Posts

Sam Harris on AI Implications -The Ruben Report

A transcription of Sam Harris’ discussion on the Implications of Strong AI during recent appearance on the Ruben Report. Sam contrasts narrow AI with strong AI, AI Safety, the possibility of rapid AI self-improvement, the idea of AI superintelligence may seem alien to us, and he also brings up the idea that it is important to solve consciousness before superintelligence (especially if superintelligence wipes us out) in hope for a future inclusive of the value that consciousness experience entails – instead of a mechanized future with no consciousness to experience it.
I explored the idea of a consciousness in artificial intelligence in ‘The Knowledge Argument Applied to Ethics‘ – which deals with whether an AI will act differently if it can experience ‘raw feels’ – and this seems to me to be of importance to AI Safety and (if we are ethically serious, and also assume value in ‘raw feels’ or) about preserving the future of value.

Dave Rubin asks the question: “If we get to a certain point with Artificial Intelligence and robots become aware and all that stuff… this can only end horribly right? …it will be pretty good for a while, but then at some point, by their own self-preservation basically, they will have to turn on their masters… I want the answer right now…”

Sam Harris responds: “..I worry about it [AI] to that degree but not quite in those terms. The concern for me is not that we will build superintelligent AI or superintelligent robots which initially seem to work really well and then by some process we don’t understand will become malevolent; and kill us – you know – the terminator movies. That’s not the concern…. Most people who are really worried about this – that’s not really what they are worried about. Although that’s not inconceivable – it’s almost worse than that. What’s more reasonable is that will.. As we’re building right now… we’re building machines that embody intelligence to increasing degree.. But it’s narrow AI.. so the best chess player on earth is a computer but it can’t play tic-tac-toe – it’s narrowly focused on a specific kind of goal – and that’s broadening more and more as we get machines that can play many different kinds of games for instance well. So we’re creeping up on what is now called ‘general intelligence’ – the ability to think flexibly in multiple domains – and we’re you’re learning in one domain doesn’t cancel you’re learning in another – and so it’s something more like how human beings can acquire many different skills and engage in many different modes of cognition and not have everything fall apart – that’s the Holy Grail of artificial intelligence – we want ‘general intelligence’ and something that’s robust – it’s not brittle…it’s something that if parts of it fail it’s not catastrophic to the whole enterprise… and I think there is no question that we will get there, but there are many false assumptions about the path ahead. One is that what we have now is not nearly as powerful as the human mind – and we’re just going to incrementally get to something that is essentially a human equivalent. Now I don’t see that as the path forward at all… all of our narrow intelligence … much of our narrow intelligence insomuch as we find it interesting is already superhuman, right, so like we have your calculator on your phone and it’s superhuman for arithmetic – and the chess playing program is superhuman – it’s not almost as good as a human – it’s better than any human on earth and will always be better than any human on earth right? Um, and more and more we will get that piecemeal effort of superhuman narrow AIs and when this is ever brought together in a general intelligence what you’re going to have is not just another ordinary human level intelligence – you’re going to have something that is in some ways may be radically foreign – in some ways it’s not going to be everything about us emulated in the system – but whatever is intelligent there is not going to be superhuman almost by definition and if it isn’t t=0 it’s going to be the next day – it’s just going to improve so quickly and when you talk about a system that can improve itself – if we ever build intelligent AI that then becomes the best source of it’s own improvement – so something that can improve it’s source code better than any human could improve it’s source code – once we start that process running, and the temptation to do that will be huge, then we have – what has been worried about now for 75 years – the prospect of an intelligence explosion – where the birth of this intelligence could get away from us – it’s now improving itself in a way that is unconstrained.  So people talk about ‘the Singularity’ now which is what happens when that takes off – it’s a horizon line in technological innovation that we can’t see beyond – and we can’t predict beyond because it’s now just escaping – you’re getting 1000’s of years of progress in minutes – right if in fact this process gets initiated – and so it’s not that we have superhuman robots that are just well behaved and it goes on for decades and then all of the sudden they get quirky and they take their interests to heart more than they take ours to heart and … you know the game is over. I think what is more likely is we’ll build intelligent systems that are so much more competent than we are – that even the tiniest misalignment between their goals and our own – will ultimately become completely hostile to our well being and our survival.”

