Artificial Intelligence, Today and in 2035, Still Not Human-Level Smart
Computers are a long way from possessing human-level smarts, but they are doing increasingly useful things with technologies that form the building blocks of artificial intelligence, and by 2035, they’ll be driving our cars, at least sometimes.
So says Oren Etzioni, CEO of the Allen Institute for Artificial Intelligence (AI2).”My contention is that human-level intelligence is not around the corner,” said Etzioni. “It’s not going to be here even in 20 years because our rate of progress is slow. … That said, there are amazing things around the corner.”
One example is Semantic Scholar, a new search service targeting the rapidly expanding body of scientific research, from AI2. The service employs technologies that are components of AI—data mining, natural-language processing, computer vision—to improve search within a specific domain, beginning with some three million computer science research papers.
Semantic Scholar “is a radically new search engine that uses artificial intelligence techniques to make the search more efficient,” said Etzioni, previewing the technology at Xconomy’s Seattle 2035 conference last week.
“A huge focus for us has been to build AI systems that try to solve this problem of information overload, focusing on scientists,” he said. “If our scientists can be more efficient, if they can discover things that they would otherwise miss, if they can make connections that are impossible for them to make right now because of the millions of papers, they’re going to be much better in helping to solve society’s and humanity’s thorniest problems.”
Services like Semantic Scholar and achievements such as matching the average high school junior on the geometry portion of the SAT are what AI can do today. But where might the technology be in 20 years?
Etzioni gamely tackled the question by first assessing the rate of progress in AI, in the context of Moore’s Law, which predicted in 1965 the last 50 years of exponential improvement in computing hardware capabilities.
“Exponentials—the exponential curve of Moore’s Law—are just incredible things,” Etzioni said. “They can radically change what you see.”
Computer proficiency at chess has proceeded exponentially, Etzioni said, citing research by Murray Campbell, a scientist at IBM Research and one of the creators of the Deep Blue computer chess program that defeated reigning world champ Garry Kasparov in 1996 and 1997.
“Laptops these days can play grand-master level chess,” Etzioni said. But, he added, “artificial intelligence is a lot more complex than chess.”
Things that are easy for humans are difficult for machines. Take, for example, viewing and interpreting a picture. Humans immediately recognize objects, even if they’re partially hidden; facial expressions; context; and lighting. “We can connect it with reasoning, with understanding, with risks, with goals, et cetera,” Etzioni said.
A computer perceives only a set of pixel intensities—essentially a bunch of numbers. Going from this abstraction to something meaningful that can be analyzed and interpreted ” is much, much harder than chess, and we’re only a very small fraction of the way towards human-level visual understanding,” Etzioni said.
While rapidly improving hardware capabilities have helped computers master chess—essentially a search problem over a set of positions on the chess board—better hardware has had less impact on more nuanced challenges.
“AI is not a hardware problem,” he said. “It’s largely a software problem, and it’s a software problem that’s ill-defined.”
Questions like, “Should I invest in this company or not? How do I write a PhD thesis? Even how do I write an article in a newspaper? Those things aren’t simple optimizations,” Etzioni said. “They’re not simple to even formulate, let alone solve. Problems that we don’t even know how to formulate, the computer doesn’t even have a shot at, at this point.”
So what do computers have a shot at in the next two decades?
“At least inside the city of Seattle—write this down so you can make fun of me if I’m wrong—driving is going to be a hobby in 2035, it’s not going to be a mode of commuting,” Etzioni said. “The same way that hunting is a hobby for some people, but it’s not how most of us get our food.”
Etzioni showed a visualization by University of Texas computer science professor Peter Stone of an intersection of two busy roads, each with six lanes coming and going, with no stop signs or traffic lights. Cars streamed through without stopping, accelerating or slowing to avoid collisions. “They all have AI software that allows them to instantaneously and correctly communicate with each other and make reservations saying, ‘I’m passing through,'” Etzioni explained. “And the algorithm makes sure that everybody passes through as efficiently as possible without any danger.”
Even this would be accomplished with clever algorithms set to work on a well-defined optimization problem. “That is not human-level intelligence,” Etzioni said.
But it is a welcome vision in a city dealing with worsening gridlock year by year.