« Libertarian red meat | Main | My new job »
AI, Consciousness and Robot Outsourcing
One of my "absurd views" is that the first computer to become conscious was Deep Blue playing against Gary Kasparov in 1997. It only happened for a moment but in one spectacular move Deep Blue performed like no computer ever had before. After the game, Kasparov said he felt a presence behind the machine. He looked frightened.
Ken Rogoff, a top-flight economist and chess prodigy, wonders whether we don't all have a little something to fear.
But the level that computers have reached already is scary enough. What’s next? I certainly don’t feel safe as an economics professor! I have no doubt that sometime later this century, one will be able to buy pocket professors – perhaps with holographic images – as easily as one can buy a pocket Kasparov chess computer today.
Rogoff thinks that the upheavals caused by cheap AI will be far more important than those caused by low-wage labor from India and China.
...will artificial intelligence replace the mantra of outsourcing and manufacturing migration? Chess players already know the answer.
Posted by Alex Tabarrok on March 23, 2006 at 07:14 AM in Economics, Science, Web/Tech | Permalink
Comments
"...will artificial intelligence replace the mantra of outsourcing and manufacturing migration? Chess players already know the answer."
Nah. Chess is a very limited domain and chess computers have been custom-tailored to that limited domain. Very little that has been invented in the course of creating grand-master level chess computers is of use for general purpose intelligence.
The original idea behind chess-playing AI was that if researchers could get machines to master such an intellectually challenging task then surely, along the way, they'd necessarily solve the problems of creating intelligent machines in general.
But it didn't work out that way at all. Chess is intellectually demanding because it is a task that human brains are not all that well suited for naturally. It turned out that the *really* hard problems that human brains solve are ones that every normal 5-year-old can handle effortlessly--recognizing people's faces, for example, or understanding language. It took years/decades of failure for AI to solve these problems (or even make very much progress) to reveal how hard these problems are (an odd sort of contribution for AI to make, but a valuable one).
So don't sweat the 'AI professor'. In fact, here's an absurd prediction for you. Computer chess will follow the pattern of manned missions to the moon. The creation of chess playing computers that can beat the best human players will eventually kill the whole high-level research program. After that, there'll be no point, people will lose interest, funding will dry up, etc. Chess software will still be around for training and amusing human players, of course, but "Deep Blue" will end up a museum piece.
Posted by: Slocum at Mar 23, 2006 7:52:01 AM
Would rather learn from an AI economist, or an economist who can also tell you about his favorite ethnic foods?
IMNSHO as a computer scientist, AI is a historical mistake. The real gains are in better ways to access the intelligence of other humans. I'd bet on a holographic ipod with recordings of Alex or Tyler, and a voice search interface which can locate their answers to your questions.
p.s. IBM junked Deep Blue as soon as the match was over. A 1-0 undefeated computer is better for advertising than a 1-1 computer.
Posted by: DK at Mar 23, 2006 8:32:20 AM
....Computer Artificial Intelligence can not handle cognitve dissonance -- a much higher level of analog data processing that humans rely upon quite heavily for survival.
Posted by: Danzig at Mar 23, 2006 8:55:01 AM
The substitution of computer software for human labor is already happenning. Tax filers buy programs instead of hiring accountants. In my field, engineering, creative work is increasingly migrating to the creation of software to be used by others to churn out results. Many firms are licensing expert systems that free users from understanding analysis methods.
Posted by: Mansfield at Mar 23, 2006 8:57:18 AM
Alex,
You don't understand the colossal failure that is strong A.I. Go do some reading about the history of artificial intelligence, the bold predictions from the 1950s, 60s and 70s, and the relentless rollback of those visions in the 80s and 90s. Go read about Cyc and then look at where it is now. We are all still on square one. In the past half-century, computer scientists have made no demonstrable progress toward understanding the nature of what we call "intelligence."
Slocum summarizes the work in A.I. quite well. Basically the researchers' contribution to date has been to demonstrate, via repeated failure, that the problem of building an intelligent machine is so hard that it's beyond our current understanding.
There have been many interesting theoretical ideas about intelligence tossed around. And essentially everyone agrees that an intelligent computer is theoretically possible. It's just that, in 2006, we have no idea whatsoever about how to begin designing one.
That's not to say it will never happen, or that predictions about "this century" are necessarily optimistic. Almost anything could happen by 2099, if the right technological breakthroughs come about. But if you think world-class chess moves have anything to do with the development of an artificial college professor, you simply do not understand the field.
