Is the world moving closer to an AI singularity? | Ep. 35
new developments in artificial intelligence continue to generate headlines along with deep fake images and videos that are causing some experts to call for a pause in the development of this technology is the world heading down a path of achieving general intelligence sooner than expected and if we are is that a good or bad thing we'll discuss the possibility of the AI Singularity on today's episode of today in Tech [Music] hi everyone welcome to today in Tech I'm Keith Shaw joining me to discuss the latest developments in AI are jodica Singh she is the director of data science at placemaker Chris Tanner he's an MIT lecturer and head of r d at kensho and Nick mate he is the an assistant professor of computer science at Tulane University hi everybody welcome to the show let's just jump right into the the discussion here um there's been a lot of developments with chat GPT generative AI a lot of these image creation tools getting lots and lots and lots of headlines um and as I'm as I'm kind of discussing these on the show we also get a sense of we're moving in a direction towards this concept of AI Singularity the idea that eventually artificial intelligence will reach or surpass the the level of human intelligence out there just to start us off uh what why don't we I get like a general definition from you guys about what you guys consider the you know the the definition of an AI Singularity jodica why don't you be start us off yeah for sure so uh when I think of yeah Singularity it's this point in the future where AI surpasses human intelligence now one way to look at it is that it surpasses human intelligence in a lot of you know ways which can really benefit uh Us in general but the other way is like it develops its own conscience in a way that it's now learning on its own it does not need human input but it's also going out of control uh and humans are unable to capture that growth all right and Chris what you know did is that the definition that you guys sort of use as well do you do you tie it with the the idea of Consciousness or is it just sort of like being able to to um like how do we know when when it surpassed human intelligence yeah I would agree with that I should also say that that in general I would say it's pretty rare for I for me to hear folks talking about Singularity AGI is the more common thing okay in some ways one can view AGI as kind of being a precursor for you know before you get the singularity okay um but yeah I would say that both of these things are would fit my definition of Singularity you can almost view it as two axes like how controllable or uncontrollable it is and then just how intelligent this thing is all right Nick anything else to add on that uh so I I didn't have a great background on sync on the definition of Singularity so I went and looked it up and did a bunch of reading yeah on my my long flight that I had yesterday so I didn't realize that it's sort of the the term and like this idea of existential risk grew out of um some stuff that John Von Neumann who's a very famous computer scientist sort of talked about in the 1950s right so this grows out of this idea that technology is out of control it was coming out of the rain preparation um in the post-war period when we were worried about nuclear winter and things like that so a lot of these concerns um sort of come from the same sort of place of you know what happened like technology run amok right and so that's why I I kind of um a little I don't like to use the word Consciousness personally um but in terms of the fear I guess or the the worry about Singularity is this idea that technology is advancing ever more rapidly and that at some point this is going to be beyond our ability to understand it and it might pose you know this very kind of cold war period existential risk type idea like it might end all of humanity right and I think that that you know Hollywood wood has taken over sort of the singularity idea as well you know the Terminator 2 there's the whole Skynet thing and and you know Skynet there was a one you know point in time when uh the AI achieved a sentience uh and and Bloom everything was downhill from that point you know I've I've heard a lot from Ray Kurzweil he's he's also said like he's been making predictions that I think the latest one I saw from him was that by 2030 which is only seven years away uh that he believes that's when you know it'll achieve this Singularity so as it sort of got confused and melded with this idea of artificial general intelligence Chris you know you brought up that that it felt that feels like there's a difference yeah yeah I mean I basically uh saying the same thing as what Nick was kind of implicitly saying that it's not really a term within my Circles of folks to talk about Singularity AGI is the the thing that most people talk about or excited about or maybe concerned about um so what what is that what like how do you then Define artificial general intelligence versus maybe what we're seeing right now yeah yeah yeah yeah good question so basically the the idea with AGI is that we know that computers have always been better than humans for at least very specific things like calculating numbers doing arithmetic and this is their Computing machines their calculators but yeah now we're at a place where they can uh rival human performance on a lot of different things so the concept of AGI is I mean I don't think there's no like strict definition but from some large large group of tasks that uh humans just struggle to be computers on that would be that would be AGI um yeah but we're already you know in some ways go ahead yeah yeah sorry oh I was just going to say you know a few examples would be like language translation right computers are able to do that so much better medical diagnosis uh and then you also you know have had these you know robots defeating uh you know these chess players and go players uh and really outperforming humans in those ways so those would be some examples to add to press the point yeah does it does it seem like in order for it to get to a general intelligence points that we would need one sort of system or app that can do multiple things because it because I always thought that the the joke was that the computer that beats me at chess for example I could still beat it at Go Fish or or I can you know I could still do other things better than that specific computer do you think it needs to we need to have like a system that can multitask or can do all of these things and that's when you would know or is it something else is is was that just too simplistic well I think for it to jump in a little bit here the the that was one of the goals of some of the