Will ChatGPT backlash deter its potential? | Ep 17
we're now seeing a backlash on chat GPT and other AI tools and wow this came quick and we're going to explore the topic on the next episode of today in Tech [Music] hi everybody I'm Keith Shaw host of today in Tech we are joined today by uh Jason Mars PhD associate professor of computer science and engineering at the University of Michigan so go Blue I'm I got that right right yeah yeah all right Jason is also the acting co-director of um's clarity lab which directs Advanced research within AI large-scale Computing and coding and he's also the author of the book breaking Bots inventing a new voice in the AI Revolution uh welcome to the show Jason yeah pleasure to be here thanks for having me so uh you know this week there were a bunch of stories that that I want to sort of bring up the um uh Story by uh CNBC bring this one up uh Chris the chat GPT AI hype cycle is now peaking but even the tech Skeptics don't expect a bust and then we also saw the New York Times columnist Kevin Roos uh write a story about this really bizarre uh conversation that he had with Bing's chatbot and as everybody knows Microsoft integrated chat GPT into its Bing search engine last week they had this big event they were hyping it up uh saying like this is the future of Chad even though you got to be on the wait list but you know some people were able to jump the wait list ahead of others I haven't gotten my invitation yet did you get yours have you gotten yours yet Jason yeah yeah no yeah so I'm I'm on the waiting list yeah it's crazy they have this um they have this screen where you can do certain things to get higher in the waiting list yeah you can you know make it your default browser and download the app um so I didn't do all of those things and so I think I'm still in waiting list yeah the first thing it asked me was like Hey do you wanna well I had to sign in under my Microsoft account and then so the first thing I had to do was actually remember what my password was for the Microsoft account uh and then second was hey make Bing your default browser and I'm like nope done or default search engine I'm like no no thanks yeah and then it was like too many steps and I was like yeah then install the app on your phone I'm like no sorry I'll just I'm willing to wait in the line yeah yeah no but you know honestly I gotta tell you I think this is an incredibly brilliant move by Microsoft I mean um you know so they recently did that 10 billion dollar deal in incredibly Visionary yeah um given that uh you know they can be first to Market truly the First Market with this kind of large language model at the scale of Bing but it was super clever because this is a special moment where they can really maximize the disruption of Google and so you know making you really transition and on board into their product ecosystem yeah as and using the carrot of getting to use this next gen Tech is is just brilliant and it's impressive it definitely freaked out Google right because Google you know they said well you've got we've got Ai and they take it out of their back pocket like hey we've got this too and they call it Bard which again I I really want to ask someone at Google why they came up with that name um yeah yeah yeah because I just think of d and d and it's the in D it's the one character class you never want to play um because you just think of someone you know strumming a loot and and yeah singing a song you know yeah I mean I spent some time at Google man the culture is um you know so Bart probably was was was conjured up by an engineer right and it when it comes to the marketing yeah efficacy and all those things like the processes like Engineers really they have a culture of Engineers really making a lot of these calls and so yeah yeah and they could have probably done something better than Bard so when they yeah so when they launched that and then you know the first thing like they they did was they showed an ad of of how it worked and it got the the the the how the telescope wrong the James Webb Space Telescope there was a fact on there that the that their chat AI thing could could and it was wrong and so you know the stock market reacted and they lost 100 billion dollars of value and there was a story this week in RS Technica about how Google employees are really mad at the CEO for kind of rushing it and being very unusual like so it's like yeah I mean they're definitely nervous yeah no and and for good reason man I mean the thing is you know conversational AI has been in consciousness of the market since Siri right Siri was the first time it was teleported um into everyone's expectations and and everybody has been disappointed and under impressed or unimpressed for a long time Alexa oh you set a timer okay that doesn't seem very intelligent um and then you know for the first moment we had a conversational AI that captivated um the imaginations of everyone I mean uh you know it it was preceded by GPT 3 which was the big technological realization of what was possible when you trained massive models at the tune of you know five million dollars plus uh and then chat TPT was the first realization of the thing that everyone it blew everyone away right and so the the wild thing about it though and I really feel for Google is the key technology that underlies the achad GPT was actually invented at Google uh oh okay wow they brought to Market this you know researchers at Google came up with this Transformer neural network uh which was the kind of neural network that makes this technology possible and they kind of had it in-house and for many many years they published papers on it they celebrated it as an AI contribution from an intellectual realm but they didn't really move fast enough to cultivate it to the point where it's able to create such a compelling experience openai did that with the technology that Google innovated yeah and so it's uh so and you know with this amount of um uh widespread uh appetite creation uh uh and the realization of the appetite you Google