Google’s AI technology company Deepmind is now I have encountered some problems. Since 2017, Deepmind has continued to lose money, raising questions about the future of AI.
Although Deepmind created AlphaGo to defeat World Go Champion, OpenAI Overcoming top human players in DOTA 2, but these AIs can only be challenged by a single task, and it is impossible for AlphaGo to play “DOTA 2”.
But AI must learn to handle multitasking when it comes to people’s real life. This is also one of the main challenges of the next phase of AI.
So, Facebook and MIT have collaborated on a project to develop an AI system that can handle multitasking in Minecraft games.
My World is currently the best-selling video game, The world has sold 176 million copies, and 90 million people participate online every day. This AI will exist as a player’s assistant in My World, interact with players in the game, and perform a series of tasks as required.
Why can “My World” train “All-rounder AI”?
First of all, this 3D sandbox video game, which is similar to the stack, is inherently “infinitely diverse”. The hidden creativity in the game is far from comparable to other games. The researchers said:
The creativity and number of items created by these players is shocking, including landmarks, sculptures, temples, roller coasters and even the entire urban landscape, where they explore, interact, trade and fight in these 3D physical worlds every day. .
At the same time, the rules of the game are simple and predictable.
So for primary AI training, this is the ideal environment for developing a versatile AI. In this process of interacting with the player, the AI will evolve “grow” and not break new skills. The researchers said:
We don’t need it to have superhuman performance in a single difficult task, but first have the ability to accomplish simple but multiple tasks assigned by humans.
However, the biggest challenge of this process for AI is to understand human dialogue.
The player enters the virtual world in the game, makes everyday tools, collects natural resources, and builds a blocky structure beyond imagination. The communication during this period is very similar to reality. When the player asks for the AI assistant, it all requires the AI to parse the language.
Even a simple request requires AI to have enough information. The MIT team said:
For example, “Build a tower with 15 blocks and put a big smile on the top.” The AI assistant needs to know what the tower is, 15 what is it, how to build the tower, how to measure the height, and understand the player at the same time. The contextual meaning, knowing how to build a “smiley face” and what it means to put “smiley face” on top of the tower.
But AI will also learn the most to handle multitasking during this period.
As early as 2015, Facebook is in MessenThe ger app has launched a AI Assistant named M, but according to early testers, it’s 70% of mobile phones need to rely on people’s help, plus a single use scenario, which also limits its “growth” in the user’s request. This time, in cooperation with MIT, it also brings a new one to Facebook training AI. Opportunity.
At present, the newly developed AI Assistant is continuing to receive training in “My World”.
In the progress document, we can already see that it responds naturally to the player’s query and is instructing Under the “circle” required by the player, the player can also give AI additional comments at any time.
