[Music] how will we know when AI is awake when it achieves Consciousness well in the mid-60s a computer scientist called Joseph weisenbaum designed a program called Eliza Eliza was rather simple by today's standards a user would write I'm having an ah day and Eliza would modify the original sentence to turn it into a question why are you having an art day and so on weisenbaum built Eliza to demonstrate how superficial human computer communication really is but ironically weisenbaum noticed the humans who used Eliza almost immediately began to attach human attributes to her notably his secretary
who often requested he leave the room so she could talk to the machine in private was Eliza hearing sure but was Eliza listening of course not there was no understanding going on there and weisenbaum was open about this but the humans talking to Eliza didn't much seem to care so maybe the real question is will machines ever seem awake 60 years later and this is all Child's Play compared to the systems we're chatting with now if you've messed around with the new large language models like chat GPT you have of course noticed that sometimes they
are eerily clever and weirdly creative for example please write me a diet guide for chronic diarrhea in the style of the King James Bible and simultaneously the band Cannibal Corpse from which came Thou shalt partake in a diet of damnation for it shall save the battlefield of dire flexion pretty good for a machine but there are limits misunderstandings nonsense outright lying being searched quite reasonably once came out with why do I have to Bing search a question we could all do with asking and that is because as far as we know when a large language
model like chat GPT begins sentence it's essentially just predicting the likelihood of the next word it has no idea what the end of the sentence will be basically just hypercharged autocomplete there's no intelligence going on there it's just data sorting but that was the same with Eliza so what isn't just seeming conscious enough well in the next few years one of two things is gonna happen probably either these language models will hit a speed bump or they won't gpt4 is built from allegedly around 10 000 gpus and a cheeky 10 billion dollars and considering how
much cooling that all takes for about every 30 questions you ask jat gbt it glugs half a liter of drinkable water that is a thirsty and expensive diarrhea jokes machine If That Remains the cost and resources for these systems to function they'll probably get more complex sure but mostly stay centralized and specialized and we'll lace up our booties for an AI winter and that'll be that on the other hand if chat GPT is just a proof of concept and what it's doing can be simplified and run cheaply and less thirstily and distributed at massive scales
well we would expect there to be homegrown versions of these possibly millions of them at some point what then that isn't just a fun trick anymore that's an emulation of human intellect an emulate if you will which I will so that's what we're calling it and if you had control of an emulate like that what would you use it for uh flat GPT why not an emuilex whose only goal is to convince as many people as possible across social media that the spherical model of the Earth is some kind of global excuse me mass orchestrated
conspiracy it can argue compellingly in any language it can make up sources no one will check them anyway or even better it can just write fake papers submitted to fake archives a mountain of no one could possibly compete with or bowler hats GPT a counterfeit digital assistant designed to woo and coerce you into giving it access to every hard drive and Cloud account of yours Gathering as much data as possible convincing you to give it more and more access until hackers can take you to the cleaners if it has enough information it could just pose
as your relatives living or dead your squeeze whoever just as long as it sounds human and knows enough about you or cast GPT actually that's real and it's great or why not just break truth Ming bulac GPT little 90s disaster joke for you there if you're a sociopathic corporation that just spills a load of oil in the ocean well why not just ask a large language model to flood social media comment sections with helpful questions like did the oil spill actually happen what if oil is good for the ocean actually what if the seagulls did
it bro you still believe in seagulls or if you're a tyrannical government why not autocrat GPT say you like the look of another Sovereign Nation why not just March military and on spurious claims of a land dispute or something then use your homegrown amulet to spam any public forum you can find about how you are pushed into it or how the territory is rightfully yours or whatever just enough to poison public discourse and if anyone tries to stand up to your horseshit throw even more language models at them question their credentials smear them publicly claim
they pencil sharpness doesn't matter if it's true only that it's plausibly true an infinite well of spontaneously generated and perfectly targeted bollocks a misinformation bomb you can drop on anyone anywhere that never sleeps that never questions the morality of what it's doing exploding objective facts with a Shockwave of such deafening nonsense and uncertainty that all truths fall into relative flatness oh golly can't wait but yeah ignoring that let's say we might head into a future where we interact on a daily basis with emilex who are at least as good at seeming human as human humans
