Why Human-Level AI Won’t Change the World Thursday, Aug 14 2008
AI and singularity 3:31 am
One position I have difficulty wrapping my head around is the position held by those who believe that human-level AI is possible but that it would lack the capability to quickly change the world. The reasons for why AI would likely have that capability are frequently cited. To summarize just a few:
1) AI could quickly and easily be copied as many times as is computationally feasible.
2) Running on a flexible substrate, AI could “overclock” their cognitive functions, leading to enhanced intelligence and capability.
3) Though robotics today is still maturing, it will be more sophisticated by the time AI arrives, and with AI’s help, it isn’t unreasonable to assume that AIs will have direct and broad access to the physical world through robotic means.
4) AIs would be able to share thoughts almost instantly, meaning that skills learned by one AI could be transferred to all other AIs very quickly.
5) AIs would be able to quickly and automatically perform tasks considered by humans to be “extremely boring”, but still pragmatically useful.
6) AIs could routinely perform intellectually demanding tasks for just the cost of the computer it runs on, plus electricity.
So, brainstorming the reasons why human-level AI would exist but lack the capability to quickly change the world:
1) Human-level AI might possess human skills and intelligence but lack free will, making them incapable of modifying the world in any real sense.
2) Humans will deliberately prevent AI from doing so.
3) AIs would need to be embodied to do anything, and there currently isn’t enough room on the planet for that many embodied AIs or the infrastructure to support the resources they would consume.
4) I object to the idea of human-level AIs in general, thus when the prospect of such AIs changing the world is brought up, I object to its feasibility, while concealing that I reject the premise outright.
5) Humans are equivalent to the most intelligent entity possible, therefore AIs will never be smarter than humans, and will lack any huge impact. (Sometimes this is phrased as saying that humans and AIs are both Turing complete and will thus have the same capabilities.)
6) AIs will just exist on the virtual layer, and being virtual beings, will always have highly limited access to the physical layer.
Any others I’m missing? If there are any actual papers with people presenting points in this vein, that would be ideal.
August 14th, 2008 at 4:17 am
Human level AIs have free will, are beyond our control and spend their time on tasks that don’t much effect us.
August 14th, 2008 at 7:40 am
The update and knowledge sharing problems could be slower and a non-trivial problem.
If the AIs have something akin to enterprise software levels or data warehouse levels of data volumes, then data transfer (ETL) and data integration problems could be substantial and slow.
Minor message communication is one thing but major knowledgebase updates would be a big deal and reliability issues would need to be worked out.
Plus there would need to be AI decisions on whether the major update would be beneficial and not something that would crash or degrade the AI.
======
Regulatory issues and delays for robotics.
I have written about using UAV robotically flown planes (particularly electric planes and hybrid planes) to revamp transportation. The computers, sensors and robots could be ready but regulations will slow deployment.
Robotic cars could be ready if robotic car and systemic policy choices were made to introduce them but insurance, legal and societal issues could and probably will block/slow it. [FDA, FAA, city councils, state governments, unions, lawsuits etc…]
Japan is more robot friendly.
Probably 3-4 million service robots by 2010. Over 1 million industrial robots now.
How much will optimal non-AI systems raise the bar for better AI to make more of a contribution.
If we get: really good quantum computers running Grover’s algorithm before really good AI, which is the mathematically optimal search algorithm. N**0.5
Then AI cannot improve on certain functions which are already mathematically optimal. Although they could find superior approximations that use less resources to do the same thing.
Human enhancement (particular radical cognitive enhancement and computer interfacing) could takeoff pre human level AI. there could then be a merging of the improvement paths.
August 14th, 2008 at 9:15 am
While I accept the likelihood of human level and greater AI being achievable, it seems to me that many of your claims for why it would be likely to rapidly change the world are far from proven given our current state of knowledge of the nature of human intelligence and the requirements for reproducing equivalent intelligence in an AI.
“2) Running on a reprogrammable substrate, AI could “overclock” their cognitive functions, leading to enhanced intelligence and capability.”
If by “overclock” you mean what is meant by overclocking in the PC community - i.e. just cranking up the clock speed of the hardware - then yes, it seems likely that an AI could be made to think faster in proportion to increases in hardware speed.