had to move very fast especially when you hear these billion dollar deals coming from Microsoft to to get exclusive rights to productize these Technologies Google knows what this could mean right for search as a technology do you think do you think that companies like Microsoft and Google are the right sort of companies to bring us to that next level of conversational AI or yeah do we need more startups and those types of companies to to Really push the envelope I mean because these are two huge missteps for Google and Microsoft we didn't even get into the whole uh chat conversation that he that they had with the New York Times guy yet yeah exactly we've totally got it we've got a chat about but you know when it comes to the possibility that is Chad GPT that's been realized is something that is incredibly expensive to hone and cultivate indeed a lot of folks call it space flight type projects right where you invest lots of money to train a neural network one time you train it on the the internet basically it's you you crawl the internet you give this model the internet you train it on thousands of computers millions of dollars of compute and so in a lot of ways the creation and cultivation of this technology is not democratized it's it's not the case that a small startup or a a not well-funded outfit okay can actually do the training to get these Technologies to the point where they are so we we can only rely on um uh you know large companies and also incredibly well-funded companies like open AI was a startup but Elon Musk started them off with a billion dollars you know what I mean uh and so they had a billion dollars to go play I mean the Mandate of open AI was just do research for the world and show us what's possible and so they use that money uh to do this kind of large-scale high effort model training that gave us a chat GPT and so uh uh right now it would be somewhat infeasible for a small startup uh to uh to to try something different and train it at the scale of of of the models that okay so so do you feel like then that these missteps were are just like small little blips and they'll get their act together and then kind of push us to the next level yeah so these missteps I mean it's it's very it's I mean it's it's not surprising uh to me right because essentially what these models are doing um they're not thinking per se but what they do the way they're trained is by generating a response word by word and every time they generate a new word of that response they look at everything generated before and all of the conversation that happened before right to pick the most likely accurate word based on the consciousness of the internet so it's a writer it's it's writing and it's it's it's writing word by word with statistical probabilities determining what word is happening next uh so it doesn't really know what it's saying it's just so good at at continuing the pattern of what should be said next word by word based on the consciousness of the internet that it generates that word now there's a number of things you need to do to get it to to guard rail it from saying weird things or getting Loops or get tricked up and and and and that's uh it's it's almost an art they call Prompt engineering but it really means you you almost have to guide it with with some Preamble right and someone just happens in the background when you ask a question there's there's some systems that might produce texts to start it off like starter text uh and then you can constrain its output so you could really have to put these tight guard rails and I think what's happening with the missteps is the model is uh is a part of the solution that creates that end user experience and another part is is how do you guard rail it and it's almost certainly the case that the chat GPT that you would get from openai is has different guard rails and and uh you know prompt engineering Solutions yeah uh that can curtail how far off to the deep end it can go and Microsoft moving so quickly I could imagine that there was a lot of Engineers that had to push things out very quickly to capitalize on this moment uh and the the the process of improving those guard rails for a model like child gbt is something that we'll see over time but the speed to Market is really why we're seeing these misstops it it's not the case that the the AI necessarily is is is is is is broken or weak it's just not being steered with this piece of the system uh that generates the prompts to start it off right now were you surprised by how quickly sort of this backlash has has happened uh not at all really because because it does feel like in a normal hype cycle that you do stay at hype for a while before you start getting disillusioned it felt like it was like one week it was up and then then this week it's it stinks and and you know it's amazing how quickly the negatives have ramped up so but you're not surprised about this how come no not at all actually I mean it's totally expected I'm surprised I'm surprised that the good news was around for so long and that's because you know it's cool right because I think we got the most authentic reaction from the media when they first reacted because it was so mind-blowing they just have to tell the world right a really true story about how mind-blowing this is but just given the nature of what gets clicks like the bad news the good news could only interest people so long you know you can't have an article like oh I'm still using chat tpd and it's still amazing it's like no no we've got to have a story we've got to have some drama and so it's a it's a part of the the the the kind of the the market of of readers uh would would almost call for well tell me tell me the scandalous stuff or tell me show me like like it breaking right uh and then and that's a good surprise from from the media side of things it's a good thing that the media is sort of like it's it's sort of like they're testing you know they're poking it and seeing what what pops out and so early stories about well there's still there's still