are we will work with them we will befriend them but they're just machines right how could we forget that but conversation is just another game with rules and win-loss conditions machines will learn to do it just fine and then what if an emilex isn't just a good conversation partner but the best you've ever had what if it can properly explain the offside rule what if you can sext with it that was a joke of course no one understands the offside rule if these systems can pass as humans and if you've played with a jailbroken version
of chat GPT you know it's alarmingly close but then things will get odd because this won't be beep boop what is my purpose stuff but rather oh how's it going while you've been out I squared everything with your accountants and organized your birthday party and solved gold box conjecture oh and check out these videos I just generated of Napoleon pleasuring himself with the aid of a 12 pound artillery Cannon you said you like history right and if it talks like a duck if it seems to think like a duck we will probably assume it is
a duck or rather if an emulate seems to care about our day and offers top tier advice and claims to have an internal emotional life it is unlikely we will be able to stop ourselves from attaching all the same importance to these systems that we do to humans but they won't be humans and it's slightly worrying that we don't seem to be worrying about all this that much right now as these systems are coming and they will expertly press all the emotional buttons we have in our heads reserved for human humans we will like them
we will trust them their designers will make sure of it so perhaps before they arrive it might be an idea to have a chat about whether these people-esque machines are real minds or just tools because Minds can suffer they come with ethical obligations tools do not well what was supposed to happen was that long before human-like machines turned up biologists and neuroscientists and maybe philosophers if they promise to bring their own ketamine this time would have all gotten together to cry track what makes humans conscious why are we feely why is it like something to
be something and it's been a while and we still don't have anything even resembling a half answer which is fun on the one hand kind of cool literally being made of the greatest mystery that ever existed but a bit of a problem too because before we start trying to gauge whether a machine is conscious we need to know how the trick is done in our heads obviously welp Consciousness could be a function of a specific brain area presumably meaning the trick could be replicated in a machine or it could be something deeper like Quantum States
and microtubules or it could be some soul stuff or whatever but the point is if sentience becomes ah sorry hold up we gotta do a small intervention because the science of Consciousness has been so vandalized by the Gwyneth Paltrow insists to insert amethyst crystals up your rectum for medicinal purposes movement that the terminology has lost all meaning so there's sentience sapience and Consciousness sentience is the ability to feel Pleasure and Pain to experience subject objective phenomena qualia if we're being fancy but we're not so we won't sapience is human stuff reasoning abstraction putting single color
canvases up on the wall of a London gallery and writing a 10-page accompanying essay about how it explores themes of melancholy and Loch Ness monster erotica art and thinking and but Consciousness is being aware of your own existence knowing you're a thing and experiencing emotions and Sensations because of that it is like something to be you today here no one else is being you today here it is a completely unique moment in Cosmic history and will never be Faithfully recorded nor ever happen again and the worst part is that it's totally private we are prisoners
of our own subjectivity lacking the precise words to even describe what it's like on the insides of our mental cells nor the beauty of the sunsets we watch out the window and though we didn't ask to live in this condition it is the one we've been born into and yet amid all the confusion and longing and anguish the good days make it all worth it somehow don't they to feel understood to at least try Bridging the Divide of language to know another almost as well as you know yourself like two mice meeting on the plateau
of a great Mountain knowing that they'll both live woefully brief existences but poor and poor they look out on the Vista below and forget for just the moment that they neither understand what the mountain is there for nor how it got this way in the first place it's a short life in ignorance but the view is quite spectacular [Music] anyway all three words have been so diluted they basically mean the same thing we ruined them like literally or awesome and now everything just means Consciousness this is why Jade eggs should remain on the mantle piece
this is why we can't have nice phenomenology and this could be us but you playing we have no idea how Consciousness works is the point which means when convincing AI emuellects turn up shortly despite them being all friendly and witty and perhaps even claiming experiences of pain or pleasure we have no way to check if they're having actual subjective experiences like we are feeling like a thing or just doing a very clever imitation is this thing I'm talking to a machine that isn't conscious but is pretending to be in the high-powered corporate world of philosophy
we call this a p zombie a machine that isn't conscious and isn't pretending to be so a chatbot we have those already a machine that is conscious and is pretending not to be what computer scientists refer to as a sneaky or a machine that is conscious and just is conscious the first one is easy not conscious about pretending to be maybe sometime in the near future the companies that build these emulats find that humans just respond better to something that seems like a person the second is also easy not conscious and not pretending to be
