the human now adds absolutely nothing to man-machine chess-playing teams.
I am pretty sure I predicted this. I certainly would have if anyone asked. Whenever you get to the point where a computer is close to human capability at something, you should bet on the computer becoming much better at it within a few years. Humans only get better slowly, and computers get better rapidly.
Imagine that you were running at pretty close to top speed, and there is some other creature that is currently chasing you. If that creature is gaining on you rapidly, then you aren’t going to stay ahead of that creature, are you? And after the creature catches up with you, if it keeps going you are not going to be able to keep up, are you?
By the way, that is why I take a bullish view of self-driving cars. Maybe we will make the physical and regulatory environment for self-driving cars as unfriendly as possible. Otherwise, I think they will take over. People will come to see driving as a waste of time. They will come to see having a car that is idle most of the day as a wasteful expense.
Truly ancient joke: two friends out hiking, they happen upon a bear which takes out after them. As they flee, one says, ‘Bob, we can’t outrun a BEAR’ and the other says, ‘Jim, I don’t have to run faster than the bear, I just have to run faster than you’.
http://www.sciencealert.com/google-s-ai-built-it-s-own-ai-that-outperforms-any-made-by-humans
Tyler is right that Google will become a hedge fund, but before that they will likely become the premier builder of AIs. Cost efficient in a particular environment.
I expect to see more AI “idiot savants” who are genius in some realm, but only used there. Like the Watson cancer doctor.
Regarding self-driving cars, the question isn’t whether or not they will become better drivers than humans. They will. The question is how many deaths, injuries, and damage will self driving cars cause on the way to that point. Gives new meaning to “Move fast and break stuff”, doesn’t it?
what do you mean “people will come to see”?
right now people HATE being stuck in traffic or having to pay a small fortune so they can be stuck in traffic
thats the rational behind developing autonomous cars
Yes, but a different situation could be when you are standing on top of a hill, and the creature starts at the bottom, quickly races up half the hill in a minute, but then takes another minute to move up half of the remaining journey, and so forth for each remaining step. So it only takes four minutes to get to the 94% mark, and only seven to pass the 99% solution, but it never quite makes it all the way.
So the question is, which of these scenarios is a better description of current efforts to automate certain cognitively complex human activities?
If the bottleneck is simple computational horsepower then yes, it’s more like your metaphor, and it’s probably only a matter of time so long as the cost per computational operation keeps decreasing.
If the bottleneck is a fundamental lack of understanding on how it is that the brains of human beings (and other animals) deal – often instinctively, immediately, and effortlessly – with messy environments full of fuzzy information and unexpected situations full of uncertainty, and how one might be able to replicate that success with computers, then it might be closer to the second metaphorical scenario.
I think a lot of people are missing the real astonishing thin about Alpha-Zero, which is that unlike all the other top programs, it didn’t benefit at all from ‘studying’ the comprehensive database of the games of the best human players, and has to use its particular approach to machine learning combined with extremely impressive computational abilities to ‘discover’ all the best moves from complete scratch by learning from ‘playing itself’.
And by doing do, it is now claimed to arguably be the best chess playing engine in the world by far, able to beat out all the other chess programs and chess-human combinations, and indeed perhaps even benefiting from being free of the ‘contamination’ of all that human fallibility.
We shall have to wait and see whether these extremely powerful new techniques have wider applicability into noisy, real-world situations that currently require some application of human judgment and common sense.
Samuelson’s introductory textbook used to have a graph of US and USSR per capita GNP, with the USSR gaining rapidly on the US, and possibly passing the US as early as 1985 (p. 831 in the 1970 8th edition). As we know, that did not happen. The same turned out to be true for Japan. And people debate what will happen with China.
I don’t know if I should be impressed by AlphaZero or not. Chess-playing is not one our natural endowments. It’s like flying, something we’re really, really bad at. It ought not be a surprise that a special purpose machine is far better at it. Nevertheless, I’m impressed.
Facial recognition is one of our endowments. Now AI can do it better and faster than any human.
Recognizing voices and transforming audio into words, is one of our natural endowments, and now computers are doing that really well, though making just a few mistakes when meaning has to be understood. That’s one of those 90% solution issues I was talking about.
Translation is certainly an art with various degrees of skill, since it’s an inherently challenging task to convey implied meaning in sequences of words with similar length, “Une traduction, c’est comme une femme: si elle est fidele, c’est qu’elle n’est pas belle ; et si elle est belle, elle n’est pas fidele.”