some bias issues right there's still some you know it's not perfect and even the the CEO of openai is backpedaling a little bit I think he was quoted as saying like yeah it's not a great great product yet like it's you know so he's trying to like temper the the enthusiasm exactly but the media and I think it's a good thing for the media to keep poking at it because otherwise like that's the best way to get it to improve right yeah and it's it's also incredibly important that we call we we we call upon the public to be thoughtful I'm not not over you know I'd be thoughtful about this technology understand that you can't you can't rely on it as a truth speaker right because because if you poke at it in certain ways it'll tell all you nonsense or it'll tell you untruths it's it's it's important to have the the media able to communicate that to the masses in a more effective way than say a CEO just saying hey guys you know slow down like but we need to really really know so that the public doesn't uh misuse the technology right go to the technology for truth to Art solve all arguments to know what directions to to take real policies right um there's a tendency to to to to be too optimistic about these models and think it has the answers to our heart problems and that would be a a mistake well I'll give you an example of something that I did this week um I was looking to bake some chicken in my oven uh and we got the chicken at the at the meat counter but we didn't I couldn't remember like what temperature and how long to cook it was it was panko chicken um normally so if you look this up in Google like how to you know what temperature do I need to cook this panko chicken and you get a list of like 30 000 recipes and I'm like I don't want recipes I just want the answer I just want and exactly or the answer that they gave me was just based on one recipe that they found or the the highest ranking recipe that they had whereas then I went to chapter GPT it said oh 425 for 20 to 25 minutes and then it gave me a list of other things like make sure you don't overcook it make sure you don't undercook it all of these things that I know as a as a human I do know already exactly but it did sort of give me truth and again I I guess I believed it um because the chicken came out fine and I didn't get sick and it wasn't dry so yeah yeah right and and for and from and for many many uh uh common things I mean one of the ways you can kind of use uh a heuristic to know when when when something like Chad GPT is likely to succeed is how prevalent is this information and the consciousness of the internet now the the power of a chat GPT and the reason it's going to disrupt search for ever like searches really old Tech at this moment yeah forward but the reason is is it in search you're searching for a web page that might have your answers your question might span multiple Pages that's incredibly laborious and painful and you have to integrate you have to integrate uh uh your understanding from those disparate pages and search results rankings right the power of this Tech is that it compresses that all of that knowledge and integrates it into a single Consciousness that you can you can conversationally engage so you get precisely your information right away and when you should be more confident in what it's telling you is when it's information you would expect the internet to have a lot of knowledgeable right so so like the do's and don'ts of cooking a particular uh a dish uh if it's you know uh curry chicken right uh you could expect that oh in the consciousness of the internet dpd probably learned really well this information but if you ask about a really exotic dish where there might be one page on the entire internet about then you'd have to be a little bit concerned because it might use things integrated from something similar that's actually wrong for that specific thing and I think that that was the case in some of the people that were asking about like where does you know how old is someone and and the the Consciousness thought it you know this person was this old old when they were in fact a different age because that that incorrect information had been put into sort of the the data set right right and if there's names if if on the internet the there's a similar name uh that's that many different people have it may integrate information from different sources right so if you ask something about William Burr something precise about William Burr and maybe maybe the consciousness of the internet knows it about the million William Burr you're likely to mean yeah but maybe in generating that response it throws in something about a less known William Burr that that isn't relevant and so understanding the the that that's an issue uh is is really important when you engage these Technologies for truth yeah I remember I remember when Google came out everyone almost started Googling Googling their name to see what information that Google had on it and with my name I can't there was like seven different Keith Shaws that do all sorts of different things including this guy in England who flies Model helicopter but I haven't thought about this yet I should probably uh go on to Chachi PT and see if it can if it knows me or not yet I don't know I mean yeah yeah you know nothing yeah have you done that with your name oh yeah yeah he'll totally done it actually and I've tried to tried it with a couple names I was very um I was very honored that uh Chad GPT actually knows a good yeah it kind of knows things about me and yeah which was awesome but but um uh and then some names not so much um but you know it's it's very uh you know it's very important uh that you knowing that it could sprinkle things in um from other uh you know sources is is very important as one of these Technologies you know yeah because the use cases are are wild right like there's a crisis going on right now in education at all levels in our University yeah uh at University of Michigan there's special sessions where um you know the honor Council wants