Microsoft releases Copilot agent and also brings native small language model

## News on May 22, 2022, on Tuesday local time, Microsoft released Copilot, a new agent at the Build developer conference. It is like a virtual employee and can automatically Perform tasks. Microsoft believes that this kind of artificial intelligence will eliminate boring work content rather than completely replace employee work. Microsoft's Copilot is developed based on the open source code model GPT-3, which uses a large number of code libraries and algorithms to help developers write code more efficiently. Copilot's ability to analyze context and generate matching code snippets greatly speeds up the speed and quality of development. However, Copilot's functionality is still limited and it cannot completely replace artificial intelligence. It usually

In addition, Microsoft also launched a native small language model Phi-Silica for Copilot+ PC. Microsoft announced that all Copilot+ PCs sold starting in June will have Phi-Silica embedded. This is a model specially created by Microsoft for the neural processing unit (NPU) configured in artificial intelligence personal computers. It is the smallest of all the company's models, with only 3.3 billion parameters.

Microsoft claims that the Copilot agent will not replace humans observing and waiting for commands, but can actively perform tasks such as monitoring email inboxes and automating manual data entry.

This represents a major change in the behavior of what the industry calls AI agents, which enables chatbots to independently and intelligently perform complex tasks. This is a major change in what the industry commonly calls intelligent agents, which allow chatbots to automate a range of complex tasks.

Charles Lamanna, vice president of business applications and platforms at Microsoft, explained in an interview: "We quickly realized that being limited to Copilot for conversations greatly limited its current capabilities. "Instead of making Copilot wait for others to chat with it, let's make it more proactive and allow it to perform tasks automatically in the background."

Microsoft showed this to a very small number of early access testers. The new capabilities are planned for public preview in Copilot Studio later this year. Enterprises will be able to create Copilot agents that can handle tasks ranging from IT support to employee onboarding. "Copilot is evolving from an assistant working alongside you to an assistant working for you," Microsoft said in a blog post.

This Copilot agent will be triggered by specific events and will operate in conjunction with the company's own data. Here’s a fine-tuned onboarding example of how Copilot works: Copilot is an intelligent assistant designed to help new employees adapt and integrate into the enterprise faster. He will provide relevant training resources and guidance based on the new employees' personal information and position requirements to help them understand the company's culture, processes and working methods.

Imagine you are a new employee. A proactive Copilot will greet you, answer your questions using HR data, introduce you to colleagues, provide you with training information, help you fill out forms, and schedule meetings for your first week. This way, HR and other employees can focus on their regular tasks and no longer need to be bothered by administrative work.

This automation could raise questions and fears about job losses and the future direction of AI. Ramana believes that the Copilot agent can remove repetitive and monotonous tasks from work, such as data entry, rather than completely replace the work. This frees up more time and energy for people to do creative work.

"Jobs and positions are made up of many different and diverse tasks. If someone keeps doing the same job over and over again, then it is likely that this job has been automated by existing technology." Lamanna said, "We think that with the use of Copilot and Copilot Studio, some tasks will be fully automated...but the good news is that most of the tasks that will be automated are tasks that no one really wants to do."

About Small Language model Phi-Silica, Microsoft stated that the first word delay of this model is 650 tokens per second and the power consumption is approximately 1.5 watts. This means it doesn't use a lot of system resources, freeing up your PC's CPU and GPU to handle other computing tasks. In addition, Phi-Silica reuses the NPU's KV cache when generating Tokens and runs on the CPU, which can generate approximately 27 Tokens per second.

A Microsoft spokesperson said that Phi-Silica is unique in that it is the first locally deployed language model for Windows. It is optimized for NPU and can bring extremely fast local inference processing to the device. This is a pivotal moment, marking the arrival of advanced AI technology directly on the Windows platform, enabling third-party developers to create superior first- and third-party experiences. These experiences will be rolled out to end users this fall, improving productivity and accessibility within the Windows ecosystem.

Phi-Silica is the fifth product in the Microsoft Phi-3 series of models. The other four products are Phi-3 mini with 3.8 billion parameters, Phi-3 small with 7 billion parameters, and Phi-3 small with 14 billion parameters. Phi-3 medium and Phi-3 vision with 4.2 billion parameters.