Image source: Generated by Unbounded AI

Zhixixi reported on November 28 that just yesterday, Orion Star officially released the open source Orion-MoE 8x7B parameter MoE model, and jointly released a large model data service - AI data with Juyun Technology precious.

Han Kun, chief scientist of Orion Star, introduced that AI Data Bao AirDS (AI-Ready Data Service) can provide enterprises with a full range of large model data services, covering data collection, cleaning, annotation, and prompt word engineering. and evaluation and other comprehensive aspects. AI data treasure is an important bridge between the underlying model and upper-layer applications.

Data, algorithms, and computing power have always been indispensable for the development of large models. Nowadays, the gap between algorithms and computing power has narrowed significantly, and the importance of data has become increasingly prominent. Compared with algorithms and computing power, sufficient quantity and high-quality data are the key to large-scale model effects and application development, and are also the core of AI application effects that can widen the gap. In the large-model business closed loop, data has obviously become the key that most directly affects its implementation in vertical industries.

Therefore, on the occasion of the release of the AI ​​​​data treasure, Fu Sheng, chairman and CEO of Cheetah Mobile and chairman of Orion Star, mentioned in interviews with media such as Zhixixi that piercing the window paper of the AI ​​​​industry will lead to hundreds of models. The battle relies on data, and data is the key to winning in the implementation of industry scenarios.

In this context, relying on Orion Star’s large-scale model capabilities, comprehensive capabilities in data collection, annotation, and prompt word engineering, as well as Juyun Technology’s understanding of the scenario needs of China’s overseas enterprises, it has become the only company in the industry currently A company that not only builds large models but also opens up large model data services.

This is of great significance to the current development of the large-scale model industry in terms of technology, ecology and many other layout aspects.

01. "Refining elixirs" is easy but "cultivating immortals" is difficult. Data is the key to success in scene implementation

From last year to now, based on the revolutionary neural network Transformer The ChatGPT architecture detonated the AI ​​industry, and then the era of computing power came, where those with computing power conquered the world. NVIDIA GPUs were snapped up... The fierce competition in algorithms and computing power has slowed down.

With the upgrading of competition in the large model industry and the acceleration of application implementation, everyone is choosing the same card at the computing power level. At the algorithm level, most companies will choose the mature Transformer architecture. These two carriages can no longer become enterprises. The key to widening the gap is no longer as important as before. On the other hand, data has become the key to victory for all princes.

A key topic behind this is: "Refining elixirs" is easier than "cultivating immortality".

More diverse AI applications have emerged and have shown their value in various industries, but this is just the tip of the iceberg in the development of large models. It is not easy for large models to truly maximize their value in all walks of life and for enterprises to make good use of large models to reduce costs and increase efficiency. The amount and quality of data are as follows:What is the key to whether an enterprise can build a good AI application.

But is it enough to just have data? The answer is no. When enterprises choose AI applications, the most critical thing is that they are error-free and can significantly improve business efficiency. But as Tong Ning, vice president of Cheetah Mobile, said, early companies did not find a suitable path when developing large-model applications. They could only see the high ranking and good reputation of the model, but could not gain insight into the application development process under the iceberg. of many problems.

These problems are often related to the specific effects of large models in-depth into enterprise business, such as whether the data is accurate and real, whether the diversity of the data is sufficient, and whether the prompt words have been optimized.

The road to "cultivating immortality" under the iceberg is long and arduous. Enterprises need to clean and label data, fine-tune and strengthen models, and process large amounts of multi-modal data such as text, pictures, videos, audios and even 3D data. Fast processing is closely related to avoiding the illusion of large models and breaking through the accuracy of large model recognition and understanding. Fu Sheng believes that large model data services are the key to determining the basic capabilities of large models in the industry chain. They need to be highly integrated with applications to find high-quality data.

This is a problem that must be solved for enterprises, but many enterprises currently have barriers to data processing, and the data processing tools currently on the market also have their own pros and cons. Therefore, what kind of data services can Connecting the base capabilities of large models with useful applications is a major problem facing companies that develop large model applications.

02. Do both large models and application development to form a closed loop of models, data and business

Since this year, the war of hundreds of models has come to an end, and AI applications have become a battleground for all companies.

So, who is the best solution to provide data services? What kind of enterprise can connect models, business and data?

We can start with today’s new release of Orion Sky.

Orion Star and Juyun Technology jointly released the AI ​​data treasure AirDS, which provides a complete set of services around data, including data cleaning, data annotation, prompt word engineering, how to evaluate models, etc., allowing enterprises to Quickly build useful applications with large models.

In fact, the data service track is not an emerging field. At present, the industry has formed a tripartite situation of technology giant companies, professional basic data service providers, and start-up technology companies. But for current enterprise data services, these three types of enterprises have their own advantages and disadvantages.

Previously, the "AI Basic Data Service White Paper" released by Deloitte Consulting, a well-known municipal research institution, mentioned that traditional professional basic data service providers are an important part of the industry, and technology giants rely on their technological strength and Strong resources gradually occupy a competitive advantage.

Among them, technology giant companies have automated annotation, professional data collection and standardization and comprehensiveStack service capabilities have the strongest comprehensive capabilities, but these services are not fully open, and some are limited to customers of these giant companies; professional basic data service providers have an early layout, have accumulated deep service experience, and occupy a large share of the market. Its biggest advantage is low-cost human services, but compared with AI annotation tools, human services currently have no advantage in terms of cost and efficiency; technology startups focus on entering the market through automated annotation and AI annotation tools to reduce labor costs, but they are relatively Compared with giant players, its customer resources are not sufficient.

