Sam Harris update on dangers of AI | Lex Fridman Podcast Clips

it would be interesting to hear how is your concerned concern evolved with the coming out of Chad GPT and these new large language models that are fine-tuned with reinforcement learning and seemingly to be able to do some incredible human-like things there's two questions one how is your concern in terms of AGI and super intelligence evolved and how impressed are you with Chad gbt as a as a student of the human mind in mind in general well my concern about AGI is unchanged and so I I did a I've spoken about it a bunch on my podcast but you know I did a TED Talk in 2016 which was the the kind of summary of what that conference and and you know various conversations I had after that did to my my brain on this topic um basically that once super intelligence is achieved there's a take off it becomes exponentially smarter and in a matter of time there's just we're ants and they're Gods well yeah unless we find some way of to permanently tethering a a self a a super intelligent super intelligent self-improving AI to our value system and I you know I don't believe anyone has figured out how to do that or whether that's even possible in principle I know people like Stuart Russell who I just had on my podcast um are oh really have You released it I haven't released it yet yeah he's been on previous podcasts but we just recorded this week because you haven't done an AI podcast in a while so yeah it's great yeah he's a good person to talk about alignment with yeah so Stuart I mean Stewart has has been you know probably more than anyone my Guru on this topic I mean like just reading his book and and doing I think I've done two podcasts with them at this point I think it's called the control problem or something like that his is uh his book is human compatible human compatible yeah he talks about the control problem and yeah so I just think the idea that we can define a value function in advance that permanently tethers a a self-improving super intelligent AI to R values as we continue to discover them refine them extrapolate them in an open-ended way I think that's a tall order and there I think there are many more ways there must be many more ways of Designing super intelligence that is not aligned in that way and is not ever approximating our values in that way so I mean Stewart's idea to put it in a very simple way is that he thinks you don't want to specify the value function up front you don't you don't want to imagine you could ever write the code in such a way as to admit of no loophole you want to make the AI uncertain as to what human values are and perpetually uncertain and always trying to ameliorate that uncertainty by by hewing more and more closely to what our professed values are so like this it's always interested in uh saying oh no no that's not what we want that's not what we intend stop doing that right like no matter how smart it gets all it wants to do is more perfectly approximate human values I think there are a lot of problems with that you know at a high level I'm not a computer scientist so I'm sure there are many problems at a low level that I don't understand or like how to force the human into the loop always no matter what there's that and like what humans get a vote and just just what is you know uh what what do humans value and what is the difference between well we say we value and our revealed preferences which I mean if you just if you were a super intelligent AI they could look at Humanity now I think you could be forgiven for concluding that what we value is driving ourselves crazy with Twitter and living perpetually on the brink of nuclear war and you know just watching you know hot girls in yoga pants on Tick Tock again and again it's like when you're saying that is not this is this is all revealed preference and it's what is an AI to make of that right and what should it optimize like so and part of the this is also Stewart's observation that one of the Insidious things about like the YouTube algorithm is it's not that it just caters to our preferences it actually begins to change Us in ways so as to make us more predictable it's like it finds ways to make us a better reporter of our of our preferences uh and to trim our preferences down so that it can can uh further train to that signal so the main concern is that most of the people in the field seem not to be taking intelligence seriously like as as they design more and more intelligent machines and as they profess to want to design true AGI they're not again they're not spending the time that Stewart is spending trying to figure out how to do this safely above all they're just assuming that these these problems are going to solve themselves as we as we make that final stride into the end zone or they're saying very you know pollyanish things like you know an AI would never form a motive to harm human like why would it ever form a motive to to be malicious toward Humanity right unless we put that motive in there right and that's that's not the concern the concern is that in the presence of of vast disparities and competence and in certainly in a condition where these the machines are improving themselves they're improving their own code they could be developing goal instrumental goals that are antithetical to our well-being without any without any intent to harm us right it's it's analogous to what we do to every other species on Earth I mean you and I don't consciously form the intention to harm insects on a daily basis but there are many things we could intend to do that would in fact harm insects because you know you decide to repave your driveway or whatever whatever you're doing you're like you're not you're just not taking the the interests of insects into account because they're so far beneath you in terms of your cognitive Horizons and the so that the real challenge here is that if you believe that intelligence you know scales up on a Continuum to toward Heights that we can only dimly imagine and I think there's every reason to believe that there's no reason to believe that we're near the summit of intelligence um and you can Define you know Define maybe maybe there's maybe there's some forms of intelligence for which this is not true but for for many relevant forms you know like the top 100 things we care about cognitively I think there's every reason to believe that many of those things most of those things are a lot like chess or go where once the machines get better than we are they're going to stay better than we are although they're I don't know if you caught the recent thing with go where and this guy actually came out of Stewart's lab yeah yeah yeah uh one time a human beat a machine and yeah they found a hack for that but anyway in the ultimately it's there's going to be no looking back and then the question