AI can beat humans only one game at a time
Despite all of the potential for artificial intelligence to solve our most vexing problems, it's still in a primitive state, according to a new report by Stanford University. But a separate paper, this one by Alphabet's DeepMind, suggests again that it has made some of its best progress in the narrow realm of games.
Why it matters: Those advances are important, but life isn't a game. AI progress outside of these areas has been harder to define and track. "The most important thing for AI is to go from exceptional promise to use in actual everyday life," Martial Hebert, director of the Robotics Institute at Carnegie Mellon University, tells Axios.
Here are key takeaways from the Stanford report, called the AI Index:
- AI investment is through the roof: Between 2000 and 2017, the number of AI startups in the U.S. grew 14X — from 47 to 649. Annual VC investment in AI rose 6X last year, to $3.5 billion.
- AI skills are in demand: The number of U.S. job postings on Monster.com listing AI as a required skill was about 11,100 last year. So far this year, the number is 31,000.
- A race is on between the U.S. and China: Kai-fu Lee, CEO of Sinovation, a Beijing-based VC firm, said in the report that China is making impressive progress in AI, and has far more data—the main building block of robust AI—to work with than anyone else. "In this age of AI, I predict that the United States-China duopoly is not only inevitable. It has already arrived," he said.
- AI is good—at some things: AI is on par or better than humans at detecting objects in an image. (Machines can do this with half the error rate of humans.).
- Yes, but: Algorithms struggle to capture the nuanced meaning in the words we use and how we use them. This is an active area of research. At the Allen Institute for Artificial Intelligence, an algorithm now scores about 42% on science questions an 8th grader might encounter, an improvement of 12% since the beginning of 2016. "We continue to move the score higher and higher," says Institute director Oren Etzioni.
The new advances in games: In a paper released yesterday, DeepMind, a top AI lab, introduced AlphaZero, an algorithm that in just 24 hours learned chess and its Japanese version Shogi, and went on to beat world champions in both.
- AI is beating us at other games, too: It has beat expert humans at Go, Pac-Man and Texas Hold'Em. Work on the latter won the award for best paper at a top AI conference underway this week in Long Beach, CA.
Why it matters: Poker is particularly interesting for AI researchers because information in the game—an opponent's hand, for one—is hidden and deception is a key part of a player's strategy. That makes the game similar to real-world financial markets and political campaigns, where AI could eventually be deployed.
Along the same lines, Facebook and DeepMind are looking to conquer StarCraft, a multiplayer videogame of complexity that dwarfs Go.
- And experts worry about hype: In an expert comment included in the Stanford report, Michael Wooldridge, head of computer science at Oxford University, said exaggeration has created an AI bubble:
"There are plenty of charlatans and snake oil salesmen out there, who are quite happy to sell whatever they happen to be doing as AI, and it is a source of great personal frustration that the press are happy to give airtime to views on AI that I consider to be ill-informed at best, lunatic fringe at worst."
What's needed: Gregory Allen, an adjunct fellow at the Center for a New American Security, said the field should establish solid long-term milestones for future progress. In the space race, the milestones were breaking the sound barrier, reaching space, reaching orbit, then reaching the Moon.
"If we want to think about future implications of AI in society, it would be helpful to have a longer runway," he said.