Slashdot Mirror
Garry Kasparov: The World Should Embrace Artificial Intelligence (bbc.com)
"Chess champion Garry Kasparov was beaten at his game by a chess-playing AI," writes dryriver. "But he does not think that AI is a bad thing." From Kasparov's interview with the BBC:
"We have to start recognizing the inevitability of machines taking over more and more tasks that we used to do in the past. It's called progress. Machines replaced farm animals and all forms of manual labor, and now machines are about to take over more menial parts of cognition. Big deal. It's happening. And we should not be alarmed about it. We should just take it as a fact and look into the future, trying to understand how can we adjust."
Kasparov has given the issue a lot of thought -- last month he released a new book called Deep Thinking: Where Machine Intelligence Ends and Human Creativity Begins. But he also says that the IBM machine that beat him "was anything but intelligent. It was as intelligent as your alarm clock. A very expensive one, a $10 million alarm clock, but still an alarm clock. Very poweful -- brute force, with little chess knowledge. But chess proved to be vulnerable to the brute force. it could be crunched once hardware got fast enough and databases got big enough and algorithms got smart enough."
Kasparov has given the issue a lot of thought -- last month he released a new book called Deep Thinking: Where Machine Intelligence Ends and Human Creativity Begins. But he also says that the IBM machine that beat him "was anything but intelligent. It was as intelligent as your alarm clock. A very expensive one, a $10 million alarm clock, but still an alarm clock. Very poweful -- brute force, with little chess knowledge. But chess proved to be vulnerable to the brute force. it could be crunched once hardware got fast enough and databases got big enough and algorithms got smart enough."
The world should embrace Garry Kasparov. I like a man who gets beaten by an AI, but then embraces AI. =)
Why did the chicken cross the road? Because Elon Musk put an AI chip in its head.
Almost everyone likes the idea of machines taking over grunt work like laundry and driving, but our society is NOT designed to distribute the benefits of AI evenly enough: many will get screwed, career-wise.
It's not so much about AI versus jobs, but how society adjusts (or doesn't). Change can be painful, especially if done wrong.
If the current trend continues, the owners of the technology will get really rich, and the rest will struggle or fail, fighting bitterly over the remaining scraps in ever uglier "culture wars", in part fueled by the rich who want to stay rich by demonizing those who complain.
Table-ized A.I.
A combination of both. The better the algorithm, the less brute force it needs.
Think of it as a lever.
Go was supposed to be a much tougher challenge, not expected to be dominated by machines for decades
I don't think many people keeping up with advances in machine learning were surprised. There were several teams working on Go, and they were making rapid progress. The hardware was also improving rapidly, and much more historical game data was available.
the pool of humans who are even capable of holding their own against AlphaGo has likely dropped to below 1000, out of 7 billion
No, the number is zero. No human will ever again beat the best Go program.
There will still be human Go tournaments, just like forklifts haven't done away with human weightlifting contests.
I don't think many people keeping up with advances in machine learning were surprised.
Most people even involved with Alpha Go were surprised at how quickly they were able to dominate human Go champions. From what I have read only Hassabis was confident they could do it in a few years. In most cases even AI researchers are often wrong about how quickly AI is getting better.
Humans are not very good at comprehending exponential increases in capability, even in their chosen fields. People have been spending too much time worrying about the end of Moore's law, and ignoring that exponential increase in algorithm performance has been much faster than even Moore's law.
There will probably be some things we assume are easy which will still elude us in 50 years (like flying cars). But most things we think will take 100 years will probably take less than 20.
-- All that is necessary for the triumph of evil is that good men do nothing. -- Edmund Burke
Here's the thing about 'brute force' in computing. Computers can go through millions of computations and thousands of strategy scenarios in a second. As we are seeing today, a computer can simply brute force its way through encryption, simply by trying *everything* until you get the desired result, simply because the machines are so damn fast.
Brute Force can be an exceptionally powerful way of doing something, if it is tweaked to and pointed at a particular problem, in Kasperov's case, it was Chess.