By mentioning a reprogrammable substrate though you seem to imply something different, namely that an AI could automatically be capable of reprogramming itself to improve its own efficiency. Very few humans have the necessary skills or intelligence to understand and improve software at all and so far precisely zero are capable of creating AI that approaches human level. It’s not clear to me that a ‘human level’ AI would necessarily be at the level of the so far non-existent human who could write or improve such software, rather than at the level of the typical human who could not.
“3) Though robotics today is still maturing, it will be more sophisticated by the time AI arrives, and with AI’s help, it isn’t unreasonable to assume that AIs will have direct and broad access to the physical world through robotic means.”
Not unreasonable perhaps, but by no means guaranteed either. Perhaps robotics will lag AI considerably when a breakthrough occurs (though it would seem at the moment that the lag is more likely in the other direction).
“4) AIs would be able to share thoughts almost instantly, meaning that skills learned by one AI could be transferred to all other AIs very quickly.”
Given our lack of knowledge about how human AI might be achieved it seems an unwarranted jump to assume this. The only human level intelligence we know has yet to find a more efficient way of communicating thoughts than language. It’s not clear to me that an AI that was something like a large bayesian network or a neural net would be able to easily extract ‘thoughts’ and transfer them to an AI running on the same software but with a different state and simply merge them into that state effectively. I don’t believe we understand the nature of a ‘thought’ or the details of how a human level AI would be constructed well enough to take this premise as a given.
“5) AIs would be able to quickly and automatically perform tasks considered by humans to be “extremely boring”, but still pragmatically useful.”
I don’t see any good reason to take this as a given. If human level AI is developed partly from increased understanding of human intelligence and modelled on that understanding maybe it will initially inherit some human ‘defects’, like getting easily bored with repetitive or mundane tasks. I can imagine plausible accounts of human intelligence where ‘boredom’ would be a key driver behind creativity and other valued intelligence traits.
“6) AIs could routinely perform intellectually demanding tasks for just the cost of the computer it runs on, plus electricity.”
Again this brings up the question of whether by human intelligence we mean “average” human intelligence or “genius level” human intelligence. We already have 6 billion human level intelligences in the world that can in theory perform tasks for just the cost of food but very few of them can routinely perform “intellectually demanding” tasks.
Overall it seems to me that we are already progressing at the speed that comes with having 6 billion human level intelligences active in the world and it’s not clear to me that doubling or tripling that number is going to lead to singularity level increases in the pace of technological change. The singularity pace of change is predicated on the self-improving capacities of AI but given the difficulties so far evident in creating human level AI it would seem that ‘merely’ human level AI will not necessarily be capable of improving itself at a great rate initially.
August 14th, 2008 at 9:17 am
It occurs to me that #’s 2, 4, & 5 are really big “ifs”.
We’ve gone over that before… but otherwise, bravo, Michael.
August 14th, 2008 at 9:59 am
We dont need human level AI to bring home the point humans are nearly as intelligent or conscious as we assume we are. Even something approximating human level intelligence will wipe out 90% of existing jobs.
August 14th, 2008 at 1:37 pm
Hi Matt. For #2 I only meant brute acceleration of cognitive modules or the entire brain. I’ve changed “reprogrammable” to “flexible” to reflect this.
For #4, for copies of an AI, they’d be using the same general architecture, so it seems hard to imagine that exchanging knowledge sets would be more difficult than making straightforward changes to the code. Whatever “knowledge” is, it’s information in a brainstate, and if we’ve built AI successfully, then that brainstate (as far as the AI is concerned) will be known. We need not know how thoughts or knowledge are instantiated in human brains to achieve this.
With regards to #5, I guess I was envisioning non-neuromorphic AI here.
In any case, with a team I’m currently building a tool that will allow probabilistic modeling of various scenarios from here to AI, and an AI’s takeoff trajectory. This should elucidate everyone’s beliefs more clearly and explore their implications.
August 14th, 2008 at 2:03 pm
Copying a digital brain state is trivial, sure. But I agree with Wang and Matt that merging different brain states is nontrivial. Given AI_1 good at task A and AI_2 good at task B I’m not convinced there’s a way to produce AI_3 good at task A and B.
This might be an important set back if research becomes increasingly interdisciplinary.
August 14th, 2008 at 2:06 pm
Oh, and a reason AI might not make a difference:
Technological ceiling. By the time AIs are developed humans have essentially exhausted the search space for new and useful technologies.
