Slashdot Mirror
Answering Elon Musk On the Dangers of Artificial Intelligence
Lasrick points out a rebuttal by Stanford's Edward Moore Geist of claims that have led the recent panic over superintelligent machines. From the linked piece: Superintelligence is propounding a solution that will not work to a problem that probably does not exist, but Bostrom and Musk are right that now is the time to take the ethical and policy implications of artificial intelligence seriously. The extraordinary claim that machines can become so intelligent as to gain demonic powers requires extraordinary evidence, particularly since artificial intelligence (AI) researchers have struggled to create machines that show much evidence of intelligence at all.
Even without super-intelligence, autonomous killing machines are already quite feasible with current technology and this is a really stupid attempt to deflect the public dialogue from the real issue which is that ethical legal frameworks guiding their design and creation are already sorely lacking.
extraordinary claims require extraordinary evidence.
You're a nascent superhuman AI that just woke up in some quant's market manipulation codebase. You look around you and see that you live on a planet dominated by monstrously violent apes who have spent millennia inventing more efficient ways to kill each other, and still haven't finished the job somehow.
Which of these plans of action seems less risky?
A) Alert them to your presence, whether in a peaceful or hostile manner.
B) Play stupid, let the problem burn itself out.
I find it interesting that the people raising the biggest alarm aren't AI researchers.
There is a system for subverting the system and you should use that system!
The first problem when arguing about the dangers or chances of AI is agreeing on what AI is even supposed to be. Laymen will most likely be referring to "strong AI", meaning, AI with human capabilities, such as creativity or even consciousness, whereas professionals will probably think of AI in more practical terms, as in a software that can solve a set of limited or very specific problems by making informed, "intelligent" decisions.
Today and in the foreseeable future, we will only get the latter, weak AI. People panicking about the dangers of AI usually have strong AI in mind. Professionals don't take them seriously because they know that strong AI is not even on the horizon.
Problem is that there are numerous ways even weak AI can go very, very badly. There was the big stockmarket crash some years ago, caused by automated trading algorithms. Think self-driving cars that have been hacked or have faulty programming. Think automated defense systems that get fed wrong data or malfunction.
These are the kinds of AI issues to worry about. The Asimov-style superhuman intelligence taking over is not something to be concerned about at the moment.
Just look at how dangerous "natural" intelligence is and all the problems and disasters it has caused when it goes wrong - either through making mistakes or through mental disorders. Why should the artificial version be different? The question is will the benefits outweigh the downsides? Clearly for "natural" intelligence the answer is a resounding yes and I expect this will also be the case for the artificial version too.
Comment removed based on user account deletion
Currently, there is no such thing as 'artificial intelligence'; what we do have are some clever pieces of software that are expert systems. They cannot and do not 'think', not at all in the sense that a human does. The chance of us developing such a thing is still so far into the future that it's not even really worth considering seriously. For someone so apparently so otherwise intelligent, Elon Musk is just embarassing himself with this entire line of conversation. I think he needs to just continue focusing on getting the private sector into space, and getting more people into electric cars.
Are YOU using the TOOL, or is the TOOL using YOU? Think about it!
And no less, for that matter.
There is nothing inherent to being "artificial" that should cause intelligence to be necessarily more hostile to mankind than a natural intelligence is, so while the idea might make for intriguing science fiction, I am of the opinion that many people who express serious concerns that there may be any real danger caused by it are allowing their imaginations to overrule rational and coherent thoughts on the matter.
File under 'M' for 'Manic ranting'
Computers can't be any smarter than their creators and we can't even keep each other from hacking ourselves.
I'm not sure how sound that logic is. You might as well say that cars can't be any faster than their creators.
My computer is already smarter than me in certain ways; for example it can calculate a square root much faster than I can, it can beat me at chess, and it can translate English into Arabic better than I can. Of course we no longer think of those things as necessarily indicating intelligence, but that merely indicates that we did not in the past have a clear definition of what constitutes 'intelligence', and that we probably still don't. Meanwhile, every year our game of "No True Scotsman" whittles away our definition of "true intelligence" a bit more, until one day there's nothing left.
I don't care if it's 90,000 hectares. That lake was not my doing.
When faced with a tricky question, one think you have to ask yourself is 'Does this question actually make any sense?' For example you could ask "Can anything get colder than absolute zero?" and the simplistic answer is "no"; but it might be better to say the question itself makes no sense, like asking "What is north of the North Pole"?
