Uncover Microsoft’s “Turing plan”: reshape the AI development mode and compete with Google and Amazon

0
42

Tencent technology news on October 13, artificial intelligence (AI) is an important part of Microsoft’s future. The company is competing fiercely with Google to maintain its leading position in the field of enterprise office tools and catch up with Amazon in the field of cloud computing.
This pressure to develop new functions driven by cutting-edge algorithms and the cost of training large algorithms of up to $1 million prompted companies such as Microsoft and Google to start building “basic model”, which is a common term in the research field, meaning a large and powerful AI algorithm that can be copied and fine tuned for different projects. For Microsoft, this represents a major change in the company’s traditional way of developing AI.
Youngji Kim, Microsoft’s project manager, said in his speech on the company’s technology in November 2020: “we are really changing our corporate culture. Not every product team (such as word, outlook, teams and SharePoint) All have their own AI teams and develop independent AI models. Instead, all our teams work together to build this large-scale and centralized AI model, and then carry out large-scale and specialization. ”
Since 2017, Microsoft has been pursuing this goal through the so-called “project Turing”. The task of this team is to build large language models and study how to use them in a large number of products of the company. Turing plan may not be a familiar name outside Microsoft, but its AI can generate text in Microsoft office products and support most of the information provided by Bing search.
If the Turing plan is successful, this strategy may amplify the efficiency of Microsoft’s research funds invested in AI research over the past few decades. It is worth noting that Microsoft not only uses the flagship model Turing NLG of the project internally, but also has begun to sell the technology to some partners, which implies the ambition of the cloud computing giant in the AI market. Take avepoint and volume as examples. Both companies use Turing model technology in their products.
Ali alvi, project manager of Turing project, said: “our job is to promote the forefront of AI innovation as much as possible.”
Turing model has been used to simplify development work
Alvey pointed out that the Turing team was established within the company by Kevin Scott, chief technology officer of Microsoft, to meet the emerging upsurge of in-depth learning. Scott encouraged the team to consider the problem from a higher perspective and cooperate with the azure infrastructure team to expand the model to exponential level. Alvi said that when the CEO Satya NADELLA saw the progress made by the Turing team, he decided to deliver it to the customer.
Kevin Scott, chief technology officer of Microsoft
Avepoint is a partner of Microsoft. It resells and builds applications on Microsoft products. At present, the company has launched two products using Turing model: one is an educational platform for teachers, which can automatically generate test questions using uploaded specific course materials; The second is the enterprise training platform, which can use Turing model to test employees’ internal materials.
Dux Raymond sy, chief brand officer of avepoint, said: “suppose I teach history at any given point in time. I would say, ‘I need a test about World War II, with 10 questions, including multiple choices’, and then it can automatically generate test questions by using Turing model technology.”
Another company using Turing model is volume.ai, which has developed an easy-to-use interface for training a chat robot called Qbox. Customers ask the robot to enter a specific “intention”, or they think people will ask the phrase of the chat robot. Benoit Alvarez, chief technology officer of volume.ai, said that the Turing model was used to generate alternative intentions to help solve similar problems that robot manufacturers may not have specifically thought of including.
The Turing model is slowly expanding
In selling these tools, it highlights how Microsoft sees this technology as a competitive advantage. But the company’s slow roll out of these tools to close partners shows that it is cautious about this new technology.
The philosophy behind the Turing model has been attacked before. Earlier this year, Google CEO Sundar Pichai announced an algorithm called “lamda” to help enhance its productivity tools such as search, voice assistant and gmail.
A few months before its release, Google researcher timnit gebru was dismissed on the grounds that she had co published a paper warning that the large language model (LLM) used in Google’s lamda and Microsoft Turing models might be racist, sexist or insulting. Her research also questions whether the benefits of these models outweigh the environmental and financial costs they bring.
Out of caution, avepoint and volume.ai have always avoided letting the Turing model work without direct human supervision. Avepoint’s test generation tool relies on teachers to approve the problem generated by the algorithm, while volume.ai’s Qbox will only display the suggestions of Turing model to the users of chat robots, not users.
Alvarez, chief technology officer of volume.ai, said that according to his own experience of using Turing model and testing it with the famous gpt-3 text generation algorithm of openai, he would not use Microsoft technology to drive a real-time chat machine that can talk to users. Alvarez said: “we have seen many disastrous stories. I think gpt-3 is like this, and so is Turing.”

Although Alvarez did not mention specific examples, it is impressive that in 2016, the chat robot Tay developed by Microsoft based on another technology began to make racist and genocidal remarks at the prompt of users, which put Microsoft in a dilemma.
Alvi, the project manager of the Turing project team, said that his R & D team works directly with customers to develop how to apply large language models to their organizations and avoid adverse results. This work is also supervised and approved by Microsoft’s larger AI ethics committee called ethics and impact in AI, engineering and Research (aether). Many members of the Turing team are members of the aether Committee and keep them informed of the team’s current products.
“Microsoft pursues the concept of responsible use of AI. We discuss all the limitations of AI and give guidelines on best practices for using the output of these models,” Alvey said
The development of these large models shows no sign of stopping. This week, Microsoft and NVIDIA announced one of the largest known language models, called Megatron Turing natural language generation. The two companies claim that the model is much more accurate than the Turing model before Microsoft in tasks such as language generation.
Alvey wrote: “the quality and results we have obtained today are a big step forward in opening up AI’s full potential in natural language!” (reviewed by Tencent technology / Jinlu)