Posted by: Jamie McCarthy at Mar 23, 2006 9:12:06 AM
I'm familiar with the failure of strong AI and the history of the field. I won't go into details about why I am currently optimistic because of the point that Mansfield makes, expert systems are already good enough to handle 80-90% of what any human does in a wide variety of fields. I see no reason, for example, why expert systems should not replace a large number of family physicians. Personally, I would prefer diagnosis from an expert system because the mean accuracy will be higher.
In anycase, the point is that in the near future AI will be to many white-collar workers what China and India are to blue collar workers today. This will happen with modest progress, no big breakthroughs are needed.
Posted by: Alex Tabarrok at Mar 23, 2006 9:27:36 AM
Deep Blue, huh. Read about Go and computers somewhere, that's a better measure of the current state of AI.
Posted by: A Tykhyy at Mar 23, 2006 9:38:17 AM
I think your prediction about "A.I. outsourcing" will depend very much on how fast humans can enhance their own intelligence. If there are breakthroughs in genetic engineering or the ability of humans to otherwise enhance their intelligence, then I predict there will still be plenty of jobs for humans to go around.
Posted by: Javier at Mar 23, 2006 10:13:31 AM
Even if this was a problem, and it's not *currently*, we had better make peace with the Cylons before they make peace with US. I don't really like their style.
Posted by: Admiral Waugh at Mar 23, 2006 10:18:02 AM
Many people, including economists, simply refuse to believe what data tells them. A computer might have a problem doing that.
Posted by: cactus at Mar 23, 2006 10:28:21 AM
"...expert systems are already good enough to handle 80-90% of what any human does in a wide variety of fields"
I think this applies only to fields that require large bodies of reference data, like a medicine or geology. It seems unlikely that weak AI would have much effect on fields that are driven by creativity or innovation - which would include all area of human endeavor to some degree. This would be a good thing too; it would free up more human intelligence and leave the drudgery to computers.
Posted by: Todd Fletcher at Mar 23, 2006 10:49:35 AM
Alex,
I am in complete agreement with you. Computers will not need to be general AI, they will be expert systems programs, something that masters an individual area of knowlege. It's going to start happening very, very soon. Within 7 years, many white collar jobs are going to be heavily impacted by the effects not only of cheap Indian and Chinese professionals, but of expert systems replicating their knowledge.
Within 3 years, keyboards are going to start being phased out as good voice recognition software takes over, so average input will go from 40 words a minute to 120-150. Within 5 years, many programmers are going to realize that their IDEs are writing more of the code than they are. What is Westlaw going to offer in 7 years? Well we can guess at that. My job is going to be largely replaced by automation, I run the operations of an electronic exchange. In fact, I am responsible for writing Excel tools that have saved a full person from being added to our department. I don't think I am smarter than the average bear, I bet every person here has done this in their own jobs, used technology to replace at least 1/2 of a person over the last two years.
As the hardware becomes more and more powerful, we're going to be using the same method that Deep Blue used to defeat Kasporov - Brute Force - to simulate human responses to very specific knowledge. Deep Blue has been replaced by a much smaller computer, one that tied the current best player in the world. This combination of simple brute force calculation and even slight improvements in developing expert systems is going to be the doom of the white collar worker. Its already happened in my former field, trading, where the people who used to arb between cash and futures and options, have been completely replaced by far fewer people running computers.
Check out kurtweilai.com. I don't think all of Rays predictions are going to be true, but one is. And that is that a $1000 computer will have the same complexity as a human brain, sometime around 2023. I don't think this computer will be as smart as a human, but in narrow areas of knowledge, this computer will be far, far superior to humans. And 2023 is only 17 years away. The scary thing is that computers will keep getting more complex. Ray makes this hilarious and scary prediction that a $1000 computer will be as complex as ALL (!) human brains, sometime in the 2050s!! But if Moores law holds, thats where computers will be.
Its a matter what we are going to do with all these now useless, idle people, in the year 2030. My son will be 29 years old.
Posted by: mickslam at Mar 23, 2006 10:57:18 AM
One more thing. We can look back 30 years, into 1976, to see where the state of the art in computers was at the time, and where they might be 30 years in the future. As - my PC is to banks of spinning magnetic tape: my pc will be to X.
Do you have an iPod? I wonder if a 40G iPod that you hold in your hand has more memory than existed in the entire world in 1976.