is what do we do in relate in relationship to these systems that are more competent than we are in every relevant respect because it will be a relationship it's not like the people the people who think we're just going to figure this all out you know without thinking about it in advance it's just gonna these the solutions are just gonna find themselves um seem not to be taking the prospect of really creating autonomous super intelligence seriously like like what does that mean it's every bit as independent and ungovernable ultimately as us having created I mean just imagine if we created a race of people that were 10 times smarter than all of us like how would we live with those people they're 10 times smarter than us right like they begin to talk about things we don't understand they begin to want things we don't understand they begin to view us as obstacles to them so they're solving those problems or gratifying those desires we become the chickens or the monkeys in their presence and I think that it's but for some amazing solution of the sort that Stuart is is imagining that we could somehow anchor their reward function permanently no matter how intelligent scales I think um it's it's really worth worrying about this I do I do buy the the you know the Sci-Fi uh notion that this is an existential risk if we don't do it well I worry that we don't notice it I'm I'm deeply impressed with Chad GPT and I'm worried that it will become super intelligent these language models have become super intelligent because they're basically trained in the collective intelligence of the human species and then it'll start controlling our Behavior if they're integrated into our algorithms the recommender systems and then we just won't notice that there's a super intelligent system that's controlling our Behavior well I think that's true even before far before super intelligence even before general intelligence I mean I think just the narrow intelligence of these algorithms and of what something like you know chat GPT can can do um I mean it's just far short of it developing its own goals and that is that are at Cross purposes with ours just the just the unintended consequences of of using it in the ways we're going to be incentivized to use it and and you know the money to be made from scaling this thing and what it does to to our information space and our sense of of just being able to get the ground truth of of Iran on any facts it's um yeah it's it's super scary and it was it's do you think it's a giant leap in terms of the development towards AGI Chad GPT or we still um is this just an impre impressive little toolbox so like what what do you think the singularity is coming or is it to you doesn't matter eventually I have no intuitions on that front apart from the fact that if we continue to make progress it will come right so it's just you just have to assume we continue to make progress there's only two assumptions you you have to assume substrate Independence so there's no reason why this can't be done in silico it's just it's just we can build arbitrarily intelligent machines there's nothing magical about having having this done in in the wet wear of our own brains I think that is true and I think that's a you know scientifically parsimonious to think that that's true um and then you just have to assume we're going to keep making progress it doesn't have to be any special rate of progress doesn't have to be Moore's Law it can just be we just keep going at a certain point we're going to be in relationship to Minds leaving Consciousness Consciousness aside I I don't I don't have any reason to believe that they'll necessarily be conscious by virtue of being super intelligent and that's its own interesting ethical question but uh leaving conscience aside they're going to be more they're going to be more competent than we are and then that's like you know the aliens have landed you know that's literally that's an encounter with again leaving aside the possibility that that something like Stewart's path is is is uh actually available to us uh but it is hard to picture if what we mean by intelligence All Things Considered and it's truly General um if that scales and you know begins to build upon itself how you maintain that perfect slavish devotion Until the End of Time the tethering those systems that tells you humans yeah I think my gut says that that tether is not there's a lot of ways to do it so it's not this increasingly impossible problem right so I have no you know you know as you know I'm not a computer scientist I have no intuitions about just algorithmically how you would approach that and what's somewhat spot my main intuition is uh maybe deeply flawed but the main nutrition is based on the fact that most of the learning is currently happening on human knowledge so even Chad GPT is just trained on human data right I I don't see where the takeoff happens where you completely go above human wisdom the current impressive aspect of child gpts that's using collective intelligence of all of us from what what I glean from again from people who know much more about this than I do I I think we have reason to be skeptical that these Tech techniques of you know deep learning are actually going to be sufficient to push us into AGI right so it's just not they're not they're not generalizing in the way they need to they're not certainly not learning like human children and so they're they're there's brittle and strange ways there there um it's not to say that the human path is the only path you know and you know and maybe there's we might learn better lessons by ignoring the way brains work but um we know that they don't generalize and use abstraction the way we do and so um although they have strange holes in their competence but the size of the holes is shrinking every time and that's so the intuition starts to slowly fall apart you know the intuition is like surely it can't be this simple to achieve super intelligence yeah but it's becoming simpler and simpler so I don't know I don't the progress is quite incredible I've been extremely impressed with Chad GPT and the new models and there's a lot of financial incentive to make progress on this regard so it's we're going to be living through some very interesting times

Channel: Lex Clips

Views: 260,494

Keywords: ai, ai clips, ai podcast, ai podcast clips, artificial intelligence, artificial intelligence podcast, computer science, consciousness, deep learning, einstein, elon musk, engineering, friedman, joe rogan, lex ai, lex clips, lex fridman, lex fridman podcast, lex friedman, lex mit, lex podcast, machine learning, math, math podcast, mathematics, mit ai, philosophy, physics, physics podcast, sam harris, science, tech, tech podcast, technology, turing

Id: nsFwyWZjh7A

Length: 15min 42sec (942 seconds)

Published: Thu Mar 16 2023