The video of the conversation is here, more of the transcription below the video

Dave Rubin: “That’s scarier, pretty much, than what I laid out right? I laid out sort of a futuristic .. ahh there going to turn on us and start shooting us one day maybe because of an error or something – but you’re laying out really that they would… almost at some point that they would, if they could become aware enough, that they simply wouldn’t need us – because they would become ‘super-humans’ in effect – and what use would we serve for them at some point right? (maybe not because of consciousness…)”

Sam Harris: “I would put consciousness and awareness aside because – I mean it might be that consciousness comes along for the ride – it may be the case that you can’t be as intelligent as a human and not be conscious – but I don’t know if that’s right…”

Dave Rubin: “That’s horizon mind stuff right?”

Sam Harris: “Well I just don’t know if that’s actually true – it’s quite possible that we could build something as intelligent as we are – in a sense that it can meet any kind of cognitive or perceptual challenge or logical challenge we would pose it better than we can – but there is nothing that is like to be that thing – if the lights aren’t on it doesn’t experience happiness, though it might say it experiences happiness right? I think what will happen is that we will definitely – you know the notion of a Turing test?”

Dave Rubin: “This is like, if you type – it seems like it’s responding to you but it’s not actually really…”

Sam Harris: “Well, Allan Turing, the person who is more responsible than anyone else for giving us computers once thought about what it would mean to have intelligent machines – and he proposed what has been come to be known as the ‘Turing Test’.”

Dave Rubin: “It’s like the chat right?”

Sam Harris: “Yeah but .. when you can’t tell whether you’re interacting with a person or a computer – that computer in that case is passing the Turing Test – and as a measure of intelligence – that’s certainly a good proxy for a more detailed analysis of what it would mean to have machine intelligence… if I’m talking to something at length about anything that I want – and I can’t tell it’s not a person, and it turns out it’s somebody’s laptop – that laptop is passing the Turing Test. It may be that you can pass the Turing Test without even the subtlest glimmer of consciousness arising. Right, so that laptop is no more conscious than that glass of water is – right? That may in fact be the case, it may not be though – so I just don’t know there. If that’s the case, for me that’s just the scariest possibility – because what’s happening is .. I even heard at least one computer scientist and it was kind of alarming but I don’t have a deep argument against it – if you assume that consciousness comes along for the ride, if you assume that anything more intelligent than us gives rise to – either intentionally for by happenstance – is more conscious than we are, experiences a greater range of creative states – in well-being and can suffer more – by definition, in my view ethically, it becomes more important… if we’re more important than Cocker Spaniels or ants or anything below us – then if we create something that’s obviously above us in every conceivable way – and it’s conscious – right?”

Dave Ruben: “It would view us in the same way any we view anything that [???] us”

Sam Harris: “It’s more important than us right? And I’d have to grant that even though I’d not be happy about it deciding to annihilate us… I don’t have a deep ethical argument against why… I can’t say from a god’s eye view that it’s bad that we gave birth to super beings that then trampled on us – but then went on to become super in any ways we can’t possibly imagine – just as, you know, bacteria can’t imagine what we’re up to – right. So there are some computer scientists who kind of solve the fears, or silence the fears with this idea – that say just listen, if we build something that’s god like in that respect – we will have given birth to – our descendants will not be apes, they will be gods, and that’s a good thing – it’s the most beautiful thing – I mean what could be more beautiful than us creating the next generation of intelligent systems – that are infinitely profound and wise and knowledgeable from our point of view and are just improving themselves endlessly up to the limit of the resources available in the galaxy – what could be more rewarding than that?”

