Do we understand ethics well enough to build “Friendly artificial intelligence”?
johncarlosbaez.wordpress.com
johncarlosbaez.wordpress.com
I'm surprised that you are so pessimistic about your research that you think ethics won't even be relevant in year 2205. Holy cow you must think AI is hard.
This is a very good point. It's always good to be reminded that we're already living in the future.
That said, I feel like aothman is discussing real artificial intelligence, that is, an entity capable of making a conscious decision that it wants to, in this case, fire the missiles. If I had to guess, if predator drones gain the ability to "decide" for themselves whether or not to fire their missiles, it will be built on a system of complex rules, and not because they're "intelligent". Potayto, Potahto? Maybe. I'm not an AI researcher and I don't even come close to understanding human intelligence, but I feel like even if it is just a complex system of rules, it's at a much deeper level than we'll be able to simulate soon.
This is not a new phenomenon; the first use of autonomous killer robots was in 1943, in the form of acoustically guided torpedoes.
To be clear about that "probably 200", are you saying that you believe we'll need trillions of times the processing power of the human brain in order to crack how it works, or that you believe that we've nearly reached the end of increases in processing power, for at least the next 200 years?
But self improving AI is not remotely predictable based on our current progress, and really, it's not even the same field as what we call AI today: extrapolating our current progress to predict where we'll be in 50 years is like asking a bombmaker from 1935 to look at a log log plot of historical explosive power in bombs to try to predict what the maximum yield from a bomb in 1950 would be. It doesn't matter how slow the mainstream research is if someone finds a chain reaction to exploit, and it's impossible to predict when someone will successfully exploit that chain reaction.
IMO there's very good reason to believe that we're already deep into the "yellow zone" of danger here, where we have more than enough computational power to set off a self-improving chain reaction, though we don't actually know how to write that software. What we really have to worry about is that as time goes by, we creep closer to the "red zone", where we don't even need to know how to write the software because any idiot with an expensive computer can brute force the search through program space (more realistically, they would rely on evolutionary or other types of relatively unguided methods). That's exceptionally dangerous because the vast majority of self improving AIs will be hostile, and we want to make sure that the first to emerge is benevolent.
So yes, there's a lot of uncertainty here, but I think it's a mistake to say that we don't need to worry about it until it's here. By the time it's inevitable and the mainstream has started to even accept it as possible, it's probably going to be impossible to ensure that (for instance) some irresponsible government won't be the first to achieve it merely by throwing a lot of funding at the problem and doing it unsafely.
This is a fascinating and broad-ranging criticism of AI, and it's interesting to me because the author is clearly considering 'what happens if we are successful?'.
Definitely worth a read.
Unfortunately, I don't think that's really going to happen.
When people refuse to fund research in a promising area, others wind up taking up the slack.
Witness America's prohibition on fetal stem-cell research. The US lost its lead in that field when researchers from other countries continued the work.
The same will be much more true of AI research, as a working strong AI would probably be seen as a goose that lays golden eggs.
Of course, it's also a Pandora's Box. So we should be careful. But I don't think scorning researchers in the field is going to be very effective.
For example, should we scorn genetics researchers who do not have a credible claim that their organisms will remain harmless after a billion generations of evolution and recombination? That's more or less what many of the anti-GMO arguments boil down to, that we ought to require genetically-modified organisms to be provably safe, both as they exist now, and in all possible future ways they could evolve and interact with other organisms. (And since that bar is very hard to reach, therefore, their arguments go, we should be careful about funding such research to begin with, and definitely shouldn't let any of its results out into the wild, e.g. into crops.)
If anything, the argument there is stronger, because evolving biological organisms that can pose a threat to humans actually exist, whereas evolving machines that can pose a threat to humans are sci-fi, and likely to remain so for a very long time. Why regulate the latter one more stringently?
Not at all true. The space of possible biological organisms is searched in a highly nonuniform manner by evolution, and the human search strategy is fundamentally different. It's overwhelmingly likely that there competitive human-constructible organisms which could never be produced by evolution in the past 4 billion years.
There is no reason to believe that golden rice will evolve in any significantly different manner than ordinary rice. In contrast, AI will evolve via a mechanism which is unprecedented.
