In the field of artificial intelligence, the existence of general artificial intelligence (AGI) has always been the focus of debate among researchers and major artificial intelligence companies. Several big names in the scientific research field, including Turing Award winners Yoshua Bengio and Yann LeCun, believe that there is no general artificial intelligence.

But some technology companies are willing to put their future on general artificial intelligence. For example, Microsoft and the artificial intelligence non-profit organization OpenAI.

A year ago, Microsoft invested $ 1 billion in OpenAI, hoping that the two parties will cooperate on the Microsoft Azure platform to develop new technologies and expand large-scale artificial intelligence functions to achieve General Artificial Intelligence (AGI). In exchange, OpenAI uniformly granted some intellectual property rights to Microsoft and accepted Microsoft’s commercial sale to customers.


The two sides handed over the first result after cooperation-a new artificial intelligence supercomputer

One year later, at the Microsoft 2020 Build conference, the two parties handed over the first result of the cooperation-a new artificial intelligence supercomputer. Microsoft claims that this computer is the fifth most powerful computer in the world compared to the world’s TOP500 supercomputer.


The computing power ranks in the top five in the world

According to reports, this supercomputer is currently hosted by Microsoft Azure and designed by OpenAI. This supercomputer has more than 285,000 CPUs and 10,000 GPUs. Each GPU server has a network connection capacity of 400Gb / s, and is designed to train a single large-scale distributed AI model.

Microsoft claims that compared to the world ’s TOP500 supercomputer, this self-developedTop 5 by shoulders. This means that it is behind Tianhe 2A of the China National Supercomputer Center, and before Frontera of the Texas Advanced Computer Center, its peak computing power can perform 23.5 to 61.4 trillion floating-point operations per second.


Microsoft self-developed supercomputer

From the performance point of view, benefit from Hosted on Azure, this supercomputer has all the advantages of a modern cloud computing infrastructure, including rapid deployment, sustainable data centers, and access to all Azure services.

The super high performance means that it can be used to train super large scale artificial intelligence models, and can provide super large AI models and the architecture required to train this model. .

Judging from the experimental results, the large models of supercomputers perform well because they can absorb the nuances of language, grammar, knowledge, concepts, and context so that they can Do: Summarize lengthy speeches, adjust content in real-time game chats, parse complex legal documents, and even generate code from searching GitHub.

Microsoft claims that it has used the Turing model in this computer (to be used in open source) to enhance Bing, Office, Dynamics and its Ability to understand the language of other products. In Bing, these models can increase the generation speed of questions and answers by 125%. In the office, the model promoted the development of Word’s intelligent search and key insight tools.

From a technical point of view, large models have the advantage of self-supervision, which means that they can generate labels through the relationship between various parts of the data, which It is essential to realize human intelligence. This is in contrast to supervised learning algorithms, which are trained on manually labeled datasets, so it is difficult to fine-tune industries, companies, or topics of interest.

“This is about being able to accomplish 100 exciting things in natural language processing and computer vision at once. When you start to see the combination of these perception fields, you will have Unimaginable new application. “Said Kevin Scott, Microsoft’s chief technology officer.


Supercomputing is the road to hope for AGI?

Whether there is general artificial intelligence (AGI) in the field of artificial intelligence has been the focus of debate among researchers and major artificial intelligence companies.

OpenAI is a non-profit artificial intelligence research institution. Since its inception, the company has asserted that huge computing power is the only way to AGI, and AGI can learn humans and complete any task. OpenAI believes that powerful computers combined with reinforcement learning and other technologies can achieve advances in artificial intelligence technology.

However, it is unclear whether the new supercomputers highlighted by both sides have enough functions to achieve the functions close to AGI. According to the report of the MIT Technology Review, a team called Foresight within OpenAI conducted an experiment to test how much they can push AI functions by training more and more data and calculation algorithms. . In addition, OpenAI is also developing a system that uses images, text, and other data for training. The system uses a lot of computing resources. OpenAI’s leadership believes that this is the most promising path to AGI.

For Microsoft, whether this supercomputer is the only way to become AGI or a leap forward, at least for now it has brought new market opportunities for Microsoft. Through large-scale distributed AI models, Microsoft has trained models on Azure AI accelerators and networks in an optimized manner to improve the performance and experience of its own products.