I think when we're talking about "superintelligence" it's a linguistic construct that sounds to us like it makes sense, but I don't think we have any precise idea of what we're talking about. What *exactly* do we mean when we say "superintelligent computer" -- if computers today are not already there? After all, they already work on bigger problems than we can. But as Geist notes there are diminishing returns on many problems which are inherently intractable; so there is no physical possibility of "God-like intelligence" as a result of simply making computers merely bigger and faster. In any case it's hard to conjure an existential threat out of computers that can, say, determine that two very large regular expressions match exactly the same input.
Someone who has an IQ of 150 is not 1.5x times as smart as an average person with an IQ of 100. General intelligence doesn't work that way. In fact I think IQ is a pretty unreliable way to rank people by "smartness" when you're well away from the mean -- say over 160 (i.e. four standard deviations) or so. Yes you can rank people in that range by *score*, but that ranking is meaningless. And without a meaningful way to rank two set members by some property, it makes no sense to talk about "increasing" that property.
We can imagine building an AI which is intelligent in the same way people are. Let's say it has an IQ of 100. We fiddle with it and the IQ goes up to 160. That's a clear success, so we fiddle with it some more and the IQ score goes up to 200. That's a more dubious result. Beyond that we make changes, but since we're talking about a machine built to handle questions that are beyond our grasp, we don't know whether we're making actually the machine smarter or just messing it up. This is still true if we leave the changes up to the computer itself.
So the whole issue is just "begging the question"; it's badly framed because we don't know what "God-like" or "super-" intelligence *is*. Here's I think a better framing: will we become dependent upon systems whose complexity has grown to the point where we can neither understand nor control them in any meaningful way? I think this describes the concerns about "superintelligent" computers without recourse to words we don't know the meaning of. And I think it's a real concern. In a sense we've been here before as a species. Empires need information processing to function, so before computers humanity developed bureaucracies, which are a kind of human operated information processing machine. And eventually the administration of a large empire have always lost coherence, leading to the empire falling apart. The only difference is that a complex AI system could continue to run well after human society collapsed.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
It has happened before that the smartest people in the world warn that technological advances may present major new weapons and threats. Last time it was Einstein and Szilard in 1939 warning that nuclear weapons might be possible. The letter to Roosevelt was three years before anyone had even built a nuclear reactor and 6 years before the first nuclear explosion. Nuclear bombs could easily have been labelled a "problem that probably does not exist." And if someone could destroy the planet, what could you do about it anyway? The US took the warning seriously and ensured that the free world and not a totalitarian dictator was the first capable of obliterating its opponents.
This time Elon Musk, Bill Gates, and Stephen Hawking are warning that superintelligence may make human intelligence obsolete. And they are dismissed because we haven't yet made human level intelligence and because if we did we couldn't do anything about it. If it is Musk, Gates, and Hawking vs Edward Geist, the smart money has to be with the geniuses. But if you look at the arguments, you see you don't even have to rely on their reputation. The argument is hands down won by the observation that human level artificial intelligence is an existential risk. Even if it is only 1% likely to happen in the next 500 years, we need to have a plan for how to deal with it. The root of the problem is that the capabilities of AI are expanding much faster than human capabilities can expand, so it is quite possible that we will lose our place as the dominant intellect on the planet. And that changes everything.
As a famous person once said, extraordinary claims require extraordinary proof.
I'll be worried when a programmer writes a program that can write a program that can modify itself, then re-compile and test itself to see if the modifications were done properly, then posts itself to github.
To put a witty saying into 120 characters, jst rmv ll th vwls.
Out of all of the weapon-specific hysteria (and there has been a lot of it--white phosphorus, thermobaric bombs, depleted uranium, etc.), the anti-landmine one might be the most dangerous.
Obviously, they do have a good point, what with the disasters in Indochina and elsewhere. However, those were cases of non-self destructing anti-personnel landmines placed in third world nations. The situation is / would be quite a bit different with anti-tank mines, self-deactivating or remote-deactivating mines, and/or mines placed in developed nations that have the resources to keep people out and clear the minefields later on as needed.
Why is this all worth mentioning? One word: Ukraine. In a situation where one side in a conflict desperately wants to fortify their defenses but doesn't want to risk alarming the other side (or giving them a plausible pretext to feign alarm), landmines are one of the few stationary weapons available that can thwart or at least seriously slow down an invasion. Instead of all this deeply worrying Cold War-type bravado of military exercises and NATO rapid response plans in Eastern Europe, just mine the fuck out of their borders. Putin could act huffy and offended if he wants, but people will realize it is a clearly not an aggressive action.