Behind this, the combination of Orion Star and Juyun Technology can effectively link the advantages of the two and avoid the shortcomings of different types of enterprises.

Compared with large model companies and traditional data annotation companies, AI Databao AirDS has the capabilities of large model research and development, large model data services, industry services, and AI application development into a system. Tong Ning said that Orion Star is engaged in large model research and development and provides large model data services. At the same time, it has carried out AI application development and delivery in the industry since last year. Juyun Technology has long served Chinese brand companies going overseas, so it has It has full-chain end-to-end capabilities, so it not only has data annotation services that combine AI and manual work, but also has a certain amount of customer resources.

In this way, models, data, and business form a closed loop, and the commercialization of Orion Star’s AI data treasure has been completed.

At present, AI data treasure AirDS has been applied to enterprises in mobile communication terminals, Internet entertainment, new energy vehicles, Internet money, consumer retail and other fields. It can serve diversified types of Chinese brand overseas enterprises.

For example, a global mobile terminal customer solves the problem of language adaptation for localized scenarios based on AI data treasure AirDS+ multi-language. AirDS completes the development and testing platform by collecting data from multiple scenarios and covering more than 20 languages. , after optimizing the prompt word project, the accuracy of the company's relevant evaluation index results exceeded 95%

It can be seen that how large models realize commercial value is a key proposition for current industry development. Orion Star has already Be the first to find a feasible path.

03. Aggregated AI technology + overseas service advantages highlight the integration advantages of Cheetah Mobile

There are two questions behind Orion Star’s release of AI data treasure and its first run-through commercialization: why Orion Star can do it, and why Orion Star did it first.

It boils down to Orion Star’s focus and persistence in the AI ​​industry and Juyun Technology’s deep insight into customers’ overseas needs.

On the one hand, Orion Star has firmly developed its own full-chain AI technology since its establishment in 2016. Han Kun, chief scientist of Orion Star, said that from the beginning, Bao Xiaomi’s intelligent voice interaction system, laser and visual multi-mode ecological system, and then to the Lucky Leopard intelligent indoor navigation system. Currently, LieTohoshizora is also conducting research on embodied intelligence.

After that, ChatGPT became popular at the end of 2021. Orion Star quickly cut in based on its many years of AI technology reserves, providing customers with services such as AI applications and model fine-tuning. Subsequently, in mid-2023, the company embarked on the road of self-developed large models and trained from scratch the open source tens-billion-parameter model Orion-14B "born for enterprise applications" released at the beginning of this year.

This year, in order to meet customer demands for fast model speed and good effects, Orion Star chose the MoE route and launched the Orion-MoE 8x7B-Base model today.

The total parameters of the Orion-MoE 8x7B model are 48B, and the activation parameters for each task execution are 14B. A comparison of the results of the main Chinese and English evaluation sets shows that the Orion-MoE 8x7B model's performance in Japanese, Korean, Spanish and other multi-lingual abilities is overall better than the Mixtral-8x7B equivalent level parameter model.

In terms of inference speed, compared with dense models with similar effects, Orion-MoE 8x7B's different GPUs and different concurrency speeds can be improved by 20%-30% compared to models with the same level of parameters. At the same time, this model has been completely open source and has been launched on GitHub, Hugging Face and other platforms.

On the other hand, Juyun Technology was founded in 2020. Its predecessor was the IT operation and maintenance service department of Cheetah Mobile during the overseas 1.0 period. It has more than 10 years of overseas operation and maintenance experience and currently serves Chinese brands overseas. There are hundreds of companies. It is the first senior consulting partner of Amazon Cloud Technology in China to obtain generative AI capability certification. At the same time, it passed the Amazon Cloud Technology MSP certification Renewal with full marks this year.

In addition, in terms of large model data service capabilities, Orion Star, owned by Cheetah Mobile, has sufficient practical experience in improving large model effects by improving data quality.

These are all due to Cheetah Mobile’s business genes and integration advantages. In recent years, Cheetah Mobile’s strategic transformation has shifted from traditional ToC business to ToB business with AI and large models as the core, and through controlling Orion Xingkong further strengthens its layout in the fields of AI service robots and AI large models. The AI ​​data treasure jointly created by Orion Star and Juyun Technology is a concentrated expression of this layout.

In this context, models, businesses, and data are truly connected to the model and the enterprise through AI data treasure, realizing a closed-loop commercialization of large models and accelerating the application of large models.

In addition, Orion Star also announced that it has signed a cooperation agreement with the School of Computing and Data Science of the University of Hong Kong. The two parties will jointly develop AI application educational tools for course teaching scenarios, and carry out "focused on embodied intelligence-related Course Project" to jointly promote the popularization and application of AI technology in application fields.

In summary, it can be seen that AI application innovation and explorationEntering a critical period, infrastructure such as data plays an increasingly important role, and it is even more critical for enterprises to make good use of data. This is exactly what Orion Star is doing now.

04. Conclusion: 8 years of accumulation build a bridge between large model development and enterprise needs

Data is becoming more and more important in the development of large models. Rich data resources allow models to learn and adapt to new changes in a timely manner to meet user needs in different scenarios. AI data services have become an important bridge between models and upper-layer applications.

Better utilization of data is an important step for large models to achieve business closed loop. Orion Star is relying on its eight years of exploration in the field of AI, linking Juyun Technology's insights into the core needs of overseas enterprises, and transforming it into a bridge between enterprise needs and the development of large models.