But for bilingual people, at a basic level, it remains “one of our natural endowments,” and google in particular has made stunningly rapid and unexpected progress in this regard in just the last several years.
It’s going to come to be the case that when humans are doing a job, it’s not that a machine can’t do it yet, but only because those humans come cheaper.
“They will come to see having a car that is idle most of the day as a wasteful expense. ”
I suggest not. My house is full of things that are idle most of the day. The house itself is empty a large part of every day. These things add up to way more than the cost of a car.
There may well be cities where people use self-driving UBER services instead of mass transit. But for something like 98% of the surface of the US, people will own their self-driving cars for the same reasons they have cars now.
And remember, *somebody* has to own the car – so one way or another you will be paying a capital expense for a given level of service.
Right. There’s a key difference for markets in which the typical timing of use is key.
There’s a big difference between a piece of capital that is used almost every day, but is idle for much of that day, and a piece of capital that is frequently idle for many days at a time, and could conceivably be rented out.
It’s easy to rent one’s house out for at least one night on AirBnB (or whatever), and for terms with typical hotel hours. It’s hard to rent it out for just the hours while one is away at work. Society often doesn’t run like a Navy ship where we hot-swap bunks and desks to maximize efficient, 24/7 utilization, even though it could in theory. There are big potential risks and costs associated with sharing things with strangers, and it’s often not worth it.
Even a certain portion of mass transit vehicles likes buses and subway trains and even taxis are idle in the hours of lowest utilization, to include 24/7 systems like NYC. During morning rush hours, the trains heading back to the exurb ends of the lines from the center are mostly empty (which is kind of like ‘idle’), and vice versa for evenings. Surge power plants are used every day during peak load (or when the renewables aren’t being renewed), but are idle otherwise.
So, the thing with cars is that that commuting is the major daily driving experience for most people, and most people commute in a very coordinated fashion, specifically because their work involves interacting with each other and they all have to converge and congregate.
I’m guessing in that even ideal circumstances of self-driving cars and proactive internet logistics, most vehicles will sit in the same parking spots all day.
Also, while this could change over time, I don’t think current suburban setups are well suited for a lot of just-in-time utilization of vehicles. A lot of big box stores would have to become a lot more like airports with lots of dropping off and picking up, probably segregated at “arrivals” and “departures”.
By the way, that is why I take a bullish view of self-driving cars.
I suspect one advantage of humans is we have great sense of knowing what to ignore than a computer. (Just think how many things you ignore driving on five lane freeway in city.) Drivers ignore a lot in what is in line of sight or around the car.
I still say they are 10 – 15 years from taking over the world. While I agree with your reasoning that driving is a pain, and our nation is full of cars not driving 90% (95%?) of the time. The sunk cost on cars is huge on the family budget. However all the programming is focusing on optimal conditions and that last 10% is going of driving hazards is going to be tough. So in five years the systems will be sold like a drone as almost a play thing but it will still need a ‘driver’ behind the wheel. Then in 10 -15 years it will be a lot more operational.
I still say the number of truck driving and delivery is not going to fall that much for awhile. I can’t imagine companies want to have trailers with the value of $10M on the road with some kind security. So they will truck security.
Remember how we were always just 5 years away from having quantum computing become a reality? We will remain just 5 years away from self-driving cars functional for at least another 15-20 years.
Steve
It’s not just about the best way to commute or locomote. At least for another generation or three there will exist all those car and classic car aficionados who just love cars, performance, speed and the aesthetics of racing and automobiles. These are the crowds at NASCAR and car shows and the mechanically inclined who like tinkering and futzing around with engines and motors. It’s a cultural thing I think beltway types might underestimate.
I’m reluctant to compare them to physical book lovers, but that might indeed be apt.
Where is the AI who can talk & listen like “HAL 9000”.
Language speech to text in real time, “with understanding”, is like the peak climbing which slows dramatically. Dragon speech was 80% accurate 10 years ago. Still nothing, including not yet Google nor IBM Watson, which is 99% accurate speech to text. And progress is slow.
Funny strange that facial recognition was so much faster for computers — we still don’t know what AI will be good at and what it won’t be.
8 billion individuals, so each face is one of them — this is an easier problem than “what did he say?”
“I am pretty sure I predicted this. I certainly would have if anyone asked.”
Oh, I pedicted that as soon as I read Cowen’s post on that type of chess game a few years back. And you’re right, there are hardly more obvious computer predictions.