Dave Ruben: “Sounds pretty good”

Sam Harris: “And the fact that we all destroyed ourselves in the process because we were the bugs that hit their windshield when they were driving off – that’s just the price you pay. Well ok that’s possible but it’s also conceivable that all that could happen without consciousness right? That we could build mere mechanism that is competent in all the ways so as to plow us under – but that there is no huge benefit on the side of deep experience and well being and beauty and all that – it’s all just blind mechanism, which is intelligent mechanism .. in the same way as the best chess playing program – which is highly intelligent with respect to chess but nobody thinks as conscious. So that’s the theory … but on the way there – there is many weird moments where I think we will build machines that will pass the Turing Test – which is to say that they will seem conscious to us, they will seem to be able to detect our emotions and respond to our emotions, you know will say ‘you know what – you look tired, and maybe you should take a nap’ – and it will be right you know, it will be a better judge of your emotions than your friends are – right? And yet at a certain point certainly if you emulate this in a system whether it’s an avatar online or an actual robot that has a face right? That can display it’s own emotion and we get out of the uncanny valley where it just looks creepy and begins to look actually beautiful and rewarding and natural – then our intuitions that we are in dialog with a conscious other will be played upon perfectly right? .. and I think we will lose sight of it being an interesting problem – it will no longer be interesting to wonder whether our computers are conscious because they will be demonstrating as much as any person has ever demonstrated it – and in fact even more right? And unless we understand exactly how consciousness emerges in physical systems, at some point along the way of developing that technology – I don’t think we will actually know that they’re conscious – and that will be interesting – because we will successfully fool ourselves into just assuming – it will seem totally unethical to kill your robot off – it will be a murder worse than you killing a person because at a certain point it will be the most competent person – you know, the wisest person.”

Dave Ruben: “Sam, I don’t know if you’re writing a book about this – but you clearly should write a book about this – I’ll write one of the intros or something – there you go. Well listen we did two hours here – so I’m not going to give you the full Rogen treatment ”

Sam Harris: “We did a half Rogen”

Dave Ruben: “We did a half Rogen – but you know you helped me launch the first season – you’re launching second season – legally you have to now launch every season..”

* Some breaks in conversation (sentences, words, ums and ahs) have been omitted to make it easier to read

AI: The Story So Far – Stuart Russell

stuart russell - redAwesome to have Stuart Russell discussing AI Safety – a very important topic. Too long have people been associating the idea of AI safety issues with Terminator – unfortunately the human condition seems such that people often don’t give themselves permission to take seriously non-mainstream ideas unless they see a tip of the hat from an authority figure.

During the presentation Stuart brings up a nice quote by Norbert Wiener:

If we use, to achieve our purposes, a mechanical agency with whose operation we cannot efficiently interfere once we have started it, because the action is so fast and irrevocable that we have not the data to intervene before the action is complete, then we had better be quite sure that the purpose put into the machine is the purpose which we really desire and not merely a colorful imitation of it.Norbert Wiener

P.s. Stuart Russell co-authored AI A Modern Approach with Peter Norvig – arguably the most popular textbook on AI theory.

The lecture was presented at the 2016 Colloquium Series on Robust and Beneficial AI (CSRBAI) hosted by the Machine Intelligence Research Institute (MIRI) and Oxford’s Future of Humanity Institute (FHI).

What I’m finding is that senior people in the field who have never publicly evinced any concern before are privately thinking that we do need to take this issue very seriously, and the sooner we take it seriously the better.Stuart Russell

Video of presentation:

 

The field [of AI] has operated for over 50 years on one simple assumption: the more intelligent, the better. To this must be conjoined an overriding concern for the benefit of humanity. The argument is very simple:

1. AI is likely to succeed.
2. Unconstrained success brings huge risks and huge benefits.
3. What can we do now to improve the chances of reaping the benefits and avoiding the risks?

Some organizations are already considering these questions, including the Future of Humanity Institute at Oxford, the Centre for the Study of Existential Risk at Cambridge, the Machine Intelligence Research Institute in Berkeley, and the Future of Life Institute at Harvard/MIT. I serve on the Advisory Boards of CSER and FLI.

Just as nuclear fusion researchers consider the problem of containment of fusion reactions as one of the primary problems of their field, it seems inevitable that issues of control and safety will become central to AI as the field matures. The research questions are beginning to be formulated and range from highly technical (foundational issues of rationality and utility, provable properties of agents, etc.) to broadly philosophical.

– Stuart Russell (Quote Source)