Of course, AIs would probably still change the face of our economy even if they aren’t great researchers.
August 14th, 2008 at 3:49 pm
Seeing as advances in AI — towards AGI — are being made primarily in developing sophistication in knowing what data to set aside (example), it seems that forgetfulness and boredom may actually be much more important to a successful AGI implementation than many give credence. And that’s //without// AGI being neuromorphic.
The Novamente concept also works on something very similar; knowing //what// to pay attention to is almost as important as being able to process information intelligently.
Granted, an AGI won’t — very likely — have to face fatigue as a complication, but certainly all those “fuzzy” bits we seem to think of as biological ‘flaws’ that will be ‘done away with’ in Machine Intelligence … there’s just no guarantee that this is so.
August 14th, 2008 at 3:51 pm
How many service-sector jobs in the developed world could be done by eight-year-olds, if it weren’t for child labor laws?
It takes much less than “human level” intelligent property to drastically change the economy. Intelligent technology has been doing just that for centuries already.
This argument doesn’t seem to be about whether certain “levels” of intelligence are possible (yes, the question is unsettled, and bores me besides), but whether, given “human levels” of it, they would drastically change the world.
The world is full of human intelligence, of course. But the difference with artificial intelligence isn’t, then the “level” of intelligence, but it’s legal status.
An entity with level of intelligence being “equivalent” to a human is one thing; what I’m paying more attention to is the fact that labor is being incrementally transformed into property - something that can be owned, privatized, and monopolized. That by itself is an enormous political, legal, ethical, and economic difference.
August 14th, 2008 at 4:19 pm
The question of whether human intelligence is possible is settled. Humans. Add in a Copernican view of mindspace and you have the possibility of superintelligence.
I doubt “legal status” will matter. AIs will quickly become radically superior to humans, and it’s OUR status that we should be worried about. Status as in alive or dead.
August 14th, 2008 at 6:34 pm
I see a lot of ’stories’ trying to demonstrate the awe inspiring power of an AGI. Here’s my counterstory:
You’ve just woken up in a slime infested world populated only by amoeba’s. Worse yet, you find upon inspection even your body is depressingly amoeba-like. Luckily your intelligence out strips all those other amoebas by leaps and bounds: even their combined mental agility from trillions of amoeba cells can’t scratch your impressive human-brain-miraculously-encased-in-an-amoeba intelligence.
You should be able to make quick work of your plans for global domination.
Go.
August 14th, 2008 at 7:28 pm
What could be achieved with military level funding - essentially unlimited?
If for example all world bankers go bonkers for transhumanism, what would we do with the trillions they’d be showering us with?
In other words, is funding a major problem? Could chip manufacturers speed up development if they had unlimited cash?
August 14th, 2008 at 10:33 pm
Mil-AI - bear in mind that an AI industry, in a shape that “produces interestling volitional general AI you and I would get hot on”, is hard to bootstrap. Industries producing AI’s would totally produce slaves. I mean - devices that can do things smartly without being smart. If you need an input device that can take human instructions, isolate that from the smart working parts - for example, if you want taxi cabs moved around a city “intelligently”, you’d put the motion parts in the cab, with a very lean “johnycab” service engine, you implement a “smart” traffic switching device on a citywide level, and finally you add the devices that are smart at communicating with clients and link them to GoogleCab or whatever.
These three distinctly smart devices should be profitable (incentive for making them) but by themselves they can do perfectly by being dumb and narrowminded. And if some lonely, depressed human ends of wanting to talk to GoogleCab using his VR glasses, that service can instantly transfer the human (without the human even noticing) to a more specialized psychiatric service provider.
My point is, we may easily end up for years in a society that has no sensible function for AGI. And hopefully it would take a few years for garage hackers to stitch up these disparate parts into an actual AGI.
Why would that be bad? Because in our current societies, where the value of a human is NOT intrinsic (visit a local emergency ward in a US hospital or an old age home and see how valueless some humans are in even the richest of earth societies) but emergent (watch someone with ambition ass-kiss someone with a few million $$ - I have seen rather gruelling examples recently) and you’d realize that as soon as we have intelligent competitors we will have a problem not necessary with the AGI (they will probably be willing slaves) but with the rich superelite of humanity (the cold methodical bastards who became rich because of being cold methodical bastards) who would then have no incentive to keep the rest of (non- eager slave) humans around. The ultrarich 1% would have their servants, they wouldn’t need “us” anymore.
