I would advise against relying on Artificial Intelligence papers published on Science. Computer Scientists look down upon the paywalled/elitist academic publishing model behind Science. As researchers discount the value of papers on the conference/journals where they are published, I would suggest looking to top conferences such as AAAI, IJCAI, CVPR, ICML, NIPS (now NeurIPS), ACL, EMNLP etc.
7
"Will such skills translate to the real world?" Relax. Once they do there won't be enough time left for anyone to worry about it.
7
It is ironic, but the problem with AI is that you cannot predict what it will base a decision on. Thus, a safety feature that is suppose to prevent stalls in jets causes a fatal crash. The algorithms and models are getting better, but it is not clear how you solve this problem with present day methods of developing AI programs. These programs are still stuck simply executing code that is totally dependent on the programmers design.
3
One of the mis-perceptions about artificial intelligence is that the same machine can be in more than one place at a time. According to this article: "In April, a team of five autonomous agents beat a team of five of the world’s best human players."
This mis-perception began no later than with the HAL 9000 in the 1968 film 2001: A Space Odyssey. HAL, it seemed, had developed a normal psychology, just like his human astronaut companions, except that 'he' had sense organs all over the ship. In reality, the psychology of HAL would be something more akin to a human with two or three heads. This might be possible to do in theory, but in practice it would be difficult to proceed through life as a normal person. The two-headed man at a carnival freak show would be closer to the truth for him.
This problem has not gone unnoticed by the psychologists. According to Peterson (06 JUN 2017): "One of the primary reasons that we still don’t really have autonomous robots, though we’re a lot closer to it than we were in the 1960’s, is because it turned out that you actually have to have a body before you can think. And even more importantly, you have to have a body before you can see because the act of seeing is actually the act of mapping the patterns of the world onto the patterns of the body.” (0:07:10)
Cite:
Peterson, Jordan. Biblical Series III: God and the Hierarchy of Authority. Uploaded: 06 JUN 2017. From: https://youtu.be/R_GPAl_q2QQ
1
"Although the result looks like collaboration, the agents achieve it because, individually, they so completely understand what is happening in the game."
Humans collaborate precisely because they have to. We are not omniscient and can't see everything happening at all times so to overcome this "handicap" we hope that through collaboration we can reach the right conclusion. AI as you describe overcomes the problem using means other than collaboration. It uses a type of "brute force" attack.
In my mind the question isn't which is more valid. The question is: which is less dangerous? We collaborate because we know the extent of our limitations, which colors how we accept our conclusions. AI does not. We have to understand that whatever the universe of experience that AI has to draw from in fact has boundaries defined by us (in other words, it's only the illusion of omniscience) and so we have to take any conclusions reached and results achieved with a grain of salt.
What happens if we leave something important out of that universe? Something we didn't know to put in. AI won't be able to tell us the result is flawed. Not so big a deal playing capture the flag. But what else will we use AI for?
2
I should become a camp counselor then, not a warehouse employee. Gotcha. Maybe a warehouse robotics technician. Now there's an idea.
My first thought mirrored Mark Riedl though. If you possess perfect knowledge of the game state, you're essentially cheating. No human player will ever have a perfect awareness of all the different game objects in real time. Humans are not part of the machine. We therefore need to rely on peripherals in one form or another. The humble keyboard and monitor being the most basic input/output.
More importantly though, the robot isn't going to have perfect knowledge of the game state if you put deep state in a warehouse. The robot can't track any object that isn't part of the network, not precisely anyway. That means humans and the occasional Hatchimal the robot drops on the floor. The test is therefore unrealistic.
Watch though. They're going to start asking human warehouse employees to implant network chips so the robots can know where they are at all times. Problem solved if you envy a Skynet sort of future.
2
For a good may years now, we have been reading the same article about AI. It goes like this:
"AI demonstrates amazing abilities to do ZZZ. But experts doubted that it could ever master the more complicated version, ZZZY." One year later: "AI has just beaten all human experts at ZZZY. 'I can't believe it,' said ZZZY Master Joe McDoakes. 'But I doubt that this skill could be useful in the real world.' Eighteen months later: "A consortium of BMW and Nissan are using ZZZY to guide self-driven automobiles by smelling the exhaust of others near them."
And so on and so on. Soon such articles will be written by AI! Maybe it's already happened? Does the NYT know for sure that I'm human?