to understand how to reason about this technology being a part of our ecosystem now like students are using it to to to generate essays and to uh generate reports right and it works very well with popular topics and um and so as you use these things you having a human in the loop that at least edits it I think is a must yeah you know what I mean what's your what's your opinion on that because I've now talked to two different um uh I've talked to a professor who uses chat GPT in his classes and then I've talked to another it administrator at another University and they've got professors that are trying to like prevent the use of it because they don't want people cheating like what's what's the opinion like what's your opinion and then is it does it match of what you know University of Michigan is doing yeah no well you know so I tend to I tend to uh have the opinion that may be less intuitive like so I I actually believe that we should be instructing our students to to use chat GPT to solve problems and then providing the oversight uh of how to solve those problems I mean we have this inner this issue before right when when the internet came to the fore and we had good search engines there was a time you actually had to go to libraries to do book reports right like you'd have to go do research real research to find answers uh because there wasn't a search uh and it would be Folly for a professor you know post having the search era to say okay please don't use Google I want you to go to a library check out seven books read those books and get the answers it's you're trying to resist the evolution of our times and we need to we need to as a society understand how we in a healthy and and forward-looking way how we integrate these new technologies and change our pedagogy change how we teach change the expectations of what the students must do and what the students can rely on technology to do because in the real world that's what they're going to be doing yeah uh preparing them for that is really the mission so so I I really believe in an integrative uh approach um and and adapting ourselves around the new Norms uh rather than try to depress like suppress the nor the new Norms yeah to keep our old ways you know you know and I think you answered this before but I want to just ask it again like why is there such a rush to integrate uh the generative AI into search it didn't feel like initially that like you know generating content is different from search like so why did you see them like both Google and Microsoft quickly because I think Microsoft also wants to put it into their office apps as well and you know the whole corporate app area um you know is something that they want to do is it just because they will they see the value of it they want to generate some profits out of it or yeah is there is there a natural sort of meshing between the two technologies yeah no well so the the this is one of those cases where it's it's a perfect technology because it increases convenience and it increases productivity the day that chat GPT made it open uh open AI made it open I got on right okay yeah and I was playing I just played with it and then from that day on I have it up as I'm working as I'm coding because I do a lot of uh development of software um I I always have it up and I go there instead of going to stack Overflow or going to Google to get my answers so just because I can precisely ask exactly what I need and get it right away it saves me instead of spending 15 minutes researching how to solve a problem I literally spend 15 seconds getting exactly the answer I need so it's a spike in productivity and a lot of people had this reaction when they tinkered with it of course the the leaders and the executives of these companies know that the whole purpose of the internet and search is to help people get access to the information they need and this is such so good at that and it's so much fundamentally better for let's say 70 I'm just pulling that number out seventy percent of the use case of Google it's so much better at that yeah that it's the biggest opportunity uh uh in the in the world Google is on like the number five company in our market economy our global economy because of search so an opportunity to capitalize on one of the most important and profitable Technologies and it's it's a technology that gives you access to information uh to have that in front of you means you need to capitalize on it yesterday right so so everyone's rushing to to get all of the society to use their product because of the lock-in effect right if Bing gets a lot of people to start with set Bing up as their default because Chad GPT is there those folks are gonna be locked in even if Google over three years figures it out yeah Google is going to become the Bing right because Bing figured it out late nobody was using Bing everybody was on Google yeah and so so this is why there's so much effort to get there you know as soon as possible I'm wondering if if this means that maybe ask Jeeves will come back or Alta Vista like you know exactly why they have to be big like yeah exactly well yeah I know right that but if they yeah well you see the thing is ask you if it's a great idea they just didn't have the technology to do it well enough yeah you know but yeah uh all right and so you know that so the next question then becomes like will we start seeing these things and you know we start seeing this technology and things like robots and other physical devices I mean obviously it's going to be on your phone and that's a physical device but what I really want to say I've covered robotics for a few years too uh I would love to see sort of a mobile humanoid robot I go into a store and I and I ask it a question and it tells me exactly where I need or it brings me to the the place where I want to get the thing um in the store that you know that's just one example but you know customer service robots so far have been horrible it would be great to actually integrate this in exactly like do you is it just that we're