programs that you saw for go um that deepmind was working on was but they're all your second example was still a game right and I think uh Chris would sort of speak to the the idea of AGI is that it's good across lots of different tasks that aren't just games so it's effective at driving and route planning and translating and playing chess and and right because I think like jodica was saying you know computers are better at us in a lot of ways right computers uh find routes faster than us computers route our packages better than you know we used to sit there and actually like plan where packages would go and how boats would drive like we don't do any of that anymore right like it's all done by these scheduling algorithms that that process data much faster than humans do um and able to process larger amounts of data so the question you know getting back to our Singularity topic is like if are the computers better at than us at very specific tasks and that they are and they've been and they're going to continue to be um the idea of AGI in some sense again like Chris it's not a super well-defined term is that you know you've got one system that does all of these things that recognizes people on Facebook and and also can drive your car and also can cook you dinner right um and and I don't personally think we're gonna we're anywhere near that uh despite the fact that these language models might try to convince us they are um but that's the uh you know that's that's the AGI concern I guess yeah I've also seen a bunch of things recently about um there's been a there's an article last week that said someone thought that Chachi PT and and these generative AI chat Bots have passed the Turing test and then I saw another one today that said no it hasn't yet but it you know here's when it will sort of he's thinking that gpt5 uh which is scheduled for later this year to be released uh that might be able to pass the touring test um I have a t-shirt I forgot to bring it in but I have a t-shirt of a robot looking over someone a human shoulder and says I cheated on the Turing test um and I think it's a funny shirt but whenever I wear it no one look no one understands what it is they're like what is this are you the robot or are you the human and which is kind of funny because now no one knows you know the human versus if you were to write that T-shirt and and human language and give it to chat gbt would it get the joke yeah yeah like hey do you understand this joke uh so but like is there do you think that we're getting closer with that in terms of like for those that might not understand uh what the Turing test is jodica can you can you sort of uh explain what the Turing test is yeah I mean it's basically a test to uh to give the situation to a robot uh and the the judges don't really know if it's coming from a robot or a human being and they're supposed to convince the judges or convince the panel that they are a particular human being so really adapting to the situations and trying to pose a convincing argument such that it makes this difficult for people to understand whether it's coming from an actual human being or not uh right and and so uh my uh my director here actually said that uh ex machina was a great movie on AI about the Turing test would you guys agree with that I haven't seen it unfortunately it's on my list I haven't seen it either Okay it definitely speaks to the topic here because I think the robot goes crazy and kills a bunch of people at the end so oh good spoiler alert oh geez I'm kidding it's probably like what 30 years old or something 20 years old I think it was more recent than that I think the dystopian robot movie I mean it's something bad happens sorry Chris well all right let's let's go down this path a little bit too because you know there was this open letter that came out last week of of Elon Musk and a bunch of other AI um uh experts that have signed you know saying like oh wow this we should hold off on maybe putting us a six-month moratorium on on the research and um first of all um have you either been approached to either sign this or um or have you signed it as or is this more of like I just wanted to hear what your thoughts on this are anyone um Nick why don't you start I'll I guess I could jump in I I did not sign it I have not signed it okay um I thought it was pretty it's it's interesting that they're getting a lot of credit for this this letter um if you if you read the letter it's kind of interesting it cites a lot of the work of uh timick group who's a researcher who was at Google for a long time who uh was fired for publishing a paper to point out some of these issues about three or four years ago um and so it's it's been interesting that that's specifically about large language models and about say gpt2 right so these are these ideas at least in the research Community been bumping around a little while I think the letter itself um is not specific enough for me like not to be the professor guy around here but like it talks about power it doesn't Define power it talks about you know research it doesn't Define what that means um and I think it's these these bands are really uh on Research I I don't know it's it's it seems tough to put this Genie back in the bottle I guess um I I was talking to a colleague here you know there's all of these new you know mix and match versions of these large language models that are already out there then there's more and more every day um and so slowing this down might be is might be a good idea but at least the letter itself without any sort of details uh just leaves a little bit to be desired in my in my opinion so that's why I sort of held off on signing up do you have an opinion on this do you do you have go ahead uh I haven't signed it either uh I have read it and uh I think some of the points that are in the letter they make sense right like the concern is the spread of uh misinformation through these AI generated content uh you know deep fakes uh potential misuse of AI for other malicious purposes uh and just the concern that the technology is developing so far is that people are not able to fully understand it yet uh and you know so what I've seen is those in favor of uh this uh particularly talk about like uh having some shared safety protocols uh for AI systems in general that they can Implement you know worldwide uh essentially um and how they can ensure uh like policies put policies in place that ensure that these AI systems uh have a positive impact have like manageable risks uh and also give time for the society to adapt to this AI different world so I'm not saying like those are bad things I think we do need some AI governance here uh and uh those are concerns like the spread of misinformation is a big