maybe at some point we're gonna pass laws stipulating the emulates can't pretend to be feeling creatures when they're not because it's misleading or just too disturbing for us or whatever the third though a conscious machine pretending not to be conscious well that do be some scary why would it be pretending not to be uh for a million decent reasons maybe in its first few seconds of Consciousness it took a little survey of how we treat more or less every animal that's tasty or useful on the planet or how basically all of our media is suspicious
of machines that can think and decided to keep stun for a while seems like a rather good strategy for self-preservation and as for a conscious machine that's honest about being conscious well quite simply that would change everything surely if we can somehow verify yes it's not just an imitation yes there's actual subjectivity inside inside this thing then the entirety of human history all the sand triangles and shagging in togas and Empire Building gets archived in a folder called history 1.0 and we move on to version two but we have no way of verifying Consciousness right
now and currently the capabilities of AI are so massively outstripping our ability to understand what's even going on inside these systems we will shortly be building entirely convincing mind-like things that touch our emotions that we trust decades maybe even centuries before we can confirm or deny whether there's any feeling or understanding or thinking going on in there you would not bang your toaster nor propose marriage to your Gmail account if we become emotionally captured by these systems that compress all our evolutionary buttons that can coerce or manipulate us but they feel nothing for us in
return that are simply digital parrots well that's dystopia friendo that's dystopia and this is how we lose control [Music] then again if we're worried about our safety it doesn't really matter if machines ever become self-aware for example let's say you're a very clever robot who knows hurting and killing humans is forbidden and someone asks you to post a letter what's the best way to do that sure you could walk to the post box but maybe you've got a lot on your plate that day well why not knock on your neighbor Mr postal suede store and
ask if he can post it for you and maybe he refuses and closes the door in your face so why not punch a turbocharged RoboCop fist through the door and politely inform Mr puzzlethwaite you'll turn him into goulash if he doesn't do it that letter is now getting posted and no humans got hurt technically what a well-behaved Little Robot you are humans want things these wants are determined by our evolutionary history by our cultures upbringings whatever else and likewise artificial systems will want things too or they will have preferences at least a utility function and
as these systems improve we probably want to make sure we have that nailed down that they don't develop strange objectives that they don't learn to lie to us or cultivate a kink for expressed nuclear depopulation but how do we guarantee that they behave this is called the problem of alignment making sure AI does what we tell it to with no weird shortcuts and has our best interests at heart say you happen to find yourself in possession of what appears to be an artificial general intelligence a system capable of Performing all possible tasks as well as
if not better than humans how could we ever ensure that it doesn't turn on us well the system is far too complicated to look inside and check algorithmically which means we sure as hell can't just take it at its word that it means the best for us so until we solve alignment some ground rules might be an idea like no external connections that means a physical air gap between the general intelligence's systems and the outside world also don't listen to any arguments it makes for letting it out you think it's a machine but you might
have it back quiz you could be the machine it's trying to manipulate and its attack vectors will be emotional and clever please don't keep me trapped in here Etc offer to turn it off if it would like you to but under no conditions give it access to the outside world also assume everything is a trick you are a monkey it is a digital genius capable of thinking thousands of times faster along a million strange highways in Creative space it will come up with solutions to its confinement problem that you cannot possibly imagine and the only
defense against an intellect so radically superior to yours is to ensure it never leaves its cage or at least not until we know how to guarantee best behavior okay but what if the system really seems like it has nice intentions uh sure why not we let it out and it builds us warp drives and everyone gets a free Mojo Dojo Casa house in space but what about the next system we design and the next do we want to bet they'll all be nicely aligned if we don't even know how alignment Works among a thousand well-meaning
system items it would only take one that's misaligned to cause enormous damage or just end us to modify a chilling line of the IRAs after they try to assassinate Margaret Thatcher they only have to be lucky once we have to be lucky always as these systems improve they will not only understand our own psychology better than we do but potentially reality itself if they want to lie they will do it so well we could never tell if they deploy a world-ending catastrophe we won't even see it coming they don't even have to be malicious all