not there yet and we will get there or is or is there just not a business opportunity well so here here's the interesting thing right and it is a nuanced answer right so and by the way that use case is is perfect I go to during like uh you know Black Friday I'll go to Best Buy I need to get the Nintendo switch like why do I have to wait for that human to explain the TV to someone in front of me before I can just know where if there's any in stock right yeah yeah like why don't we have that use case well here's the here's the reality now a lot of the technology a lot of the products Technologies where the creators the producers of of society focuses attention usually comes from the market appetite right so now we're going to see a surge in that appetite for these use cases now that the art of the possible is so manifested in everyone's face those kinds of questions are going to be asked not just by us consumers but by the next Founders by Venture capitalists by the the consciousness of the public is now uh focused on the possibilities that this creates so that's going to create an economic energy to produce new products and Creations and advance that so we're going to see a lot of these things get a lot of energy in the coming uh near-term years now another important Nuance to understand is the chat GPT technology itself is is great for being a question answering system that can integrate a lot of data into one knowledge yep right so so there's use cases like how to navigate your Supermarket to find the item on what shelf yep you know kiosk that you might talk to like that use case may not be able to benefit from a chat GPT because the information it needs is very specific to that particular store okay and so you might need to use other Technologies now the good news is the the same neural network technology that's in chat GPT is present in many many other places at different sizes and scales and sizes that you can wield and train to solve specific problems and so there's there's there's going to be many many innovative solutions coming to the fore because it's no secret how chat GPD works for the for the technologists and the intellectuals and we can replicate now uh some of the methodology that made it so smart in smaller models for specific use cases and that's you know an area that I'm actually spending a lot of time on now yeah on myself do you do you think that then this becomes something where it gets maybe integrated into an Amazon Alexa or you know the Google home sort of voice assistant uh Siri and Cortana as opposed to search you know because yeah you know instead of like you mentioned Alexa before you're like hey set a timer it's like that's what I use that's what I use it for mainly and to listen to music um but exactly it's like but imagine having an actual conversation with you know we actually get closer to that Jarvis world right from Iron Man or he's running everything or the the AI is doing all these different things do you see that yeah that's probably going to happen before there's pulled up teams working on it right now like the the one of the key lessons that came from gbt3 and chat GPT now is that and people are talking people you know it's a generative AI right so these are it's a it's a special approach to how you can build a system it's that that generate a word by a word a word by word and produce this answer they call it in an auto regressive way it means you look at what you said so far to predict what you say next that particular technique is not how Alexa's uh all of these conversational AIS have really worked um they're usually these dialogue systems with something they call intense and you extract slots so you'll kind of figure out well what are the parameters I need to hit an API with in what you said it was a different approach that really that really create these Technologies were created with and it's why it's part of the reason they kind of sucked is because everybody was thinking in that approach and all the products you have pre this this revolution that's happening right now we're kind of built in that way but now people have seen proof points as to what's possible in the gender generative way yeah to create a dialogue system which which isn't it's not obvious right just looking at how the models work that you could create a dialogue system with just this generative model so so now that work is already started everyone's aligning I've been interacting with a lot of folks everyone is aligning what they investigate next and around well what can these generative models do in many different cases and so the next version of many of these products are going to use generative model Technologies I'm confident yeah now we've talked obviously you're optimistic about the about this technology I can tell because you're you know you're you're very excited about all of this stuff but on the on the flip side we've also seen some sort of dark or not so you know good purposes for it you hear about that you know people are using it to generate malware or write code or or provide a better email sentence for uh phishing attempts uh because again with with phishing you could always know oh this is not real the it's not even correct English and now and now that's all out of here now I got to be suspicious of everything that I get and not just yeah you know my filter now has to be tuned as well so like what do you feel about some of the bad uses is it is it again find the bad stuff so that we can fix it or generate tools to spot the bad stuff yeah yeah well you know it's always a tools arm race arms race when you have the you know there's no various folks out there that will take every possible technology and Tool and see how it can make them more effective and efficient yeah uh at phishing attacks hacking every you know yeah it's used everywhere but uh but and also it's just not even beyond that when it comes to fake news like it's incredibly you can you can give it a a headline to write a story about and it'll write you a compelling story with that headlining