Home About us

For the "100 model war", almost all the bigwigs' tone has changed 180 °?

Intelligent relativity 2024/07/08 13:48

Text | Intelligent relativity

Author | Chen Bocheng

At the 2024 World Artificial Intelligence Conference and the High-level Conference on Global Governance of Artificial Intelligence, Robin Li, founder, chairman and CEO of Baidu, talked about his views on AI models, which shocked everyone.

Intelligent relativity, for the "100 model war", almost all the bigwigs' tone has changed 180 °?

He first pointed out that "the war of 100 models has caused a huge waste of social resources, especially the waste of computing power." But at the same time, it has also built up our ability to catch up with the world's most advanced basic models. ”

Then he emphasized that "without application, there is a basic model, whether it is open source or closed source, it is worthless." At the same time, Robin Li also said that it is necessary to jump out of the thinking logic of the mobile era and avoid falling into the "super application trap", not only the application of 1 billion DAU is called successful.

It can be said that Robin Li's speech was quite intense. This seems to be the first time that a bigwig has spread out the "100 model war" and the development of large models on such a high-level occasion.

Of course, Robin Li is not the only one who holds a similar view.

Zhu Xiaohu, managing partner of GSR Ventures, also mentioned in the dark horse class in June that many entrepreneurs blindly invest in the underlying AI technology. Although it has created a grand occasion of the "100 model war", it has also caused a waste of social resources.

Intelligent relativity, for the "100 model war", almost all the bigwigs' tone has changed 180 °?

He emphasized, "Obviously, the direction of making money in AI entrepreneurship has completely changed. ”

How to change?

In addition to Robin Li and Zhu Xiaohu, there are also Fu Sheng, chairman and CEO of Cheetah Mobile and chairman of Orion Star, Zhang Fan, COO of Zhipu AI, Wang Xiaochuan, founder of Baichuan Intelligence, and other bigwigs have also talked about the direction of competition in the large model industry on different occasions, and the key points that can finally reach a consensus are "scenarios" and "applications".

It seems that the "100 model war" caused by focusing on the "volume" of the basic large model should be stopped, and the focus of the large model is still the application of the "volume" scenario.

The bigwigs reached a consensus on this point. Since the beginning of this year, everyone's tone has changed!

Don't overly compete with the base model, the "volume" scenario is applied

In the past period, the United States has flocked to a large number of startups focusing on the development of large-scale model applications, such as Adept, Stability.ai, Runway, BettrData, Tinybird, UnSkript, and many more.

At the same time, leading companies such as OpenAI and Anthropic, as well as technology giants such as Google and Microsoft, are also committed to using open-source models or self-developed basic models to develop solutions for various application scenarios.

The launch of GPTs and OpenAI's announcement of a series of concessions to developers are aimed at attracting more entrepreneurial teams to participate in the innovation and application of GPT technology, thereby enriching the GPT ecosystem and helping OpenAI to occupy the advantage of large models in the field of scenario applications in the next time.

Intelligent relativity, for the "100 model war", almost all the bigwigs' tone has changed 180 °?

Judging from the trend of the foreign large-scale model industry, the change in the tone of these domestic bigwigs is not groundless.

At present, the average daily call volume of Baidu Wenxin Yiyan has exceeded 500 million, and Baidu officially announced that the daily call volume of Wenxin Yiyan exceeded 200 million just two months ago.

During the period of 2 months, the number of calls has changed so much, which shows that the application of large models to "volume" scenarios is not only promoted by manufacturers, but also the demand of the entire market has been put on the table, showing an explosive growth trend.

A similar signal was unleashed on Alibaba Cloud's home turf.

At the World Artificial Intelligence Conference, Zhou Jingren, CTO of Alibaba Cloud, announced the latest progress of the recent Tongyi model and Alibaba Cloud Bailian platform - in the past two months, the number of downloads of Tongyi Qianwen open source model has increased by 2 times, exceeding 20 million times, and the number of Alibaba Cloud Bailian service customers has increased from 90,000 to 230,000, an increase of more than 150%.

Intelligent relativity, for the "100 model war", almost all the bigwigs' tone has changed 180 °?

When it comes to large models, compared with the comparison of parameters, the domestic bigwigs seem to be more willing to tell the market how easy to use their own large models, how many people use them, how to use them next, and a series of things related to the implementation of scene applications.

Investors represented by Zhu Xiaohu have also begun to look for investment opportunities in large models at the application layer.

The direction of the market has changed, not only the tone of the bigwigs is changing.

Where is the application of "super capable"?

