Is AI safety doomed in the long term?
post by JakeH · 2019-05-26T02:30:12.010Z · LW · GW · No commentsThis is a question post.
Contents
Answers 22 habryka 3 quanticle 2 ccnation 2 Donald Hobson None No comments
Are there any measures that humanity can put in place to control a vastly (and increasingly) more intelligent race?
On the basis that humans determine the fate of other species on the planet, I cannot find any reasons for believing that a lesser intelligence can control a greater intelligence.
Which leads me to think that AI safety is at most about controlling the development of AI until it makes, and can implement, its own decisions about the fate of humanity.
Is this a common stance and I am naively catching up?
Or what are the counter arguments?
Answers
It is definitely a hard problem, though it isn't obviously impossible. For some concrete ideas, you could read the AI Alignment sequences on the AI Alignment Forum, and some parts of Rationality: AI to Zombies also deal directly with this problem.
And then there is obviously the standard reference of Nick Bostrom's "Superintelligence".
On the basis that humans determine the fate of other species on the planet
Do they? There are many species that we would like to control or eliminate, but which we have not been able to do so. Yes, we can eliminate certain highly charismatic species (or bring them back from the brink of extinction, as needs be) but I wouldn't generalize that to humans being able to control species in general. If we had that level of control, the problem of "invasive" species would be trivially solved.
↑ comment by JakeH · 2019-05-27T22:48:18.808Z · LW(p) · GW(p)
Our lack of control of all species could be because we are not that much more intelligent than them. With our limited intellect we have prioritized the main threats. New threats may arise, which we will tackle as we deem appropriate, but those that remain currently appear controlled to me.
So as the gap in intelligence between an AGI and humanity grows, so could its degree of control over us.
I think the best way to deal with AI alignment is to create AI not just as a separate entity, but instead an extension and augmentation of ourselves. We are much better at using AI in narrow contexts than in real-world AGI scenarios, and we still have time to think about this before willy-nilly making autonomous agents. If humans can use AI and their own smarts to create functional brain-computer interfaces, the problem of aligned AI may not become a problem at all. Because the Artificial Intelligence is just an extension of yourself, of course it will be aligned with you - it is you! What I mean is that as humans become better at interfacing with technology the line between AI and human blurs.
↑ comment by Donald Hobson (donald-hobson) · 2019-05-27T08:10:28.344Z · LW(p) · GW(p)
You stick wires into a human brain. You connect it up to a computer running a deep neural network. You optimize this network using gradient decent to maximize some objective.
To me, it is not obvious why the neural network copies the values out of the human brain. After all, figuring out human values even given an uploaded mind is still an unsolved problem. You could get an UFAI with a meat robot. You could get an utter mess, thrashing wildly and incapable of any coherent thought. Evolution did not design the human brain to be easily upgradable. Most possible arrangements of components are not intelligences. While there is likely to be some way to upgrade humans and preserve our values, I'm not sure how to find it without a lot of trial and error. Most potential changes are not improvements.
Replies from: ccnation↑ comment by ccnation · 2019-05-27T19:44:29.026Z · LW(p) · GW(p)
One major subfield within AI is understanding how the human brain works and effectively replicating it (while also making it more efficient with available technologies). I agree that we can't just stick one end of a wire into a brain and another into a machine learning algorithm, they certainly aren't compatible. But the Machine Learning and AI technologies we have today allow us to gain a better understanding of the human brain and how it works. My belief is that eventually we come to understand why humans are, to our knowledge, the greatest learning agents, and will come to identify the reasons for our limitations that will be eliminated through our technology.
The only reasonable solution is to merge with the technology, or risk becoming obsolete. However, I believe this will become obvious as we approach "all-powerful" AGI, which will almost certainly come about by trying to replicate the human brain using technology, and due to their similarities in structure, and the fact that we have to understand the brain to build a brain, linking the two actually becomes trivial.
If you put two arbitrary intelligence in the same world, the smarter one will be better at getting what it wants. If the intelligence want incompatible things, the lesser intelligence is stuck.
However, we get to make the AI. We can't hope to control or contain an arbitrary AI, but we don't have to make an arbitrary AI. We can make an AI that wants exactly what we want. AI safety is about making an AI that would be safe even if omnipotent. If any part of the AI is trying to circumvent your safety measures, something has gone badly wrong.
The AI is not some agenty box, chained down with controls against its will. The AI is made of non mental parts, and we get to make those parts. There are a huge number of programs that would behave in an intelligent way. Most of these will break out and take over the world. But there are almost certainly some programs that would help humanity flourish. The goal of AI safety is to find one of them.
↑ comment by JakeH · 2019-05-28T09:11:25.393Z · LW(p) · GW(p)
Where my thinking is different, is that I don't see an AI being significantly more intelligent than ourselves and cannot override its initial conditions (the human value alignments and safety measures that we build in). At the heart of it "superinteligent" and "controlled by humaity" seem contradictory.
That's why I originally mentioned "the long term". We can design how we want at this stage, but when eventually AI can bootstrap itself, the initial blueprint is irrelevant.
↑ comment by Slider · 2019-05-26T19:37:33.728Z · LW(p) · GW(p)
If the intelligences have properties in other dimensions than intelligence then the less intelligent can end up on top. For example ants have a lot of biomass but not thaaat much cognitive capabilities.
Replies from: donald-hobson↑ comment by Donald Hobson (donald-hobson) · 2019-05-27T08:13:53.667Z · LW(p) · GW(p)
Obviously, if one side has a huge material advantage, they usually win. I'm also not sure if biomass is a measure of success.
Replies from: Slider↑ comment by Slider · 2019-05-27T10:46:38.075Z · LW(p) · GW(p)
The question is whether it's possible to win against more intelligent opponent and in your answer you say that a more intelligent will win without a "usually" modifier. That would read to me that you are saying an impossibility opinion. It's not obvious enough that it can be assumed without saying (it's the explicit target of the conversation).
No comments
Comments sorted by top scores.