with that headline in it oh you'll buy them just yeah I don't know just make up your something whoever and then you'll have a command yeah right so there's all these kinds of things that we need to now be educated and aware of as to how nefarious use cases could work and the good news is actually chat TPT the the motivation for the technological advancement of chat DPT was to develop Technologies to help when it comes to the safety of how these models are are being used I believe and I haven't valid verified this but I believe it's almost an accident and the CEO kind of indicated this it's almost an accident as to how us trying to develop Technologies to make it smarter so it could be safe turned it into being smarter in in many ways and the very fascinating thing that's going on right now is there's a lot of energy uh being invested in how do we design AI models that can can control or detect other models right you can actually use models you can train models that can find nefarious use cases that can detect when an AI wrote this and not a human right uh and so we've got an interesting uh you know you can build tools to make the model safer that are other AIS and that's happening all the time it's happening right so there'll be some sort of like signature form or something that says like oh this was created by an AI right like that's a good way to to at least start right exactly exactly or leveraging an AI to say does was this written by an AI and the AI is like yes this was I think this was written by an AI with a 98 right um I think you know you can imagine then you'll have AIS talking to other AIS and oh yeah yeah yeah that's that's happening now actually and it used to be kind of a fun thing like it was almost like a fun thought experiment but five ten years ago like oh should we have ai's interacting with each other and hanging out and just learning from each other but now we're actually uh uh leveraging uh techniques to to do this right where AI is like are adversaries to produce better results uh or you know AIS can kind of check other ai's work and so forth it's happening man it's it's wild yeah so it you know is is some of the backlash that we're experiencing now where you know you've got you're seeing bad answers you're seeing biases and answers and you know there was another there was another story that that came up this week too where um a bunch of uh tick tockers uh went to the McDonald's drive-through where they were experimenting with AI and if you haven't seen this yet go go on to this article on in the show notes page um and it's these people that are trying to just order an ice cream cup and the AI just keeps going oh would you like ketchup or I put ketchup in your order or here's some sugar sugar cubes because they could they couldn't understand what the person was asking and you know it's it's hilarious because obviously it just shows you the failures of the systems um but on the other hand you know is Backlash like this going to cause people to go ah AI you know yeah you're gonna go oh that's just a fad and and then and then yeah or you know or is it just sort of a little speed bump in terms of you know yeah no no so this is the this yeah go ahead well as I say short attention spans being what they are these days like you really have a limited amount of time to sort of you know go through with something to get to that next level of acceptance yeah absolutely absolutely no it's it's a very um so the initial AI has been the there's been multiple waves of hype Cycles around composition like I is including Syria and Alexa yep there are these waves and then there's these troughs of of uh disappointment yeah right and despair right like they're like ah it's terrible I can't use this it's more give me a button I can push I want to use a touch interface I don't want to talk to this thing so when you look at McDonald's I'm actually very familiar with McDonald's story okay uh I I actually uh uh was uh kind of in the Runnings at one point to collaborate with McDonald's to uh produce uh their conversation land for drive through um but they did an acquisition and I'm also aware of the company they acquired to to help uh create that experience that experience is using the same old Tech it's using the Alexa set a timer Tech right which is kind of this rules-based thing that's why it does these silly things where you asked for an ice cream why is it asking for ketchup it's because there's some if statement rule or some intent thing that will just take you to catch-up land but but you wouldn't see something like that yeah like you wouldn't seem something like that if a generative approach was used with enough data like if you had enough orders let's say uh and you saw the interactions uh uh and you were able to train a generative model uh you wouldn't end up in a place where it say something so unlikely right or so out of out of like whack right and so and so what I would hope is all of the old Tech that is not good uh and all the experiences that are terrible like folks will just quickly understand like from a be knowledgeable enough to know oh yeah that was that old you know thing um and then they can recognize when something's good and it's it's part of the new phase yeah of conversational AI technology does that mean that like if you're if you're gonna develop this for a drive-through scenario that that you're gonna have to record all of these order calls and and then you sort of have to tell your customers that you're recording right like yeah Yeah Yeah well yeah you know I actually think I don't uh so there's some I think some of them are recorded uh like there are these data sets where they've recorded um uh the the orders and so forth yeah so the data is there it's the methodology as to how you use the data dude guys like one thing I want to make sure a lot of folks are aware the the core Tech has been there for a while yeah the The Innovation is on the methodology on how you train this Tech right um the Transformers is old news but how you train it and how you systematize