Posted by: mickslam at Mar 23, 2006 11:03:19 AM
Is solving a game of complete information all that spectacular? It's a function of processor speed, isn't it? It's when AI can solve games of incomplete information with a stochastic process that we need to be afraid. The team at Alberta is busy at this, too, though. But so far, I think their bots are great ring game players at limit poker against average and slightly above average players (maybe), but can't compete yet at the highest levels of skill. At least, that's what I've heard.
http://www.cs.ualberta.ca/~games/poker/
Posted by: scott cunningham at Mar 23, 2006 11:46:27 AM
Is solving a game of complete information all that spectacular? It's a function of processor speed, isn't it? It's when AI can solve games of incomplete information with a stochastic process that we need to be afraid. The team at Alberta is busy at this, too, though. But so far, I think their bots are great ring game players at limit poker against average and slightly above average players (maybe), but can't compete yet at the highest levels of skill. At least, that's what I've heard.
http://www.cs.ualberta.ca/~games/poker/
Posted by: scott cunningham at Mar 23, 2006 11:46:34 AM
I am skeptical that software written in the usual way will displace humans from very many job tasks in the next half century or so, but by mid century or so brain simulations will be feasible, and those surely will displace humans from most job tasks.
Posted by: Robin Hanson at Mar 23, 2006 11:47:02 AM
> I am in complete agreement with you. Computers will not need
> to be general AI, they will be expert systems programs,
> something that masters an individual area of knowlege.
'Expert systems' to replace experts have been 'just around the corner' for 20-25 years. Their impact has been limited. That's not saying that in particular domains, expert systems might not work reasonably well. But consider family physicians--there are various alternatives that could do what American general-practitioners do at much lower cost including (nurse practitioners, foreign doctors working remotely, etc). The hurdles there are legal, not technical.
> It's going to start happening very, very soon. Within 7 years,
> many white collar jobs are going to be heavily impacted by the
> effects not only of cheap Indian and Chinese professionals,
> but of expert systems replicating their knowledge.
Where's the betting market where I can bet against that happening (white collar jobs being heavily impacted by expert systems)?
> Within 3 years, keyboards are going to start being phased out
> as good voice recognition software takes over, so average
> input will go from 40 words a minute to 120-150.
Again, where do I bet against that happening? I work on a computer most of the day -- no matter how good voice recognition gets (and, BTW, it's not going to get that good any time soon) I dont *want* to talk to my computer all day. Can you *imagine* how annoying it would be to sit in an office full of cubicles or a coffee shop and listen to people talking to their computers? Morons on their cell phones would be nothing in comparison. And most of what I do on the computer does not involve generating streams of text as fast as possible. What I get paid to do is think and generate sometimes a very small amount of text, but it's the right text in the right place. And how are you going to edit with voice input?
> Within 5 years, many programmers are going to realize that
> their IDEs are writing more of the code than they are.
That's not a new phenomenon -- development tools generating more and more of the routine 'boilerplate code' has been going on for close to 20 years. That's one of the main ways that the productivity of programmers increases.
Posted by: Slocum at Mar 23, 2006 11:59:32 AM
Also, voice recognition seems highly impractical for deleting and making corrections, at least in my experience. I have a much easier time composing silently at the keyboard then speaking the corrections aloud. Voice recognition seems better suited for navigation then composition.
Posted by: scott cunningham at Mar 23, 2006 12:15:48 PM
Ditto the naysayers on voice recognition. A whole host of problems there, noise pollution of the work environment one of the biggest. Far easier to use a keyboard in most cases where a keyboard is used today.
Voice communication works well between humans because of shared contexts and understandings and body language, etc. Consider the relative non-productivity of conference calls, for example. If we had to speak in full sentence form all the time speech would be far less productive.
Posted by: Matthew Cromer at Mar 23, 2006 12:22:41 PM
I work in AI & Robotics.
One thing you're neglecting: human enhancement. It won't be "us and them". It will be untouched humans, enhanced humans, pure machine intelligence.
Robots should work. Humans should think. It's that simple :)
Posted by: Ivan Kirigin at Mar 23, 2006 12:46:34 PM
> Its a matter what we are going to do with all
> these now useless, idle people, in the year 2030.
> My son will be 29 years old.
retirement at birth?
http://www.marginalrevolution.com/marginalrevolution/2005/12/must_i_retire_n.html
my current thinking is that eventually, humans will only be able to market their creativity. we need better idea / prediction markets to make sure you can derive a lifetime of income from your one great idea. maybe patent law needs to be reformed along those lines?
the other, a bit tedious scenario from 'the end of work' by jeremy rifkin is that we will all be working for NGO, which are somehow protected from the market forces. i'm not holding my breath on that one.