concern I think it was 2020 when the number of deep picks out there was reported to be more than 85 000 but today after the development of all the technology since then it's potentially going to be in the millions if not yes so there needs to be some regulations but I don't know if a six-month moratorium really does that um yeah yeah but it did you know I think the best part of that letter was that it did sort of raise awareness that that there are people that are concerned about this um if it was just a bunch of people that maybe were were known within the AI space but were not known sort of broadly um then that might not have has made a bigger splash the other thing that concerned me is that it did feel like there was a lot of Doom and Gloom around this it's like yes if we don't do this the world will end tomorrow and it's like okay well you can't really like that's not gonna get much attention right Chris did you did you um were you approached to to sign it or did you look at it uh yeah definitely look at it I was not approached uh I have not signed it although I very much support the the sentiments um not so much about like the Doomsday but just the the reality that these are very important um very important aspects that we need to reflect on within society and consider uh reallocating resources not a policy maker I have no idea the right feasibility of any of these things um but yeah my my biggest kind of uh concern or question marks or whatever is kind of to Nick's point that uh how would this play out and what are the details I think it's just very unrealistic unfortunately like there's no easy solution for any of this and the letter is very forward thinking as it should be the reality and even even some of the the experts within this space um have kind of criticized it because the reality is uh even our current situation or even six months ago we should have been dumping more resources into uh better you know having better explainability you know helping mitigate bias and all these things I mean people have been saying this for for years but the reality is that all this is a Continuum and uh we should have we we could have and should have been putting more emphasis on this right a while back right well let's talk about explainability and the the bias part of it is that are those the two biggest concerns that a lot of people still have about uh this technology or or is it something else everybody else just jump in with yeah okay you think it's something else I mean it's in addition okay to explainability and bias well I mean kind of yeah it's for let's take deep fakes for example of course there are tons of bots on any social media platform but now because the technology is getting so realistic like what's the prevent just just incredibly massive amounts of bots going everywhere and it being really hard for any user to discern what is real or not that's just one tiny example but you have a point like there are multiple issues that people are concerned with rightfully so okay well what about the explainability part like like again every time someone tells talks to me and says explainability is important in Ai and then I go well can you explain that more and then they go no my car but we don't know we don't know why some of these things are producing the the results that they are and that just starts I start scratching my head I'm going well if you guys don't understand it well now you can understand why we would be concerned from a from a Layman's perspective uh Jonica do you have any thoughts on sort of the explainability thing like how do we explain explainability better yeah it's a very challenging one right it's true like the people who build the models are not going to be able to tell you exactly what the model is gonna output for a particular input like we just don't know because the models are developed in a way that they are learning from these patterns that they are Auto uh understanding Auto detecting from the data so explainability is difficult uh and I think you know what might make people uncomfortable is that if they don't understand how something works at all uh and then there are all these concerns about AI getting smarter uh than humans and taking over the world yeah I think it's just important for them to understand some process that goes into building a technology like that so it may not be in detail like what are large language models and what do they do it may even be that but just an understanding of you know what data is used to train this what is the overall principle uh just educating people on what that is increasing visibility uh I think would help people a lot there yeah Nick do you have anything to add on that one or I gotta because I got another question for you um no no I I think explainability is a good one explainability mean it's like the word fairness right it's one of these Concepts that really come to us from Philosophy from moral philosophy that have lots of different meanings and I think that's why it's tough to really pin down um a lot of times most of the conversation I have a talk that I give sometimes in class it's like here's all the definition of like ethics and AI or fairness and AI that people want to talk about it's like 37 things yeah um explainability to a lot of the conversations I've been having around explainability especially for these large language models are really about um what someone from software engineering might call traceability right or might call like why is it saying like where did it get this thing that it's saying to me right so can I take an output of the system like a sentence like you know the moon is made of cheese and figure out where in all the training data like where is the sentence that it's using to justify that right so that's a lot of times what most folks are thinking about when they're thinking about explanations Within These large language models is why is this saying it because I think people do have some most people that I interact with have some some intuitive understanding that these things are basically reading the internet and then spitting things back out at us right and they want to know by explanation like where did you read this thing or where did this come from you know what is your Source material for why you're saying this thing right the reality is it's just a giant text prediction machine like the thing on your phone and it is just there's probabilities on it and that's what it comes from but people want this explanation of like where did that where did you get that right and that's really kind of some most I think most of the thing that people are concerned about it's like why are you saying this like where did it come from when I figure when I get an answer from from chat EPT