Siri, Google Home, and Amazon Echo all share a few common traits. For one thing, they are all identified as virtual assistants; application programs that use limited artificial intelligence (AI) to understand and digitally interpret voice commands to complete tasks. However, unlike their human counterparts, AI assistants have historically been capable of executing only one task at a time.
That’s about to change.
Programmers from Facebook Research and researchers at MIT have co-launched a pilot project aimed at understanding how natural language is understood — or not understood — by AI, and how it can learn from dialogue as opposed to traditional algorithms or programming. The end game is a virtual assistant that can execute a wider number of tasks concurrently, and the implications could be mind-blowing.
Why Use Minecraft to Test an AI “Intern”?

According to the online MIT Technology Review, Minecraft represents a perfect platform for AI to learn a wide range of tasks. “Instead of superhuman performance on a single difficult task, we are interested in competency across a large number of simpler tasks, specified… by humans,” says the research team, led by Arthur Szlam.
The world of Minecraft is ideal because it’s a building and crafting game that offers variety, but with a set of predictable and simple rules. This kind of environment is a great breeding ground for machine learning.
“We believe we can make progress towards a useful assistant without having to be able to succeed at every possible request. If true, this could pave the way to further learning during deployment,” says Szlam’s team.
The Struggles of Machine Learning

While the commands in Minecraft might seem simplistic, breaking them down so that a bot can decode and understand them is a much more difficult task. Szlam uses this Minecraft example:
“Build a tower 15 blocks tall and then put a giant smiley on top.”
Researchers argue that in order to succeed in executing this command, the AI assistant must understand nonrepresentational concepts such as “tower,” “15,” and “blocks high.”
“It (also) needs to know what a ‘smiley’ is (and how to build it) and understand the relative position ‘top’,” researchers say. These are all concrete notions to humans, but abstract nonsense to a learning technology.
This is where understanding natural language voice commands would be infinitely useful in multitasking. “An AI assistant … can interact with players, perform tasks on request … also learn from these interactions, and develop new skills,” says Szlam and his team. In short, if AI can learn from its interactions, it can create a whole new range of tasks and perform them.
Siri will do what you ask, so long as the task is simple and uses language already embedded in its repertoire. It has a female voice but lacks the capacity for open dialogue and cannot problem-solve. Traditionally, machine learning was dependent on what creators programmed the machine to learn.
While still in its infancy, the science of machines learning from humans will open the door to limitless opportunities. It may not be long before natural language and self-improvement are not exclusively human characteristics.