Posted by: Gregor J. Rothfuss at Mar 23, 2006 1:06:07 PM
> I am in complete agreement with you. Computers will not need
> to be general AI, they will be expert systems programs,
> something that masters an individual area of knowlege.
>'Expert systems' to replace experts have been 'just around the corner' for >20-25 years. Their impact has been limited. That's not saying that in >particular domains, expert systems might not work reasonably well. But >consider family physicians--there are various alternatives that could do >what American general-practitioners do at much lower cost including (nurse >practitioners, foreign doctors working remotely, etc). The hurdles there are >legal, not technical.
In many, many fields there are no legal barries, like my field, trading. In the past there were literally hundreds of people standing around, making markets in individual stock options and individual stocks, individual futures. Now, about 1/2 as many people are making markets on computers. Thousands of extremely high paying jobs disappeared. Additionally, the biggest hedge fund in the world is a quantitative hedge fund, D.E. Shaw. In other words, they just run numbers and simulations, then program a computer to execute the trades. Here is a cool quote from Nature magazine, in there most recent issue.
"For example, within the past decade researchers in my group have used inductive logic programming (a subdiscipline of machine learning) to discover key molecular substructures within a class of potential cancer-producing agents2. Building on the same techniques, we have more recently been able to generate experimentally testable claims about the toxic properties of hydrazine from experimental data — in this instance, from analyses of metabolites in rat urine following low doses of the toxin3."
Can you say 'expert systems'?
> It's going to start happening very, very soon. Within 7 years,
> many white collar jobs are going to be heavily impacted by the
> effects not only of cheap Indian and Chinese professionals,
> but of expert systems replicating their knowledge.
Where's the betting market where I can bet against that happening (white collar jobs being heavily impacted by expert systems)? Longbets.org
> Within 3 years, keyboards are going to start being phased out
> as good voice recognition software takes over, so average
> input will go from 40 words a minute to 120-150.
>Again, where do I bet against that happening? I work on a computer most of >the day -- no matter how good voice recognition gets (and, BTW, it's not >going to get that good any time soon) I dont *want* to talk to my computer >all day. Can you *imagine* how annoying it would be to sit in an office full >of cubicles or a coffee shop and listen to people talking to their >computers? Morons on their cell phones would be nothing in comparison. And >most of what I do on the computer does not involve generating streams of >text as fast as possible. What I get paid to do is think and generate >sometimes a very small amount of text, but it's the right text in the right >place. And how are you going to edit with voice input?
In most offices, people are talking constantly on the phone anyway. Also, you type in full sentences, so talking in full sentences won't be a huge streach. Edit with a PS2 style controller to move around the screen. If you are only writing a little bit anyway, what does it matter how it gets into the computer?
> Within 5 years, many programmers are going to realize that
> their IDEs are writing more of the code than they are.
>That's not a new phenomenon -- development tools generating more and more of >the routine 'boilerplate code' has been going on for close to 20 years. >That's one of the main ways that the productivity of programmers increases.
Posted by: mickslam at Mar 23, 2006 2:20:53 PM
As a chess addict and a not very good player, I am very familiar with chess programs/computers. They can be quite frightening. In games under 5 minutes, they can beat grandmasters of any level without almost any effort. Currently on internet chess servers, some computers can play 40 moves in under 5 seconds towards the end of the game.
All that said, it took 60 years for that to happen and when it did computer scientists found out that they couldn't use what they had learned for other problems. For an indepth discussion on this (2.5 hours), checkout video.google.com and search for CHESS. There you will find a conference by the Museum of Computers on the History of Chess Computing.
Who knows? Maybe the Singularity is coming. I really do hope so... enough of competition. I want to be an art`ist!
Posted by: RWP at Mar 23, 2006 2:30:24 PM
Computers will think when cars fly.
Posted by: bjkmp@mp.com at Mar 23, 2006 2:41:21 PM
The experience from computer chess is that in some problem areas, there's a minimum amount of computer power that's necessary to solve the problem. Early chess program researchers thought that smart programs could avoid doing so much work, but those smart programs were eventually overwhelmed by systems that did massive search (with some algorithmic cleverness to improve the efficiency with which the space was searched, but without excessive pruning).
What does this imply for AI? Perhaps the general AI problem is similar, in which case lack of progress to date merely means we don't have computers anywhere close to fast enough to let us make progress. In this case, lack of success so far would not mean there will be no success in the future.
ObSF: _Accelerando_, by Charles Stross.
Posted by: Paul Dietz at Mar 23, 2006 3:22:56 PM