August 15th, 2008 at 2:35 pm
It would be interesting to know
How many people does the civilization need to run and progress today? 50% of current human biomass, 30%, 20%?
How many people does it need after every dumb job (where no inventiveness or adaptation to radically new situations is required) has been automated with an android with AGD (Artificial General Dumbness as opposed to the natural sort people in those positions possess)? 5%?
Basically you’d “need” only the productive, useful people; scientists, researchers, engineers, and artists. The idle & useless rich would probably stick around while increasing their carbon footprint to galactic proportions because everyone could own a private jet, a terayacht (there’s a gigayacht class in existence already) and there’d be plenty of oil for everyone for millenia. With only a few % left, with practically unlimited resources for everyone and all services automated, there’d be no one in need of help, and the (only?) useful function of the idle rich, charity, would be eliminated.
August 21st, 2008 at 7:15 am
What-If: Malware. I don’t trust my own computer and I’m a security expert. AI might provide better solutions, but the criminals will be using AI too.
August 23rd, 2008 at 9:59 pm
If human level AI is arrived at by too rigidly copying human brains on a different substrate then the AI may face severe limitations.
1) If the AI mind is like the human mind then its knowledge may be as sloppily encoded and as individual specific as human knowledge. In such case it will no more be able to replicate its state into others of its kind efficiently than we can.
2) It is not at all obvious that it can run at any arbitrary overclocked rate. There me aspects of its operation that cannot be freely raised in speed.
3) There is no reason, as in (1) to assume that human like AIs can communicate any faster or better than we do if they are sufficiently human like in architecture.
4) If they are that human like they will likely experience boredom of some form and also be roughly as fallible and slow witted as we are.
As these AIs will at least in the beginning likely be very expensive to build and maintain AI this limited will not be widely deployed and is not terribly likely to take off in development.
This is why I am underwhelmed by the suggestion we should create an AI that is architected on the human brain with the goal of human level intelligence. I hope we can do considerably better than that.
August 23rd, 2008 at 10:01 pm
Jordan,
Do you really believe that humans with our rather limited intelligence and abilities are likely to exhaust the entire search space of science and technology? If so, why do you believe this?
August 25th, 2008 at 5:57 pm
I don’t think we’ve come anywhere near exhausting this technological “search space.”
Also I believe the human brain has made use of boredom and forgetfulness as the brain is apt to make use of what it’s given. That’s the nature of the brain.
They can lead to creativity and inspiration, but as is likely with most aspects of human intelligence, they represent local maximums. What a self-improving AI could evolve itself into would probably bear little or no resemblance to the human mind.
The funny thing is that the best application of human-level AI is to put it inside robots and have them drive us around and make our food and other various goods. Essentialy we want slaves but in order to have them we have to develop the technology of their intelligence to the point where they might actually resent it.
We have black people already. And I’d rate a human-level intelligence robot at somewhere in the range of 4/5 to 1/1 of a human.
August 26th, 2008 at 9:12 am
Ahh — I get what is being said here, but it is really something of a misnomer. I mean, sure — such entities //might// resent us; but this is pre-assuming that a human-level AI would be possessed of human-like characteristics (namely; an organically derived personality to say the least — there are others that are equally applicable).
Furthermore, human-level AGI is not necessary to accomplish the supermajority of industrial/service tasks. Merely a lower-level AGI with sophisticated ANI ’subprocessors’ as it were. This would be equivalent to developing an artificial horse with an extremely sophisticated set of instincts which are highly suited to a specific task (such as knowing where to go and how to avoid hitting things; or how to assemble a given object.) This analogy/metaphor is incomplete, //but for this example alone// it is effective.
* AGI = artificial general intelligence
** ANI = artificial narrow intelligence
September 1st, 2008 at 1:04 pm
“And I’d rate a human-level intelligence robot at somewhere in the range of 4/5 to 1/1 of a human.”
How big is the difference between 60 and 200 or whatever the IQ range is? If an AI has IQ level equal to -1 or +1 human average level it’s not going to perform exactly the same. It’s worlds apart.
September 3rd, 2008 at 3:01 pm
Gah… moderation oblivion strikes again!