in terms of uh my question of like oh what is this what is this technology um about or you know when I type in something like hey explain to me zero trust security for example it scans all of the articles that are out there that have P of people that have already written this stuff and then it waits whatever it assigns weights to you know some of these answers and that's what I get back and then I get back sources as well I get links and things like that but when I ask Chachi well you hope so so it's not doing what I think it's doing okay because right right because my next one would be then I I tell it to come up with 10 fake names for my Dungeons and Dragons character and now does that mean that it's going to these these random name generators that are already out there on the internet and then running those or is it just throwing out four random names and you know I don't know it's I mean in general it's it's designed to uh put together strings of words that are syntactically reasonable right so what I mean by that is like the sentence that one of the famous examples that GPT the chat GPT can handle now is the question did Aristotle own an iPhone okay right so Aristotle is it clearly a historical figure the sentence Aristotle did not own an iPhone is probably not present on the internet right because it's not a sentence that anybody bothered would would have bothered to write down right but you and everybody here knows that clearly like cell phones were invented after this guy was alive right so so semantically right it makes no sense um because it's it's but it's likely because everybody owns a cell phone so if you think about it as dead person owned cell phone then it it semantically makes sense to put those words together right it's semantically very common but it's not it's not it's not meaningful right um or sorry syntactically very common get my words mixed up right um uh and so chat GPT actually can handle this question now this was my go-to example but if you ask it that question now it'll work um but if you find the older version uh gpt2 won't answer that question correctly so oh okay um that brings out a thing right these things are changing right like even if you have these I have this key example that I used for for like two years and now I can't use it anymore now you need a new example yeah I was just gonna say you know just understanding these these little like high-level principles of how these models work will also enable people to know that you know if it's giving you a source that may not be right like the information may not be accurate uh so know what you want to use it for and know what you don't want to use it for uh and just that knowledge will just help people use the tool better as well right so that goes a little bit into explainability too just or I would say understanding like people able to understand uh how to use a tool what to use it for and what to expect and what to not yeah and and so the other the other question I wanted to ask the you guys was um do you think that sort of AI in general and maybe some of these companies need a better job at sort of explaining the benefits of this you know we've attached sort of this idea of the singularity to a lot of negatives like you know Terminators taking over the world blah blah blah blah blah Killer Robots Etc but I've never seen really any sort of good um powerful benefit statements to this is why we're doing it we're not doing it because we want to achieve we don't want to get to that point where we have Killer Robots but we're going to keep developing it so that we do so that we get a b and c and so you know I've asked individuals and and they've given me some great answers but do you think there needs to be sort of uh hey this is all good and this is what we're going to do we're gonna you know this will help us cure cancer or this will help us get to Mars uh you know et cetera et cetera do like you know do you agree with that statement or or is it just me being paranoid about like why we need a better statement about this that was a really open-ended question I guess Chris why don't you jump on on this one does AI need better PR is that maybe that's the question I think it could yeah it's tough to answer because because basically I think that the more obviously the more that Society is informed period it doesn't matter if it's good or bad like the negatives are the the pros of the technology the better right the more informed we are the better but I I don't know if it's the researchers or scientists responsibility to play a large role in that that branding that that PR and in fact one could argue that intrinsically we already do because what we're working on for example let's just take machine translation like we're trying to say yeah we want to get better at providing this good to the world of translating languages from one language to another or somebody who's researching a slightly different area within machine learning same exact concept like this is what their focus is and like that should be evident by the work that that they're doing but then you know once these Technologies get really good of course it's just up to the wild west of what this reception is and what all gets circulated you know on the Internet it's like that's we kind of can't control that and I don't know whose responsibility that would be but I guess to your point um it would be good to have uh good reliable trustworthy resources where people could go to see kind of the wide spectrum of the benefits and the cons of this technology yeah yeah and I think it's almost like a resource like so it's a resource you can use it really well and you can also use it negatively so it's it's really about understanding that as well like it has a lot of potential for doing really great to the society right for example it's already shown uh evidence of helping medical diagnosis really significantly so that's a big deal uh and then even other Technologies like a robotic vacuum cleaners like they save me time so that I can do other things yeah so there's a lot of good that has already happened and there's a lot of good that can you know happen in the future as well simple example chat GPD it's just so good at like grammar and structuring sentences uh and it does surpass a lot of uh you know people's skill of writing such great sentences so hey that's a great application you know you can use it to communicate better create more understandable documents uh and really use it to help you in that way uh but you know on the other side I was just reading this article yesterday uh which was about how AI has also helped uh you know enhance some of the negatives like uh cyber crimes and uh you know trying to guess people's passwords and using ML and