10
@joel Bergsman in fact, there is media content generated right now through AI available through many outlets. Coverage of market returns, sports, etc. where the content summarizes past events definable by statistics and scores. Your question is the right one: "Maybe it's already happened?" Indeed it's happening all around you. Another question might be: "Where will it lead?"
7
Everything done so far has been done in domains -- like these games -- with strict rules and clear goals. That's not real life. Call me when AI can play games whose point is unclear even to humans.
3
In the evolution of things we will get to a point where we transcend the body and can download ourselves into whatever medium we want. Iain Banks' sci-fi masterpieces, a series named The Culture Novels ,have people living forever. Want to be a man for a while? Want to be a woman and have some kids? Go ahead. Then add a third arm to help your sculpture technique.
But until that time, the issue that plagues me is that manufacturers who were making a decent profit while employing human beings opted for increasing their profits by losing humans in favour of automation.
Growing pains? Worth it?
1
Okay, but A.I. will never win at Strip Poker.
I hope.
13
Unintended consequences seem to have no play in the world of Wall Street greed.
Only one presidential candidate is addressing the next revolution in the world of work. He believes that we have to do better otherwise large percentages will be starving in the streets.
Unless you are in the elite educated world there is no place for regular citizens. We have intentional poverty ,poor education ,racial issues supported by politicians and their short term thinking policies across the world.
So what happens to the work force of today? Walmart is using robots now slowly merging them into the daily work routine. Japan uses robots to care for seniors due to a work force shortage based on their racial ethic issues.
Should robots build more for profit prisons,since they do better at construction jobs than humans.The FDR safety net is being stolen from the American people. Every person wants to feel useful and part of a network . This article reinforces the belief only special people need to be part of life.
Well science fiction has become our New reality.
Welcome to the new world of A.I.
1
Once computers can play Stratego, humanity is doomed.
2
I'm excited about AI being the personal teacher for a human - teaching cooperative teamwork along with perfect substantive instruction.
Imagine if we could increase the average knowledge and skillset of a nation by a standard deviation?
"Now A.I. is winning at capture the flag. Will such skills translate to the real world?"
Capture the flag is the real world.
8
Game theory is designed to model the real world. Just ask John Nash.
So A.I. is not there yet. It will get there soon enough.
The wealthy will want it to exploit the rest of us. We will all want it for medical advances so we can be healthier and live longer. The military will want it for dominance. Humans will keep working to improve it. We will never be able to let it go.
And in the end, it will be the end of us.
10
Perhaps the 737 MAX software patch should take the human pilot out-of-the-loop, since autopilots are statistically safer and crashes occur when humans and autopilots are in conflict.
4
Unimpressed. Games (or any tasks) that are hard for humans are easy for computers...and vice versa. When this is the no longer true, I'll care. And I doubt I will ever see it.
3
What all AI systems lack except in rigid rule games is situational awareness. As an example, self-driving cars only react to what they can see in the direction they are traveling. They do not, for example. exhibit awareness of a vehicle approaching from behind and driving erratically, or the significance of a ball rolling out from between parked cars from a playground not visible, even if their map shows that there is a playground there. After all, the playground is not an obstacle to be considered a threat to the car's safe progress.
But a human driver familiar with the location, if paying attention, should immediately realize its significance and hit the brakes in anticipation of a child, oblivious to the danger, running out to retrieve the ball.
Their is an old residential area in Austin known as Rosedale. Narrow streets, few sidewalks -- and Ramsey Park, a favorite for families. The posted limit is 25 MPH -- but, passing the parked cars that create a visual barrier between the single traffic lane and that park: I never exceed 20.
Google found that their vehicle's religious adherence to speed limits interfered with traffic on entering an expressway -- drivers routinely exceed the posted limit here, even in the right-most lane -- or caused the vehicle to come to a stop at the foot of the on ramp, waiting for a break. After several years of experiment, they packed it up and left.
5
that is not true. the lidar on a car (like a Waymo) sees 360 degrees quite some distance in every direction and would handle the situation much better than a human who can't see through objects or in all directions at once.
7
The process of complexity and intuitive feedback learning allows for algorithms that build upon dynamic interaction. A situational awareness HUMINT approach to problem solving and why DeepMind is a grey matter maestro.
