Will “interactive” AI for writing Minecraft code be the first step in “computers of the future” | WIRED.jp

How can artificial intelligence enter into a variety of programs? Microsoft recently showed one way to do this. The AI ​​flexibly writes the code according to the situation.

CTO Kevin Scott at the Microsoft Developer Conference, Microsoft Build 2022, on May 25 (US time)provenIt’s a game“Maine Craft”He was an AI assistant to . The non-player characters (NPCs) featured in Minecraft in this demo were using the same machine learning techniques Microsoft tested to automatically generate code.

This result indicates how the development of artificial intelligence in recent years could transform personal computing in the next few years. It has the ability to replace the traditional tap, type and tap interface with a conversational interface.

AI trained in code and natural language

This AI uses the Minecraft API to convert the entered command into code that works “behind the scenes” in Minecraft and returns the appropriate response.

The AI ​​model that controls the robot is trained with a massive amount of code and natural language scripts. Then, the API specification for Minecraft is rendered to the AI, learned along with some usage examples. For example, when a player says “come here,” the AI ​​model reacts in the background and generates the code needed to move the agent toward the player.

In this demo, the robot was able to perform more complex tasks such as taking out items and combining them to create new ones. The model is also trained in natural language as well as code, so you can answer simple questions about how things are made.

It is unclear how reliable this system will work outside the scope of demonstration. However, similar technologies may allow other applications to respond to commands they type or speak to.

AI Coding Tools Development

Microsoft relies on the same technology,「Github co-pilot」We have already built an AI encryption tool called . The tool automatically suggests response code for developers when they start writing or adding comments to parts of their code.

CTO Scott describes GitHub Copilot as the first example of an “AI-first” product to emerge from Microsoft and other companies in the coming years. “The AI ​​that writes the code allows us to think differently about software development. We can express our intentions about what we want to achieve,” Scott says.

Scott does not give a concrete example of this. However, for example, Windows will search for a specific document as requested and send it to a colleague by e-mail, or “Microsoft Excel” from the “AI-equipped version” turns the data set into a graph that will one day appear. .

“We should be able to expect significant productivity gains in cognitive tasks like routines of all kinds that none of us particularly enjoy,” Scott says.

OpenAI model application

Modern artificial intelligence has proven to be good at tasks such as categorizing images, transcribing audio, and translating texts. Under these conditions, the recent development of algorithms, combined with tremendous computer power, has led to the creation of new artificial intelligence programs with higher skills. Creating coherent text, such as program code, is one of them.