Chapter 1

Getting Started with Microsoft Copilot

IN THIS CHAPTER

Seeing how Copilot works

Learning about Copilot's capabilities

Accessing Copilot

Experimenting with basic commands

Using Copilot responsibly

Microsoft Copilot is an umbrella brand name for all of Microsoft's AI-powered chatbots. Chatbots such as Copilot and similar products from OpenAI, Google, Apple, and many others have the potential to change the way people get work done. At their best, AI chatbots can enhance productivity, learning, and creativity. At their worst, they can produce low-quality text and images, confidently answer questions with fabricated data, and displace human jobs.

In this chapter, you learn some of the ways that you can access Microsoft Copilot, you get an overview of its capabilities and limitations, and you learn about using AI responsibly and ethically.

Defining Copilot

In 2019, Microsoft invested in the then-tiny AI startup called OpenAI. Microsoft provided billions of dollars, and OpenAI ran its systems on Microsoft's computers. In 2021, Microsoft exclusively licensed OpenAI's GPT-3 model, which was used to create OpenAI Codex. OpenAI Codex was subsequently used by GitHub - a subsidiary of Microsoft that provides tools and hosting for computer programmers - to create a computer programming assistant called GitHub Copilot, shown in Figure 1-1.

FIGURE 1-1: GitHub Copilot.

Although GitHub Copilot and Microsoft Copilot are similar, in that they both use OpenAI's technology for understanding and generating language, they're two different products. GitHub Copilot is optimized for helping with the writing of programming code and Microsoft Copilot is optimized for chatting with people and generating written words in human languages.

Several months after GitHub Copilot was rolled out as a plugin for Microsoft's Visual Studio Code Editor, OpenAI released the first version of ChatGPT for use by the public. ChatGPT became the fastest-growing consumer Internet app of all time - gaining 100 million monthly users in just two months.

With its unprecedented ability to respond to user queries with human-like text, ChatGPT became a cultural sensation and possibly even a threat to the traditional search engines created by Google and Microsoft.

Microsoft responded to ChatGPT by redesigning its Bing search engine. Bing Chat, as it was called, was rolled out starting in February 2023 and gained its first 100 million active users within months. The early version of Bing Chat had a tendency to produce false data (also known as hallucinations) and troubling responses during chats, including, as reported by Kevin Roose in a The New York Times article, acting like a "moody, manic-depressive teenager who has been trapped, against its will, inside a second-rate search engine."

Hallucinations, in AI lingo, are defined as incorrect or misleading information generated by AI. They're caused by a variety of factors, including insufficient training, incorrect assumptions, and biases in the data used to train the AI model.

Microsoft clamped down on much of Bing Chat's tendency to go off the rails and rebranded it as Microsoft Copilot. The current homepage for Microsoft Copilot (https://copilot.microsoft.com) is shown in Figure 1-2.

FIGURE 1-2: Microsoft Copilot on the web.

Overview of Microsoft Copilot

Copilot has been integrated into many of Microsoft's products and can be helpful with a wide variety of tasks. In fact, there are so many possible ways to use Copilot that the possibilities can sometimes seem overwhelming. AI chatbots are a fundamentally different way of interacting with computers than most people are used to, so it can be helpful to look at them as if they were traditional computer software and start by talking about the features and what makes Microsoft Copilot different from its competition.

Core functionalities and benefits

The most basic function of any chatbot, whether it's powered by artificial intelligence or not, is to respond to human speech or writing (which is also known as "natural language") with easy-to-understand text or speech. The quality of a chatbot can be measured by how human-like its responses are.

The current crop of AI chatbots can all generate highly convincing natural language responses to people's questions and requests.

The voice or text input a user of a chatbot gives to the chatbot, whether it's a question ("How tall is Mount Everest?") or an instruction ("Summarize this email.") is called a prompt. The primary way for people to interact with chatbots is through prompting.

Beyond its core ability to respond to prompts in natural language, Microsoft Copilot has exciting additional capabilities that make it stand out in usefulness, especially when it's integrated into other Microsoft products.

Some of the features of Copilot include:

• Performs web searches using Bing.
• Integrates with other AI tools to create original images and music.
• Writes original text or rewrites existing text.
• Cites the sources of the text it generates.
• Personalizes its interactions with you based on previous interactions and documents you work on.
• Translates text between different languages.
• Supports plugins that expand Copilot's capabilities.
• Supports user-created chatbots.

Key differentiators from other AI assistants

The main thing that makes Microsoft Copilot more useful than other AI assistants is that it is integrated into Microsoft Windows and other Microsoft programs. This integration gives Copilot the ability to not only generate text and images, but also to control certain aspects of the software it's integrated into. For example, using a non-integrated chatbot, such as ChatGPT, you can ask for text for a PowerPoint slide that you then need to copy into PowerPoint and format manually. With Microsoft 365 Copilot, you can ask for PowerPoint slides or an entire presentation, and the Copilot assistant will create the new slides, format them for you, and insert them directly into your presentation. Even better, Copilot can access and use other documents you've created while creating the new slides.

Another key factor that distinguishes Copilot from many other chatbots is that Copilot has access to the data in Microsoft Bing. By augmenting the data it was originally trained on with search results from Bing, Copilot can answer questions about the latest news and other developments, whereas other models have a "cutoff date" beyond which they can only speculate (or hallucinate).

Understanding how Copilot works

Chatbots like Microsoft Copilot and ChatGPT are far superior to their predecessors, such as Office Assistant, also known as "Clippy" (shown in Figure 1-3). Microsoft integrated Clippy into Microsoft Office applications from version 97 to 2003 and it proved to be more annoying than helpful in most cases.

FIGURE 1-3: Original Office Assistant, also known as "Clippy."

The reasons for Clippy's failure have been studied exhaustively, but the crux of it is that Clippy was intrusive and would appear whenever it detected that you were doing something (such as writing a letter) that it was supposed to be able to help with. But then, when you agreed to let Clippy help you, all it could do was reference official Microsoft Office documentation, which wasn't helpful for much of anything.

The two most important factors that contributed to making the latest generation of AI so much better than Clippy (and all subsequent AI assistants) are:

• Vastly more data (and computing power) was used to train them.
• They take advantage of new AI techniques that allow them to consider context when generating responses.

Learning from all the data

The AI model behind Microsoft Copilot is named Prometheus. Prometheus is OpenAI's technology combined with Bing's search index. The result is that Copilot has learned from and has access to a tremendous amount of data.

Although the relationship between training data size and a model's performance isn't simple, in general, larger models are able to gain a better picture of whatever they're designed to simulate (such as communicating using natural language, in the case of a chatbot).

Context is key

Even more important than simply throwing more data at an AI system is a technique known as attention that was invented by Google in 2017. In short, what attention techniques do is allow AI models to look at different parts of your input and their own output while figuring out what to say. For example, consider the following sentence:

"The bank can guarantee deposits will be safe because it has invested in secure vaults."

An AI model that uses attention mechanisms will know that "bank" refers to a financial institution rather than the bank of a river because of the other words in the sentence, such as "guarantee," "deposits," "safe," and "vaults." A model with...