From the MIT GenAI Summit: A Crash Course in Generative AI

professor of computer science at Brown University and also a research scientist at Google um I work on large language models and these other Vision models as well um and I primarily work on trying to understand kind of the how they work under the hood so some of this issues of opening up the black box trying to um see if we can get a more precise understanding and be able to control them better um but I was asked to give a fairly high level talk today to get everyone on the same page about what is Gen AI so I'm sorry if some of you are as technical as maybe Eric had assumed everyone is but if there are some of you who are here because you're kind of just interested to learn like what is this chat gbt I'm going to give a very high level um uh intro to basically what are we talking about what do we mean by generative AI as opposed to maybe the AI that you might have heard about uh even just one year ago um and then again a super high level understanding of just kind of how does it work so what are the main components that um that go into producing a system with this kind of abilities um and I'd be happy to talk about more of the lower level technical stuff on how it works afterward if people are interested um and I'll talk briefly about some of the opportunities although I think that's actually more you guys expertise I'm not an entrepreneur um so that's where like I think your creativity comes in I want to spend more time highlighting kind of what are the major risks so if you're thinking about trying to deploy these in some new kind of setting some new startup um or adding some uh generative AI components to existing company I think it's really important to keep in mind for these kinds of fundamental limitations things we don't yet understand that maybe we will five or ten years out but we don't understand now and so it entails some significant risk I would say okay so what do we mean when we're talking about generative AI I would say kind of my definition is we're talking about AI models but that can produce kind of open-ended and creative content so typically when we thought of AI we think about things like prediction it's giving you like a yes or no this review is positive or negative something like that or very uh constrained set of possible outputs with generative AI we're thinking about things that are open-ended and fundamentally what we would think of as creative so we talked a lot about text I don't know if you can read this yes so this is asking chat GPT to like make an announcement for this Summit and it actually does a I think quite a nice job all right so I'm sorry for whoever did the marketing for this Summit but I guess you could have you could have watched Netflix and let chat gbt do that work for you um we're also talking about image generation so this is a really popular use case where um so here I was I actually tried to spend I spent a bit of time trying to make it make a logo for this Summit and it did quite a bad job they were like really generic and things like that so there's some jobs that are safe but something like this a robot painting at an easel um it's able to produce pretty nice images of these kinds of things and it'll provide quite a bit of variety and you can kind of refine it there's also things like music generation so it's like text and images are really like the the big ones right now that people are talking a lot about but there's also applications to things like music um I was going to click um and play some of these but I think in time I won't so um but I would encourage you to Google this it does quite a nice job um generating text or music from a description um and other one that I didn't have an image of but code generation is a really huge one right so like uh we were just heard in the panel you think of a code as a language and these models are really impressive but I think that's something that surprised a lot of people is how good they are at generating codes so things like software engineering a lot of those applications could be um potentially automated okay so how does this work um there's really kind of three components to think about with all focus on language models on generative language models um because the basic principles are similar for image and music and code um although with some details different um so it's kind of like three maybe like buzzwords to have in mind the first is that they're neural network models so we've been hearing about deep learning we've been hearing about neural networks a lot um and the neural networks are very old technology actually um if you are familiar with neural networks great if not um this is kind of the picture you have in mind when delphine's talking about being able to write them down on paper um you can still write down them on paper you'll just need a lot of paper right but like the actual principles are still there so if we're thinking about something like a neural network language model what it's doing is learning associations between words right so all it's doing is learning for example when I see a word like generative what types of words tend to come next and it's going to learn some kind of distribution over this so generative is likely to be followed by AI right now maybe also design in Psychology apparently generative design and generative psychology are things like Googled it yesterday um and it might learn a different set of associations for different types of words right um so when we take this kind of basic model the like the key part I think what's been doing really the heavy lifting and generative AI in particular language models is what we call the language modeling task which is just this task of predicting the next word in a sentence for a very very large amount of text so you know literally what the model is asked to do is you're just like hey model it's had no training yet you say I'm about to write a sentence what do you think the first word is and it's just going to take a guess right it's gonna be like the and then you're like nope it was I right and it like makes some updates to its weights so that next time it's slightly more likely to say I instead of the or it puts them at equal mass or something and then you say okay you said I what comes next maybe it does a good job here I guess it's something like am because that's likely to follow I and you do this repeatedly repeatedly and every time it gets a new word it's making some updates and it's remembering something about the distribution of words that it's seen um in text right and so this means that when you basically take one of these language models you can just start writing and then at some point tell it to take over and it'll do a pretty good job of just coming up with a plausible continuation of text um and this seems very simple and it is actually incredibly simple like there's a lot of kind of you know hardware and software things that happen under the hood to make it happen but the principle is incredibly simple I think that's something that surprised a lot of us in the field um and uh what really gives it power is this scale so we talked about increasing complexity that complexity isn't actually a change in the underlying model it's just making it bigger and by bigger we mean a lot more of these neurons so it's learning a lot more of associations between words and groups of words and groups of groups of words um and it's uh reading a lot more data right so there's different ways to make it more complicated but this is kind of giving you a sense so I would say uh generated like I would put the Elmo model spring of 2018 as like the first one of the kind of modern group of generative AI models or these