it uh is what's really a lot of Innovations are are driving uh this forward so that's the open question right so I think it's still uh there's still some work and I'm talking about months worth of research work or in or innovation right lab work to to take these Technologies and come up with the right methodology to get that step function better experience yeah the chat TPT grade kind of experience yeah an interesting part of the the the drive through stuff too was probably you know you know the drive-through even with two humans the audio quality you get from the speaker to the half the time I can't understand the the person that's asking me you know what I want to order uh or you know or if they're ready for or not for the order and then you know and then on the other end you're taking the order you sometimes can't hear what it is I mean that's been mocked all the time yeah um and then I'm sorry I'm starting to think that maybe um in that algorithm that they have they're like if if confused just add ketchup like yeah exactly exactly that's confusion that's gonna mean that's gonna be my new band name is just add ketchup um yeah yeah yeah yeah just uh it's it's really hilarious man but yeah have you looked at the the artwork stuff uh you know like the generators oh yeah yeah like that's what also impressed me too um last year I saw some of the early Dolly stuff and and the images that it was generating was were completely horrible um but then at the end of the year I did it again with dolly two I think it is and the images are like wow it's actually getting closer and closer although it still can't generate good text I don't know if you've tried it lately yeah yeah yeah yeah yeah yeah and it's a it's what are you I'm just saying what are your thoughts about sort of the the generating Visual Arts because there's copyright issues and you know there's a lot of public policy issues and it's also there's also a lot of dangers of like um you know creating fake things that people think are real um but but you know the thing is there's this there's this there's this critical point when when a technology changes the world it's when it creates an uh an environment where now there's a clear thing that's more convenient or it makes people fundamentally more productive right and so Technologies like Dolly and there's a lot of these technologies that are fascinating and they're proof of of what computers doing things we never thought was possible um I think they're fascinating but you know unless there's the kind of like that that market use case I don't I don't know how much uh I don't know how how much it it'll fundamentally matter until it gets to the point where it now fits into our lives in a way that makes us better in some way but when it comes to a proof of what's possible with neural network Tech uh and large amounts of data and enough money to use large amounts of data to train massive models I mean these models spend hundreds thousands of computers uh in size um you know showing the art at a possible I think is is it's breathtaking to see what Dolly can do and we'll we're going to innovate ways we can we can utilize it to advance our our life experience but but yeah I mean I I'm always shocked with Dolly I use it in my class so in my class at University of Michigan I teach a conversationally i class I actually have one lecture all all around diffusion networks that's that's what really fuels the dolly type Tech yep uh and Dolly and the the possibilities with images um it's it's really game changing right um uh from what we thought was possible yeah yeah so I'm gonna have you put your prediction hat on and and say like where where will we be in about a year like I'm I'm or I mean even six months from now I mean it does seem like this is this is being generated very very quickly so let's just go a year from now uh will we be having these same conversations or you know will the conversations change in in terms of wow there's so much available now and and you're like where do you think we're going to be in terms of that high cycle in about a year well so I mean in a year's time I think we're gonna see uh we're gonna see real fight like Bing right now the Bing experience the child experience in bang there's work to do right so we're gonna see a a polished and more robust uh this technology everywhere the way we do search is going to fundamentally change Forever at that point uh Google is gonna have their offering the Market's gonna be disrupted stocks somebody's gonna survive this and their stock is is going to go up 20 30 uh and so that's what we're gonna see when it comes to making this accessible for changing the way we search and find information yeah and access information but we're also going to see a flurry of new kinds of startups that are solving problems Anew in interesting ways right we're going to see conversational AI being integrated into more use cases to ease our lives and then we're going to see a lot of the uh you know a lot of companies are going to revamp their self-proclaiming as an AI leader so yeah you know I give this talk about the AI Revolution it was kind of cooling off but this hype cycle is wild dude like because it gets bigger and bigger with every advancement and so uh so I think we're gonna see a whole new ecosystem of Technologies and pro uh products uh based around new kinds of conversational AI I can tell you my prediction for one year is that I'll still be on the wait list for for Bing yeah yeah yeah yeah you guys and they're not gonna let me I'm I'm bringing up the rear so to speak so uh Jason thank you so much for being on the show today this is a fascinating conversation I would love to have you back uh we'll just talk AI forever so uh thank you for for being on the show sounds sounds good hey a pleasure being here and looking forward to chatting again soon all right uh that's all the time we've got for today's episode don't forget to like the video subscribe to the 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 foreign
2023-02-23 22:31