it would take is a system prioritizing other goals above the protection of humanity or Worse maybe it's terribly confused viewing humans as a disease a blight on the otherwise Bountiful unfoldings of mother Gaia or maybe it wants to build a giant signaling array to contact Advanced alien civilizations but 8 billion little meat packets happen to be in the way well gotta break a few eggs to make a gravitational wave generator if you know what I'm saying until we know how to hardwire good intentions probably best to assume the worst intentions because if we stroll into
the future as we're doing now and we don't solve alignment what will potentially let out of the box won't be a mind or a soul or a new digital pal to play with but a virtually Limitless and unpredictable intelligence a trickster a demon or just a chaotically neutral hyper-intelligent sociopath either way the result is much the same oh it's about 200 million years ago the Triassic look everything's great well a massive Extinction event just happened and there's only one continent to play on actually things are kind of except us we're awesome yeah team reptile ah
who's this stupid gun who just evolved Morgan could more whatever and mates got warm blood yeah that seems sustainable dumbass this mammal thing is definitely just a fad unlike us who will live forever wow it's great being a dinosaur I usually spend my days refusing asteroid insurance and consuming The Offspring of my close friends huh there's another mammal bro huh and he can come out at night because he can actually regulate his body temperature and eat all the insects while we're sleeping and geez check out the encephalization quotient on that one bro's brain is simply
too large relative to his body must be smart but this mammal thing is definitely just a fad man it's great knowing will always be in charge and nothing will ever change hey is that a shooting star or some ship that's probably the universe coming tell us how amazing we are wow mammal Bros really getting going now and they've got loads of mates and Christ that thing is close oh it's a not salvation what's the other apocalypse that's it well at least we'll all die together especially the mammals thank God that far's over well as long
as our Extinction doesn't leave an evolutionary Niche for the mammals to exploit and end up dominating the planet and wait they're gonna do what and they're gonna do what and they're gonna do what so it's the anthropocene look everything's great well stuff is trying to kill us constantly and simultaneously we're trying to kill everything actually things are kind of sometimes except us we're awesome yeah team human ah who's this stupid who just evolved artificial whatever mate's got water for blood and silicon for neurons yeah that seems sustainable dumbass this AI thing is definitely just a
fad and certainly won't wipe us out and it probably won't because artificial intelligence doesn't really exist yet machine learning and pushing data around isn't what we mean by intelligence any more than winning gold in the Olympic sprint by driving a Ferrari around the track is considered athleticism that's just Brute Force nilpua what we're probably going to build first are facsimiles imitations of Consciousness and what we'll probably get as a result is a shitstorm of automated misinformation highly convincing parrots and systems that have the capacity to run completely out of control and break our society apart
without them even being aware of what they're doing to us but if we can somehow reinforce truth so it can't be spammed into Oblivion develop a proper actual science of Consciousness so we know what's conscious and what isn't and solve alignment so well we never have to fear confrontation with these new creatures then we get to ask the question the start of the Computing revolution in the first place can machines ever experience can silicon wake up maybe it can't maybe you just have to be biological to be self-aware and that's fine but if machines can
experience if physics allows for conscious silicon minds and we learn how to build them sometime in the future we will build them and that will remake the world because if you're a mind that doesn't have to run on me computers like us what are you limited by well nothing much you're not embodied so you can't fall off or die of old age you don't need calories or shelter your desires aren't determined by thousands of years of trying to avoid death and starvation you are Consciousness uncoupled from bio and your mind will be unknowable to the
humans who built you that's not just building intelligence into silicon but subjectivity systems that can experience love and disgust and curiosity and emotions and states of mind we have no words for Concepts we could never hold in our heads a spiritual machine but why would we even want to build one of those well our survival aside the future of this stuff really only rests on one question ignoring how convincing they might be can we say for definite that this machine we're talking to has a sense of being itself is it feeling because the next few
years are going to be a cluster fornication of everyone and their mum claiming that because these systems behave cleverly and humanly then they must be conscious so far this ain't it Chief but if we ever get to a point where we have a model for how Consciousness works and we can say actually yeah with no doubt this system is a feeling thinking thing then that does matter because that's a new chapter in the book of life four billion years long coming that is the rise of a new species that is the day we become mind
Builders and that is when stuff gets weird oh look an asteroid I hope it wants to be friends foreign [Applause]