AI you are able to do that also much better uh right you're able to write more realistic spam uh emails uh and really go up with it that way as well so I guess it's it's technology which is amazing and it it can have a lot of great users that come out of it but at the same time you know people can use it for uh malicious purposes and that's where uh you know I think some of that conversation of you know why is it scary and we really understand uh the negatives that it could bring on with it and what do we need to do to address it you know perhaps deep fakes is such a big problem so should these AI Technologies be watermarking any product of AI in a way that can help us uh you know with that problem uh yeah yeah like a lot of such examples yeah and again we've we've we've dealt with this before I mean the early days of the internet you know we're all thought hey the internet is great we're gonna be able to send emails to everybody we're going to connect with all these people and and it was all these great things uh and then a lot of bad guys discovered there's a lot of bad things you can do on the internet and then you know stealing money and passwords and all this other stuff um it felt like any sort of Technology development we've always seen a list of good things that happen but then we've also recognized that there's bad things that happen too but like we didn't try to ban the internet when it when it came out we there wasn't there wasn't sort of an open letter sort of you know so this feels a little different but it also but also we've got history on our side of like yeah we you know people stealing money is bad you know for those people that lose their money but we we didn't shut down everything because of that um I I don't know maybe I'm just I'm just rambling at this yeah I don't know if we some people tried to ban the internet there's still plenty of people that like don't use Facebook we're still talking about regulations for some of these communication Technologies um um for for teens I think there's a bunch of law some laws of stocks in California maybe Utah um so I I think that's true you've got all the tick tock stuff too going on yeah right exactly yeah so I I think it's I think your larger Point Keith is that you know these Technologies come out and there there's always a conversation about how we integrate them into society um and who are the benefits accruing to and where are the costs right and that's a conversation that doesn't just happen from technologists that also needs to include communities that are affected by it it needs to include people that have no idea what this Technologies are um you know and it needs to include government and policy makers and things like that and I think these conversations do happen they just happen very slowly right and that's the you know the it's a lot easier for uh for a future of Life Think Tank to come out and and you know put this letter out there it's much harder to really write rules around um how these things get done um how the how this regulation happens I know I've been on these uh National Institute for standards and Technology panels for the last three years trying to write these large policy documents like the you know about like how we govern different aspects of AI decision making yeah and it's you know it's slow and it's very unsexy um but coming out and saying oh let's ban it um you know this is a quicker way but this it is a big conversation um and the costs and benefits of any technology is is a is an ongoing one you know it's not one that we sort of say Okay up we're done with that right like you know it's always something that it's always where the rules are changing and where the people who are being harmed or benefited you know change and we need to revisit that conversation right right so um Chris I wanted to bring this up with you because I know that when we talked ahead of time um there was a discussion of like where are we if there was if this was a curve and you get that hockey stick sort of uh effect in terms of where AI is at the moment are we still at that sort of the Baseline of of the hockey stick or are we now like on that trajectory upward or you know you know if if you had if you had to make a guess of you know April 2023 this is where we're at on on the on the graph or is it is it even too hard to do this at this point while we're going for it of course it's yeah of course it's like essentially impossible to predict these things uh and things have been accelerating at an exponential rate I would say for the past I mean like in some ways always right you're always on that Continuum but especially the last 10 years have been absolutely nuts um but if I if I had to guess all right it's just a fun thought exercise I would say we're we're at the early slash mid point of the uh the crazy high derivative yeah before any Plateau okay um because yeah because it's just so hard to um so hard to even guess what this is all going to enable and I think like not to backtrack but to the previous question I love that analogy of comparing it to the internet and to Nick's Point like yeah even things like the internet took a long time the first internet connection was in 1969 and then you know it really started to enter a lot of people's households in like the mid 90s yeah but it was it was uh it received I think slower pushback on the internet than the pushback we're starting to see with for example chat gbt I think it's because it was hard to anticipate and imagine what all the internet would afford whereas which had gbt it was almost like a stepwise function to the public that technology has been building all of our researchers knew the capabilities of NLP but all of a sudden bam it was like very in our face very easy for anybody with a computer to just play with chat GPT and see the benefits of it and thus also you know the pros and the negatives right now I yeah does anyone else want to jump in like are do you agree with Chris in terms of where we're at right now or are we still really at the early days and you know it's going to be even crazier next year I'm sure it will be crazier next year like every day I see a new story I'm going oh my gosh like I can't believe I can do that now yeah I agree with Chris and uh you know large language models for example have existed for a few years it's just this has been this first time where something like this like charge GPT has been opened up to the public and so many people are using it and that's why you know there's there's a lot of attention that it has been getting but I agree with Chris I think we are somewhere we are yet to Peak for sure all right Nick I you know you asked me about this question I still haven't really I don't know right because it's