, to contribute to the new future of AI in China.
In 2024, Zhipu AI will launch an open source Big Model Open Source Fund, which includes three "1000s": Zhipu AI will provide 1,000 computing cards to the Big Model open source community to help open source development; provide 10 million yuan in cash to support open source projects related to Big Models; and provide 100 billion free API tokens to outstanding open source developers. Zhang Peng said that the purpose of the Big Model Open Source Fund is to promote the great progress of Big Model research and development and promote the great prosperity of the entire open source ecosystem of Big Models.
Facing the global big model entrepreneurs, Zhipu AI will upgrade the "Z Plan" and jointly launch a big model entrepreneurship fund with ecological partners with a total amount of 1 billion yuan to support original innovation of big models, covering big model algorithms, underlying operators, chip optimization, industry big models and super applications.
In addition, Zhipu AI has also invested in many AI industry chain companies. As of now, Zhipu AI has invested in and acquired shares in more than 10 start-ups through industrial investment, with the scale of single investment at the million yuan level. These include the AI model layer company "Lingxin Intelligence", the AI model layer company "Mianbi Intelligence", the intelligent legal service product provider "Milu Intelligence", the enterprise large model service platform "Zhiyi Bi", the software and information technology service provider "Shudao Zhisuan", the generative AI application provider "Shengshu Technology", and so on.
Zhang Peng is very optimistic about the future of Zhipu AI. "In 2024, the big model market will return to calm from wild growth, the investment and hype of big models will come to an end, and the industry focus will shift from the model itself to finding applications. However, this does not mean that the technological evolution speed of big models will slow down, and the ceiling of upward exploration is far from being reached."
This article comes from the WeChat public account "DoNews" (ID: ilovedonews) , author: Zhang Yu, and is authorized to be published by 36Kr.