"In the era of AI, 'super capable' applications are more important than 'super apps' that only look at DAUs." At the World Artificial Intelligence Conference, Robin Li tried to draw conclusions for the next trend of large-scale model application development.

However, "super-capable" apps may not be difficult to understand, but the unanswered question in the market is how such apps are developed and how they are brought to the masses.

Based on the current performance of the industry, "Intelligent Relativity" believes that there are several considerations worth exploring.

1. Behind the application of "super capable", the iteration and adaptation of large model technology is necessary.

Most of the industry trends are in the same direction, and the iterative trend of MoE architecture in the field of large models since the beginning of this year represents that in terms of technology, large models are supporting the application of "volume" scenarios.

Nowadays, OpenAI's GPT-4, Google's Gemini, Mistral AI's Mistral, xAI's Grok-1, Kunlun Wanwei's Tiangong AI, Inspur Information's Source 2.0-M32, and Tongyi Qianwen's Qwen1.5-MoE-A2.7B have all adopted MoE architecture.

The MoE architecture realizes the sparseness and modularization of the model by introducing Expert Networks and Gating Mechanism, and has quite good feedback in terms of data processing, computing power resource allocation, and output result optimization. This provides a very key technical support for the implementation and promotion of large-scale model scenarios.

For example, Microsoft has proposed an end-to-end MoE training and inference solution, DeepSpeed-MoE, which deeply optimizes the communication of MoE in parallel training, reduces communication overhead, and achieves efficient model parallelism. In addition, DeepSpeed-MoE also proposes an expert ranking mechanism based on fine-tuning, which can dynamically adjust the allocation of input samples to experts according to the loss of experts in the training process to improve the effect.

Second, the application of "super capable" means a more commercial ecological competition.

The technology is fine, but the unclear path to commercialization will still face a collapse in today's market. A few days ago, Microsoft's official website updated a notice - "GPT Builder is about to be discontinued". GPTs, which once set off countless heated discussions and climaxes in the AI circle, seem to be heading for failure.

Who would have thought that the press conference when the concept of GPTs was born was described by the outside world as "OpenAI's iPhone moment".

OpenAI originally wanted to use low-threshold technical capabilities and global developers to jointly create a number of "super capable" applications, but due to the experience flaws caused by technical problems and vague monetization policies, the commercialization path of the concept of GPTs has never been successful, and it can only be "cool" in the end.

Most of the "super capable" applications are built on mature business ecosystems, and perhaps AI vendors around the world need to recognize this. It is worth mentioning that on the other side of the ocean, the open-source AI model community led by Alibaba Cloud in November 2022 has just won the 2024 SAIL Star Award.

After more than a year of development, the community has become the largest and most active AI model community in China, bringing together more than 5,500 high-quality models and thousands of datasets, providing models and free computing services for more than 5.6 million developers. Perhaps, the ecological path that OpenAI failed to take will have new vitality in China.

3. The application of "super capable" will inevitably sprout in the industry scene.

Zhu Xiaohu's advice to large-scale model entrepreneurs, "Don't be superstitious about AI, focus on the sharp knife scene and land as soon as possible." Scenarios are the cradle for incubating "super capable" applications, but in more depth, we can't just look at the scenes, but ultimately we have to look at user feedback and value presentation.

Industries such as healthcare, education, finance, manufacturing, transportation, agriculture, etc., are "high-incidence" scenarios for large-scale model applications, but the created agents or solutions are just "like people drinking water, knowing their own warmth and coldness".

To B's project looks at efficiency. In the field of express delivery, at present, through the large model to help process orders, it is possible to achieve "a picture, a sentence to send express", no longer need other cumbersome processes, the time is shortened from more than 3 minutes to 19 seconds. And more than 90% of the after-sales problems are also solved by large models. ——This kind of efficiency improvement can be called "super capable".

To C's scenario depends on the user. Previously, at its peak, Baidu's gaokao agent had to answer more than two million questions a day. For the 10 million candidates nationwide, this percentage is quite high. - Such a number of users can also be regarded as "super capable".

Today, large model applications cover various general and vertical scenarios such as text generation, data processing, PPT production, marketing, customer service after-sales, and medical diagnosis. In fact, there is no shortage of scenarios in the market, but a lack of capable and effective applications, and "volume" applications must find users and value in the scene.

*The pictures in this article are all from the Internet

This content is original to [Intelligent Relativity],

It only represents a personal opinion and is not allowed to be used by anyone in any way, including reprinting, excerpting, copying or mirroring without authorization.

Some of the pictures are from the Internet, and the copyright ownership has not been verified, not for commercial use, if there is any infringement, please contact us.

This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com