Yes, the computer wasn't intelligent, but then again, neither are half the people I meet. Those people are simply brute forcing their way through life, without a single thought in their heads.....
If telephones are outlawed, then only outlaws will have telephones.
will it be able to edit?
Humans playing chess is like a dog riding a bicycle: it can be done, but it's not what the organism was designed for.
The organism was not designed, it evolved.
And the only thing it evolved for is to survive long enough to replicate under a narrow (on a cosmic scale) set of conditions.
These issues are very deep and potentiall deceptive. Even the cleverest of people can get hopelessly misled.
In Genna Sosonko's excellent book "Russian Silhouettes", a series of in-depth sketches of great chess players whom Sosonko knew personally, there is a very instructive anecdote about Mikhail Moiseyevich Botvinnik, multiple world champion and considered the "father" of the mighty Soviet School of Chess.
As well as being a superb chess player - although an amateur by modern standards, as he strictly limited the time he devoted to the game - Botvinnik's "day job" was electrical engineering. He launched projects to study the potential of computers for a wide range of important types of work. Sosonko tells the following instructive story.
[Botvvinik declared that] "... to write a program for managing the economy is easier than for chess, because chess is a two-sided game, antagonistic. The players hinder each other, and the devil knows what that means, whereas in economics that is not the case, and everything is simpler".
It's not so often that one catches a world-class expert in such an utterly mistaken declaration. Today in 2017 computers play chess better than any human, but the problem of managing the economy is still not understood at all. And until it is understood, it cannot be programmed.
I am sure that there are many other solipsists out there.
They were. Go is quite resistant to the brute force and play dictionary techniques used in the past on checkers and chess, which is why people wax poetic about the complexity of Go.
AlphaGo is trained using reinforcement learning, which, frankly, is such a twitchy thing that it's still surprising how well it can work.
Kasparov was beaten by a big computer programmed to play chess. AlphaGo is a very different thing.
What chess playing programs do can pretty much be described as brute force, not to an end of game solution, but choosing the best move based on examining all plausible lines, and using an evaluation function to determine how good each line is.
What is exciting about the (still narrow) AIs developed recently, based primarily on multi level neural networks, is that they can work in situations where no one knows how to create a hand crafted evaluation system. Basically, the system works out for itself what are promising actions, based on its learned knowledge of the probability that each action will lead to desired outcomes. The huge difference is that such an AI can be built without the creators knowing much about the problem domain, or being able to understand the basis upon which the AI is coming up with its solutions. For the most part, the AI is just presented with objectives and large amounts of historical data relevant to the target domain, and works out how to make good decisions in an unsupervised fashion. In some cases, the AI can then refine itself via reinforcement learning where it generates its own data and determines the best solutions. This is still narrow AI, but looks to the observer much more like an independent thinking entity.
There will probably be some things we assume are easy which will still elude us in 50 years (like flying cars).
Flying cars have not eluded us, we have chosen not to make them.
It is not a question of how hard the problem is, it is a question of how valuable the end result is (what is the user experience?). The designs end up being too much of a compromise or too expensive or just too heavily regulated compared to having both a car (or cars) and a plane (or planes).
There will probably be some things we assume are easy which will still elude us in 50 years (like flying cars). But most things we think will take 100 years will probably take less than 20.
I get that you said "most". One exception sadly seems to be space exploration. I'm pretty sure if we could go back in to, let's say, 1965 and get President Johnson and the very top NASA and private industry space experts in a room and told them the following:
"I've got good news and bad. The good news is that we're going to get men on the moon in 1969 and bring them safely back multiple times. (Sounds of cheers from the room)
The bad news is that the last time we'll go will be 1972 and we won't try again and the next 45 years after that will mostly be spent dealing with one space station (much less impressive in reality then you're likely to expect it to be) and we will be decades, at absolute best, from ever going back to the moon, let along putting someone on Mars."
It's hard for me to imagine any one of those people would believe the bad news part of that, yet here we are.