it's it's the idea of like do you think the curve is going up or do you think there's eventually going to be sort of like diminishing returns right like is it going to sort of slow down and like Crest off right because if we are we gonna accelerate are we going to slow down and I I honestly have no idea yeah so it feels like we're accelerating at the moment given that you know the this first started in December of or November 2022 and you know since then it's just been non-stop news and non-stop sort of discoveries right but I mean you could you could say the same thing about like you look at something like plane flight or the speed of cars right like it was going faster and faster and faster and faster we got on rockets and then all of a sudden we're not really going that much faster as humans anymore right so it's I I think it's really tough to Chris's point it's really tough to to know you know are we going to keep going up or are we going to Plateau off like I mean it's it's it's really tough and I think the only constant is that it's going to keep changing right and that our perspective tomorrow is going to be probably different than our perspective today which um I don't know all right so is it crazy and liberating at the same time that's what makes this this topic so fun to talk about um exactly I think one way of trying to ground it would be to think about what are the big remaining things that we've barely been able to make any progress on and kind of to kind of to what you're saying earlier Nick I forgot to which question but to think of like multimodal stuff you know like it's so hard to mix video with with language and we've been doing you know we've had incredible success over the last three years but there's still so much room for improvement and I think we're gonna see some really impressive gains in the next two to three years well well does that you know that made me think of like do we need an AI moonshot sort of sort of approach maybe from and again I don't want to bring the government into it but you know when um you know back in the 60s there was like Hey we're gonna go to the moon and you know they had a goal and and they got there and they landed on the moon and that it galvanized everybody do we need something similar in AI or is it is it is it just that's not a big that's not an idea because we don't know exactly what the goal would be like but again practitioners huh yeah Chris what I'm sorry what'd you say well I didn't mean to dominate here but I was like I think we already have like just millions of practitioners who are already doing moonshot approaches which is good I don't see this disparagingly right yeah just so many folks in high school right without even taking formal classes in this like we've democratized it and we've lowered the barrier of entries so I think people are trying outlandish things and stuff is happening progress is happening all right so so I want to and again I think go ahead I got another question I think the um I wanted to tie it back to you talked about a moonshot I think the Turing test as a as a concept which we talked about before right that was a moon shot right it's that this this you gotta remember at the time computers was this thing you ran with a crank yeah like that you like you punched some cards and you turned to crank and like the idea that this thing was ever going to be able to convince someone that it was a person right that it was intelligent which is really what the attorney Test's all about it's saying okay well if you don't know what intelligence is I don't know what intelligence is but if I can convince you I'm a intelligent thing AKA another person then I must be doing some intelligent stuff but the Turing test on its own could be understood as a moonshot and that's why it's it's really been an animating concept for the field for so long um is because you know in order to do all the stuff that you need to pass the Turning test you have to you know get over all these little challenges yeah that's why it's been such so it's why it's captured so much imagination there like you said because it is a moonshot and I think it's you know we maybe need to refine it a little bit right because because now we're kind of at this point where it's like okay well the computer can kind of convince us it's a person sort of uh and you know is that you know what's sort of the next thing but you'll see some of these right there's the idea some of these protein folding things some of these um AI for science type ideas like there's a few of these in the research Community but they're not as like big and grabby I think as as kind of um like you said going to the moon or maybe the Turing test originally was so yeah maybe it's time to rethink a like what's a new sort of cohesive one um would be a good question yeah I'm still not convinced that I'm I'm a human it's like so I I may be a robot at this point there's still there's still some debate um especially for my family uh guess what you you passed the touring test I think you I I thought you're a human well yeah some of my family might disagree with some of the decisions that I make um jodica I'm going to ask you sort of a hypothetical question I'm going to put each of you um you are now in charge of AI like we we I made this decision because I'm emperor of the world okay so I said all right jodica you are now in charge of AI development what is sort of the first thing you would do like do you do anything do you do you know like where do you if if you could steer the direction of where AI goes from this point what would you do well that's a very difficult question uh do you have the power that's a very tough one yeah assume that you have the power to do whatever you want and and so where where do you and I'm gonna ask the uh the Nick and Chris the same question so I'm not putting all the pressure well I think I'm going to first spend some time just thinking about everything and every aspect that I haven't already right because that is like really a lot of power to have um I think yeah development is has the potential to do a lot of good but then as mentioned earlier right that there there is a downside to it there needs to be some regulation so it would perhaps be a lot of conversations about how we can uh develop AI as well as uh these policies and AI governance and some regulation alongside AI development so there's not that big gap that we already made so many advancements and now there's so many problems but the policies and the regulation takes time to really get there so how can we uh you know do that hand in hand like more together more in parallel uh I think that might be you know one of the one of the concerns that I do have today so that might be the direction that