1
From the article: (Even mere ants can collaborate by trading chemical signals.) Mere ants don't need 45,000 years of training to play their game. Take that from the lovers of social insects. Oops there goes another rubber tree.
1
Let’s ask E. O. Wilson how many years it took for the ants to evolve their means of communication.
5
I want AI to take over Public Defenders responsibilities in every rinky-dink courtroom across America...Let every poor defendant have the best legal mind at their service...Wouldn’t that be great? And safer than AI driving us through midtown traffic...
7
How are the military implications not even mentioned???
10
Perhaps the Deepmind can be repurposed to solve the increasing incidences of autism among the US population. Researchers at the Princeton University are reported to have had some success in identifying the DNA anomalies in the 99% of the human genome that have not yet been studied fully. AI is being used to unravel the complexity created by various permutations and combinations of interactions and mutations among the 99% of the genome that cause autism.
https://www.nature.com/articles/s41588-019-0420-0
2
Grat. We ding a ling humans are paving the way for AI to take us out of the picture.
4
And this is supposed to be a good thing?
4
"Go, the Eastern version of chess." What, pray tell, is the Western version of chess?
19
Isn't this type of machine learning exactly what was done in the 1983 move WarGames?
4
But after winning, will AI "know" it won, or "celebrate" winning.
1
Not in my lifetime, not in my daughter's. But in maybe one more generation articles like this will look quaint.
We are inevitably headed towards robotic influence in every facet of life.
Shakespeare had it right: nothing is good or evil but thinking makes it so. So it is with science. For good or evil ? Our thoughts ( motivations) will determine the outcome.
Cross your fingers.
5
Very fascinating. To dumb this down, it appears this AI is "learning" the same way a rat does in a maze: By trial and error. It is a simple idea, while simultaneously incredibly complex to achieve. If you are a TV Sci-fi fan, this reminds me very much of the Replicators on "Stargate Atlantis" (2004-2009). It was and still is, a very popular TV series; the middle of 2 others which ran consecutively for 15 years (over 320 episodes). All are based on the smash hit film "Stargate" (1994), starring Kurt Russell and James Spader.
4
The "human skillset" in gaming is trying to outwit your human opponent as to whether they will go left or right around the vehicle they're taking cover behind. Or do they pop up in the middle? Or are they grinding the camouflage challenges on a weapon they're not used to so they aim too high for a headshot and miss? The reason playing against humans vastly outweighs playing the computer AI, as it currently is, is because humans are just so unpredictable. Or completely predictably camping in the same corner for the whole match. Can A.I. really capture the essence of that chaos?
1
@OnABicycleBuiltForTwo
It will.
4
“Chess and Go were child’s play”?! That’s a first! Never heard those games referred to in such a way! To the contrary.
4
The advances in AI are incredible, no doubt, but I'd quibble on calling Capture the Flag a game with a "human" skillset. It requires following set rules in a contained area, to achieve set goals. We know AI is capable of learning in that fashion.
"Human skills", what distinguish us from other animals (and from AI) rely on understanding other humans and meta concepts like society. Discretion, art and creative writing, humor, irony. These things all require adaptability, which computers do have, but something in addition: an understanding of humanity and human circumstances.
Irony: What do people expect? Challenge that.
Discretion: why might a person make one choice over another, if the pros and cons are weighted equally?
Humor: who holds societal power and what are the established social conventions? How can they be subverted or exposed as foolish?
I'm not worried about computers replacing us everywhere. There will always be a place for humans in a human society, as there are some things computers can never do as well. But we must learn to value those things and ensure we don't leave people behind during this massive technological shift.
5
the darker undertone is the power this technology entails. the political context this will be borne within (dwindling resources due to climate) implies a post apocalyptic potential. when destruction occurs through mechanical and ai brilliance I fear an outcome baked in unprecedented destruction. prayers for peace
1
Just a point of information, (i)go is NOT the Eastern equivalent of chess. The stones in go do NOT make specified moves according to a role (pawn, queen, bishop, etc.) You are thinking of Japanese shogi, which is very chess like.
8
@Robert Hancock
And Chinese chess, as well.