What if we create bacteria that can digest anything, survive in environmental extrema, sporulate, and kill off competing strains? The grey goo scenario comes to mind.
Chemical and biological state space is infinite. There is much room for good, but also for bad. Misuse of biology is much more dangerous in the short term than AI.
Eliezer (the interviewee) has an HN account so he can comment for himself.
Depends what you mean. There certainly exist biological organisms that can pose a threat to individual humans; but AI can pose a threat to humanity itself, and is thus very very dangerous.
GMO happens slower and is somewhat more manageable.
Basically, I'm thinking something along the lines of Hanlon's Razor: "Never attribute to malice that which is adequately explained by stupidity."
We'll see serious damage caused by "weak" AIs long before we have a "strong" AI capable of causing similar damage. For example, 2010's "Flash Crash" seems to have high frequency trading at its core.
My hope is that through the growing pains we experience from "weak" AI systems doing something stupid, we'll be better prepared for a "strong" AI system that may try to do something malicious.
Absolute power corrupts absolutely. The goal is to make AI more powerful than humans is it not? We're not going to be able to control it, no way no how.
Real world resources are finite, and real world processes with real world materials take a certain finite amount of time. The singularity therefore ignores the realities of physics. It would even be possible to add artificial constraints on the total resource use and rate of resource use.
So I always take issues with these kinds of esoteric debates about how to engineer ethics into an intelligence that can learn and become conscious.
Haven't any of these yahoos ever had kids or owned a pet dog?
You don't "engineer ethics" into your son or daughter. You teach them through examples of good behavior, punish them when they misbehave, and reward them when they succeed. Over the course of a few years, given a good environment, the end result is a new young intelligence that knows how to behave well and get along with others. That intelligence often goes on to bootstrap itself up into adulthood and eventually goes on to create later iterations of itself. If it was raised well, then the new ones tend to get raised well too. We call them "grandkids".
So lets assume in 10-20 years something descended from IBM's Blue Brain (simulating cat cortexes) leads to something that is analogous in intellectual range from a dog to an elephant.
Most people will agree that dogs and elephants are pretty damn smart. Dogs are able to perceive human emotional states, understand some language, do work for people, and fit nicely into our social structure. Elephants aren't that close with people, but are highly intelligent, have active internal emotional states, and even grieve for their dead. In some societies, people and elephants have worked together for thousands of years.
In both these cases, we have a long history of working with other intelligences of varying scales for thousands of years. In general, if you don't mistreat them they turn out to be socialized pretty well. Its only when you mistreat them that they learn to fear and hate you. The same is true for people.
So as @aothman said in another comment in this thread, AI researchers are just trying to get their projects to not fall over. There's no thought of "engineering ethics". This problem is going to be solved one little bit at a time. Artificial neural architectures are going to more and more sophisticated over time. But there is a key difference between the underlying architecture and how you go about training these new minds.
If you raise them well, then most of these angels-on-a-pin discussions are just that, meaningless.
Your kids can be punished and rewarded in ways that matter. It's hard to imagine, though, someone with the Ring of Gyges staying moral for very long. And with strong AI, that being could easily be equally free from repercussions.
If AI Blue can't be socially ostracized or punished, your "it's just a kid" story quickly breaks down. And it turns out "follow these examples" runs into problems when you ask "why?" about 3 to 5 times.
Humans are actually similar -- we have some sort of an innate ethical sense, though human ethics is a lot more complicated than dog ethics. We're smart enough to realise that our own short-term best interests may be served by acting in a non-ethical manner, and we've evolved all sorts of defence mechanisms to cope with this fact, including the moral outrage and desire for revenge which we feel when we see someone behaving non-ethically.
So in conclusion, while you can't necessarily engineer ethics into a mind, you can hard-wire in the structures necessary to care about ethics. Humans and dogs both have some sort of ethical sense wired in.
Nor do you have to wire your children to empathize with others: asking "how would you feel if Timmy took YOUR toy?" is enough.
Teaching a being with a moral sense is very different from creating one.
I should say something clever to the fatalist OP to avoid downmod... something about hardware requirements for agi likely being high (imho) and that we have a few decades to work on the really hard problem and we haven't yet worked on hard problems for a few decades since maturing as an information processing species.