I would be thinking okay all right Nick I've now placed you in in charge of of AI you are the man so what's what's what's the first couple things you do like do you do you go to Congress and go we need more policies I mean it feels like regulation is going to be way behind on this but it's like what what's the first time that's kind of there's a there's a it's called policy vacuum rate is that technology moves so fast that we have this like vacuum where there's no policies that exist um that seems to be what we do in the US right that's why we still don't have like a coherent crypto policy and things like that we sort of let things happen and then um and then figure it out later and then clean up yeah we could have the best after it's a very American way to look at things but I kind of love um some people do some people think it's terrible uh I just got back from an overseas trip so I have you know very different perspective than I normally would yeah um the uh if I'm in charge though if I'm king of the world what would I do um king of AI yeah uh you're only king of AI Nick you're not you're not I'm the emperor of the world sorry sorry sorry I'm stuck at a university now you know so you think we're Kings of everything um I I really like jonica's point about trying to um uh and this is something that we that we try to do here you know a lot of is is really do more Community engagement with this development stuff right I think a lot of the the AI development you look at chat GPT right like what it costs to train one you know one iteration of chat GPT is something like four million dollars or something that came out and sort of said it right so it's it's so inaccessible to so many people um and I think that you know and and Chris picked up on this earlier too like really democratizing these Technologies and democratizing one in terms of making it available and you know online open source Publications things like that but two and like making sure that the communities who are going to be impacted the communities that that maybe don't normally have access to these things do and and that includes the knowledge how to use them and the resources to use them right and so I think really trying to push into um those spaces is is is a good way to go I think all too often you know the these a lot of these AI developments you know by necessity because they're big technology projects are really housed in these large tech companies yeah um are really you know things that happen at universities with you know in the closer Labs where we all think we're king of the world anyway um you know so this is getting out of that right and like really making sure that that's a first and foremost kind of policy and sort of design space um would be would be really good okay Chris you're now you're now head of AI or king of AI what yeah I I haven't thought about this before but the last ago so I've had more time to think about this because anybody else here but yeah the way that I've always uh kind of framed AI machine learning in general is that there are a few pillars and this is kind of the next point that some of those pillars are towards like kind of democratizing things some of those some of those pillars that have really allowed a lot of great Innovation especially over the last five years are things like uh allowing reproducibility of models and historically I'm going to try to make this sustained but historically somebody could research a model and come up with a really great model but then nobody could reproduce it right the source code isn't available or just be really messy and you would spend years of your life I've spent years of my life trying to reproduce somebody else's code that I needed to compare against so this is like one of the pillars you know so like hugging face for example is one company that has really kind of LED this space and and has really enabled a lot of progress within machine learning because they make models available and they make data available I'm not trying to like just advertise for certain companies here but I'm addressing that there are certain things that have really allowed Innovation to fuel and I think we're getting to a point this is also to Nick's point that that there are a few companies that are training these very large language models and of course we don't have to limit our conversation to just large language models but it's a compelling example yeah and and I think the the danger is that that it could shift things to just essentially being a monopoly of kind of the same way that we saw it with operating systems but at least with Linux there's this open source operating system and you could Tinker around and make things your own we do have some players within the space of ml who are making open source great large language models Bloom is the best example that I know of but that's one solution and I think to really kind of help make sure that that everybody gets a fair shot and that things are as democratized as possible maybe you would also involve things like um yeah focusing on on like the computational power right not just the organizations who can make these things available but allow it to be possible for any group of dozens or hundreds of folks to get together so I don't know how that would play out but you know maybe just making gpus like maybe more people competing in the GPU space just making it easier so that it's not just you spend four million dollars and then you can get something I don't know how to enact that I don't know how to do that I would also like to put more resources kind of to what we were talking about earlier with explainability I would love to somehow make that possible maybe dump tons of funding because it's it's not it's not glamorous to work on yeah feasibility it's not glamorous to work on fairness and bias everybody just wants the flashiest best performing models because that's fun um yeah I don't know how to invoke this change but seeing changes towards that would be good all right so what I'm going to do is I'm we're going to bring you all back uh at the end of the year and we're going to sort of say you know we're going to probably see if like if you were still king of AI um or you know that that's offensive on my end I'm sorry king or queen of AI sorry to make sure that you're the head of AI uh you're in charge of everything um but that you know I also wanted to ask sort of as a final question um kind of getting back to the original ideas you know are we getting closer to the AI Singularity I was going to say like by the end of the year do you think we'll have achieved it but I'm getting a sense of based on the definitions that we've had and talking about general intelligence that we feel that it's probably we're probably