2
Put AI to practical use: maybe the AI techniques should be used by one or more of the Democratic candidates to learn the characteristics of Trump as well as the reactive behavior of the different media outlets, essentially a "competing team", that can in turn exploit his online and base weaknesses in the media world with the correct countermeasure messaging. This interactive strategy, similar to a game strategy, could be used to put him continually on the defensive as a "game winning" strategy. As opposed to the Russian approach, it would use real information about behaviors to produce revealing responses.
12
This underlines the importance of play time for kids and adults to improve their overall skills, too. Less time in the classroom and office and more time on the playground having fun!
7
Computers and machines will always out perform people in routine regimented actions that can be reproduced and run in loops. That’s what computers do and they do it extremely well and quickly on the software level. The hardware would have to catch up.
Some day if we had the tools to move electrons and atoms around with computers we could literally print whatever we wanted.
Just imagine if we could scan every atom of the human body and compare it each year to last years profile. We could identify abnormality and growths almost immediately.
The software is there in my opinion but the tools to do this aren’t yet there.
2
@Mathias The way neural networks learn to play video games, compose new music and stories, paint new scenes in the style of a given artist, recognize health problems from scans, etc. is nothing like "routine regimented actions that can be reproduced and run in loops." That's like saying humans are good at cooking because our brains work by bringing organic chemicals together.
2
All your base are belong to us. And soon, all your jobs too. But as a software tester, I think this is fantastic and that we have to do it. Aren’t we better off with Excel than a million accountants? And what are all the displaced accountants from the 70’s doing now? Something more lucrative and interesting? Able to do more accounts? I think this will only lead to good.
12
@CCF
I largely agree, but at some point we'll run out of those "more lucrative and interesting" things that humans are solely capable of doing. It won't be for a few decades, but within the next 50 years I would expect that nearly any task that (1) requires data input, (2) involves some sort of process that can be modeled, and (3) produces some clearly-defined output will be automated. This will start with basic tasks, but eventually we'll automate doctors (no need for humans to inaccurately diagnose, when machines are better at reading tests and scans), lawyers (textual analysis of millions of legal texts and case law docs, identification of key issues to the current matter, application of those issues in a connected way that advances a core argument - all done on a computer in minutes rather than by attorneys billing $800 an hour), engineers (pretty simple to give a series of requirements and constraints to a machine, have it analyze billions of possible designs, and produce one that optimizes for a series of desired outcomes), and even journalists (synthesizing important points from a body of information, adopting the style fed into it via a database of existing sample articles - say, the NYT archives - then composing all of that into a logical text).
Eventually, nearly all human intellectual labor will be automated. There will be nowhere to "go" next, unlike with agricultural and industrial revolutions, and service industry shift in 20th century. What do we do then?
28
@Alex
We die out except for bands of foragers living in biological preserves that haven't been flooded or dessicated by climate change.
4
@CCF
Pretty soon all your human are belong to us.
5
Great article. Shows the advances of this technology. But this will not be taking over the world anytime soon. There are no A.I. dogs, cats, or even mice. The reason is it requires entirely different skills not even remotely related to gaming. Automation is perfect for A.I. , intelligence is truly a different kind of animal.
8
@Pete Artificial Intelligence *is* intelligence that achieves goals for us. I'm not sure what part of being a mouse you think an AI can't match (finding a power socket to "feed on"?), but the fact that animals remain qualitatively different from robots is scant comfort to the hundreds of millions of workers soon to be replaced by AIs and robots that do their jobs better.
4
"Today, human workers handle such tasks" And when the humans no longer do, have we a plan for them? I rarely read a tech article that asks this basic question.
16
@Adam Possibly because the answer to this question can (does, and will continue to) fill several books worth of information. There isn't space in every article to devote to this issue. And it isn't always relevant to the topic of the article.
Mr. Metz concludes with:
"Though machines can now win capture the flag in the virtual world, they are still hopeless across the open spaces of summer camp — and will be for quite a while."
This may be small comfort to humans. As we become ever more existentially dependent on the Internet, A.I. may soon come to view our "real world" abilities as irrelevant in their virtual world.
5
@Paul Central CA, age 59
"... A.I. may soon come to view our 'real world' abilities as irrelevant in their virtual world."
And what if we're already living in someone's virtual world? Maybe A.I. is just a simulation within a simulation?
In any event, if we want to worry about the "real world," we need to come to grips with the evolutionary fact that humans are not designed to be around forever. In this real world, anyway.
5