Facebook's artificial intelligence research director says today's 'best AI systems are dumb'
So it's easy to imagine AI as super-intelligent machines that are leaps and bounds ahead of human intelligence.
But we just aren't there yet.
Not even close, Yann LeCun, the director for Facebook's artificial intelligence research (FAIR) told Popular Science. Today's artificial intelligence is quite, well, stupid.
"Right now, even the best AI systems are dumb," LeCun told Popular Science's Dave Gershgorn. "They don't have common sense."
Researchers have created artificial narrow intelligence (ANI) - programs that are very good at very specific tasks, like IBM Watson's Jeopardy-playing supercomputer or a world-champion-beating chess program.
But no computer scientist or AI researcher yet knows how to make artificial general intelligence (AGI), programs that are capable of exhibiting human-level intelligence on multiple tasks.
For AGI to exist, computers would need to be imbued with common sense reasoning, the ability to encounter new situations and make inferences about what's happening and what needs to be done.
LeCun used an everyday human behavior to describe how common sense works. When asked to imagine a person leaving a room with a bottle, a human would describe the scene in a pretty specific way: the person would stand up from the table, pick the bottle up, walk to the door, open the door to the room, and then walk out of the room with the bottle.
LeCun said that a human could deduce all those details "because [we] know the constraints of the real world," but an AI without "common sense" knowledge of how the world works would struggle to envision the scene playing out that way.
The phrase "person leaving a room with a bottle" simply doesn't contain not enough information for a machine to come to the same description.
In an email interview, LeCun told Tech Insider that one of the biggest obstacles to endowing machines with this kind of common sense is getting machines to learn in an "unsupervised manner, like babies and animals do."
"Right now, the way we train machines is 'supervised,' a bit like when we show a picture book to a toddler and tell them the name of everything," LeCun said.
This kind of learning can teach the program to make connections, but not the "common sense" of why those connections exist and how they apply to other real world situations. We still haven't figured out the best way to teach AI that kind common sense yet, LeCun told Tech Insider, which means we still don't know how to make our AI any smarter.
And, that, he said, is "why it's very difficult to make a prediction as to when 'human-level AI' will come about."