not going to be there by the end of the year but do these developments get us closer so yes or no I guess from for me to the things are are we are we moving towards that idea of either HEI or or a singularity why don't you start yeah yeah I mean every development is getting us closer to that yes I think that these developments help okay Nick uh yeah I would I would I would typically agree with jodica I yeah all of these developments are moving us in a Direction Where We have sort of more technology right and that's functionally what the singularity is kind of postulating at the end of the day is that like there's more technology and things are going to get faster and faster right so yeah these are all sort of moving Us in that direction and Chris yeah I mean I I agree but if you focus on the control aspect like losing control for Singularity then yeah that's to be determined that I'll just relies on how much uh power we give the models in terms of what they have influence over but definitely in terms of AGI and the capabilities yeah we're definitely making serious strides we're definitely not there but it's headed in that direction yeah and I guess I probably should have asked sort of a secondary question of this like are we getting there and is it a good thing like do we think that we are on the right track at the moment or do we feel like that that we're going in the wrong direction without a couple of course Corrections that's what I I get a sense of sort of from The Angst of the AI community at the moment I mean yeah go ahead okay a tricky question right but of course of course there's a stuff that's not really good that's happening right now so uh I guess we have a bit of both uh right and uh I do think there needs to be some character measures or just again AI governance things we need to think about I love Chris's uh point on you know the shareability of these models as well uh not to drag race too much but you know every time a large language model is uh you know built there it just is associated with a lot of carbon emission as well so there's impact on the environment but just in in general like for example uh we we don't stop medical research so anytime a new drug is developed it goes through certain testing and safety uh you know guidelines that it has to follow uh before it's actually released uh to the public for open use uh not saying that's the exact path for AI but we definitely need to think in the direction of how we can make it more safe so again great stuff associated with the charge GPT for example I've heard people use it for so many you know really good applications that saves people time uh it's amazing but then also uh you know we do need to address uh the other side of things as well okay and Nick are we headed on are we on the right track or are we heading over that cliff on the train I don't I mean I I guess maybe I'm overly optimistic I don't know right like I don't like the this sort of this x risk you know the technology is going to take over right like I just I don't I don't see how we ever get there with better language models or you know jodica always talks about like what ai's doing for us like I just got back again you know all the routing of airplanes and you know getting me to work on time because it knows all the bus schedules and all the stuff like I don't I don't understand how my bus scheduler is going to take over the world um I guess and that's that's where I kind of come from like and I don't I don't see how you know these as they're called sometimes stochastic parrots like are going to get better and take over the world right just because they get better at imitating human language um it's a it's a fun thing to think about but I like I guess I've kind of an optimistic slash naysayer in the room it's like I see all these benefits from technology getting smarter and I guess maybe I'm driving at the cliff much faster than I should be but okay all right and did I ask you Chris did you ask did you have any final thoughts on this like we are headed in the right direction yes yeah I definitely think that we're heading in the right direction but I'm cautious of our current climate of how we will use this technology the technology itself is amazing but yeah I don't know go back to my point about social media right like we've already seen that it can have huge adverse negative uh impact on society people don't know what's real information what's fake information and this is just fueling the fire more and more and more um so yeah to Nick's Point like it's probably not gonna be anything doomsday for train transportation for example yeah um but in terms of social media you know and whatever we hook these things up to like our emails or whatever yeah like it's a it's a bit alarming we need to be cautious I think but the technology is great all for it all right all right so we're gonna we're gonna reconnect in about six months and we'll we'll see how or at the end of the year and we'll see how uh everyone feels about it at that point um because again I just said my chat GPT representative sure just the GPT and like let it do the talking yeah sure sure we'll we'll we'll just we'll just do that and because I I'm sure that by then we'll have you know voice to text to translation and it I'll just be able to talk to other computers and there'll be an avatar representing you Nick and so um yeah yeah personality they downloaded into the internet and we'll be all we'll be all we'll be all set there or that that might be a little too ambitious but hopefully you guys will come back onto the show is that you know can I get you guys to say that at least at this point yeah of course yeah all right and was that too ambitious or not ambitious enough and that's the moon shot we need to have a deep fake of Nick by next week well I all right I could probably have a deep fake of Nick probably by the end of the year I think we have enough audio and video of him now that we could actually create this might be actually now I'm scaring myself like thinking that I could do this like I'm just gonna try to find tools on the internet that could do this for me like I said like you know you you know once once the AIS take over for you know podcast hosts you know then I'm doomed so yeah all right I think we're I think we're we're good thank you guys uh so much for for being on the show today and uh we will we'll catch up in about six months sounds good thanks for having us all right sounds great yep all right that's all the time we have for today's episode don't forget to like the video subscribe to our Channel and add any comments that you have below join us every week for new episodes of today in Tech I'm Keith Shaw thanks for watching thank you foreign
2023-04-10 21:08
