Early Indications May 2015: Centaurs

“We expect a human-robot symbiosis in which it will be natural to see cooperation between robots and humans on both simple and complex tasks.”

-George Bekey, University of Southern California, 2005

“The second machine age will be characterized by countless instances of machine intelligence and billions of interconnected brains working together to better understand and improve our world.”

-Erik Brynjolfsson and Andrew McAfee, MIT, 2014

Who performs better: a computer or a human?

The short answer is obvious: it depends on the task. Computers are now unquestionably better at chess than even a grandmaster-grade human player, and the highly visible triumph of IBM’s Watson over the best Jeopardy players shows how artificial intelligence can be applied to a linguistically rich trivia contest.

What might come next? Just this month, four of the top 10 poker players in the world played a marathon against a Carnegie Mellon computer. Given the complexities of no-limit Texas Hold’em, the result was not a Jeopardy-like rout, but the statistical tie elated the researchers. Each player played 20,000 hands; a cumulative $170 million in chips was bet over the two-week competition. In the end, the humans came out less than $1 million ahead — even though the computer did things like betting $19,000 to win a $700 pot.

Physical-world tasks are farther behind pure thought. In soccer competitions, robots are still decades away from impersonating, much less beating, human championship teams. Physical machine movement has yet to follow anything like Moore’s law, and team play is harder to model than individual contests. As these examples show, the race between people and computers/robots plays out differently, depending on the tasks being contested.

The long answer to the "who's better" question is emerging: a team of both. We will see the origin of the term “centaurs” presently, but I think this is going to be the most amazing domain, one in which each party does what it does best. We are seeing that teams of humans AND robots outperform either humans OR robots. Here are four domains in which progress is being made more rapidly than might be widely understood.

1) Audi has teamed with Stanford’s self-driving car lab to develop a TT that can beat a club-level human driver. There have yet to be head-to-head races, apparently, so human adrenaline hasn't played a role, nor have racing tactics come into play. The car simply follows a pre-programmed line and parameters around the course: it hasn't raced anyone and won yet. The centaurs are well developed here: stability control, anti-lock brakes, and sophisticated all-wheel-drive control systems all digitally amplify the skill of a human driver, so finding a "purely" human-driven car is less than straightforward.

Earlier this month Mercedes showed a driverless truck that can operate on public roads but still needs human drivers for navigating the start and finish of a trip as well as any diversions from clear, open highway such as snow covering lane lines, police officers directing traffic, or construction areas. It’s early, but eventually might the analogue of a “harbor pilot” carry over from sea to land?

2) The Internet is awash in images, some of them incredibly beautiful. Researchers at Yahoo Labs and the University of Barcelona have taught an algorithm to trawl through image databases and find beautiful but unpopular (under appreciated) images using the results of training sessions with human “votes.”

As The Economist recently noted, the process of machine learning is itself undergoing rapid improvement, in part through the process of “deep learning” as developed by the giant web businesses with both massive data and effectively unlimited computing resources. Google and Facebook are familiar names on their list; Baidu is a newer entrant into the field, having made some high-profile hires.

3) Chess has never been the same since Deep Blue defeated Gerry Kasparov, in part because of a software bug that led the human to infer that the machine was substantially smarter than he rather than allowing for the possibility that it was a dumb move. Since about 2013, teams of average players and good software have been able to defeat both grandmaster humans and computers. This type of match is where the "centaur" terminology first took hold.

4) Exoskeletons are common in Hollywood sci-fi, but robots that encase a human body and amplify its capabilities are coming into use in several scenarios:

-Rehabilitation for stroke patients, amputees, and paralytics, among other populations

-DARPA wants soldiers to be able to march or run longer, with less fatigue

-In military and other similar scenarios, able-bodied humans can be augmented to increase their lifting capacity, for example

-The Da Vinci surgical robot is a specialized exoskeleton of a sort, extending a doctor’s finger manipulations into more precise movements in the surgical field.

One big challenge for all of these efforts is in making the power source light enough to work at human scale. In warehouses, to take a very rough approximation, a forklift truck typically weights 1.6 to 2 times the intended weight to be carried. If a human is intended to carry 200 additional pounds, that puts the exoskeleton in the 400-lb range, unloaded, so the whole package would be about 750 lb. Lowering the battery weight is the quickest way to shrink the total assembly, but physics is tough to cheat: a lot of battery power would be expended in carrying the battery, and carrying a frame sufficiently robust to support the battery.


It will bear watching to see how roboticists and computer scientists design the cyber side of the centaur, optimizing around human strengths that might be expressed in unpredictable ways. Similarly, training a human to leave part of the task to a machine, and not to overthink the transaction, might be tricky in certain situations. In others (traction control on the car for example), people are already augmented and don’t even realize it.

At the same time, centaurs will have to deal with the infinite supply of human stupidity: what will self-driving cars do when a drunk driver is headed the wrong way on a divided highway? Wall Street is one big centaur, as the recent charges in the 2009 flash crash reveal: a day trader in England apparently spoofed enough orders — manually rather than algorithmically — that programmatic trading bots reacted in unstable, unpredictable ways. The gambit seems to have worked: the day trader (who lived with his parents) made $40 million over four years. The point here is not what one Navinder Singh did or did not do, or when other actors in the flash crash might be identified, but simply that the interactions between clever (or less than clever) people and computerized entities will be a most complicated territory for the coming decades.

Once you start seeing the world this way, the potential possibilities expand far beyond websites, apps, or algorithms — there’s so much human work that can be done better. Consider travel: I would love to have a computer assistant work with me to book a trip. I have several free weekends, let’s say, and want to know the best, cheapest trip I should take. Right now it’s possible to spend hours looking at maps, air fares, hotel rates, weather predictions, and events calendars. A computer can’t tell me what I want — my preferences are dynamic, conditional, personal, and fickle — but right now the computer can’t really do a great job of letting me discover what I want either.

I have a suspicion some of these limitations are far from being solved. at the same time, whether it’s in the realm of computer-controlled tools (whether scalpels or lathes), transportation (personal drones are worth a book all by themselves), or human augmentation, the various tandems of human and computing capabilities will have far-reaching impact sooner than most anyone expects.