large language models um and at the time this was model was considered to be huge and it was so exciting like I remember reading the paper and everyone was like so excited and it's just taken off literally exponential growth in the size of these models and with it this these kinds of impressive capabilities um there's another component that comes up and I don't want to go into a ton of details but you're probably hear about it or we'll hear about it more um in particular because it's been associated with chat gbt and other kinds of models and this is a reinforcement learning component so basically a model like this is just generating plausible sequences of text on the internet which doesn't mean it's necessarily good at doing useful tasks so if I were to ask it a question for example if I just go to a large language model train this way and I say like um what is a you know who are good speakers to invite to my Summit it might give me an answer like good speakers but that's actually not the kind of document that exists on the internet more likely it's going to give you like a list of questions you should be asking yourself when you're planning a summit so if I say who are some good speakers I should invite it should say where should I host how many people should I invite what platform should I use like just a list of questions right which isn't what you were going for so like this generative component by itself isn't particularly useful at tasks in and of itself unless your task is to generate documents so instead we do this reinforcement learning component where basically it generates lots of examples and people give up or down votes so this is important for a couple reasons um it's important for getting the model to basically do the task you actually want it to do um it's important for being able to give it negative feedback so tell it examples of things that are bad that you don't want it to do so this has been a main mechanism used to for example prevent the model from saying um harmful things from giving instructions on how to you know build a bomb or something that uh for example companies like open AI don't want the model to produce right so that's really the reinforcement learned component and if you're thinking about something like building um one of these models from scratch or doing some customization for applications likely the reinforcement learning component is a place where that would happen um I think it's also important to highlight because it's also associated with a lot of the bad behavior so once we start running reinforcement learning there's often a lot more uncertainty in how the models will behaves it makes can make them much more creative it can also make them higher risk so it's just a thing to keep in mind all right so my kind of um three my three uh sentence summary or three-part summary of what are something like generative language models would be their neural networks they're first trained on this predicting the next word and then they're kind of optimized for your task through reinforcement all right so like I said um we can talk kind of briefly about major opportunities um but I would actually say that's more you guys's expertise so I'll just throw out some of the things I've heard that I think are exciting to think about um I like this analogy I forget where I heard it so I don't know who to credit for this um but this idea of thinking as kind of like calculators but for open-ended content right so in the way that you might have um like people who previously spent a lot of time you know like human computers back in the 50s you know just doing actual calculations and that uses a lot of like uh mental energy from very intelligent people on somewhat routine and formulaic tests once you have calculators and computers to do those computations for them their time is freed up to work on more challenging more creative types of things so now if you can imagine replacing some of the kind of more mundane parts of creative tasks so um then you can free up your your people's time to work on things that can't be automated um so the kinds of things that I hear a lot of uh people throwing out as kind of possible um things that are ripe for for this kind of automation or like components of customer service um a lot of software engineering and code generation right like you wouldn't want all of your code written by a model but there's a lot of really for people who are software Engineers there's a lot of very routine stuff that time is wasted on um anything that requires writing text so reporting and Drafting and um writing outlines and making slides like these kinds of things um design and illustration and much more things um but like I said I want to focus primarily on the risk because I think this is something that's really important to be thinking about and especially when we get all swept up in the optimism right like we don't want to be doomstayers about the the huge risk of them but we also want to be realistic about um where we are in this kind of development of this technology and what kinds of changes we're likely to see um in the coming years that could entail some just realistic risks and also some Financial risks as you invest heavily in the current technology and things change a lot right um so one of the the primary things that um that was mentioned already is their ability to hallucinate and this can be possibly a strength maybe if you're working on protein synthesis that's where the creativity comes from um but it can also be somewhat of a concern right so for example what I mean by hallucinate is um you can ask the model I asked it to generate a speaker buyer speaker bio for me um it actually did a better job than I expected so it knows like where I got my PhD and where I work and things like that um probably just lifted it off my web page but it also like gave me a ton of awards that I've never actually received so I'm very flattered chat GPT I would love to have only with these grants and maybe it knows something I don't know and I should be uh here's what the faith comes in but I'm quite sure it just filled in the blanks by thinking of many of my more heavily awarded peers who I am similar to on other dimensions and said these seem like plausible things to put at this point in a bio right and what I hope so let us hope so yeah um and so this is the kind of thing where really it's quite application dependent right so this ability to kind of Imagine or fill in the blanks based on what it does no can be a really really powerful tool but in other applications entails very high risk so if you think about this in something like a search setting um you definitely don't want uh uh you definitely don't want models saying things that sound quite plausible sound quite true and are actually ungrounded in reality um the verification aspect of this might be as labor-intensive as having done the search yourself right so having you can think about automated processes people are thinking about automated processes for trying to fact check um but you're now kind of building hallucinating AI on top of hallucinating AI so it's not it's not a solved problem it's a thing people are actively working on but it's definitely not a solved problem um I would say this is also a thing that you can imagine kind of legal consequences if you have it doing things like customer service relations or um you know share shareholder briefings or something where it might say things that are just false and there could be big risks associated with them um an article I'd really recommend kind of on this issue that I thought was nicely written um it was this um from The New Yorker that kind of Likens models like chat GPT to like a blurry jpeg of the web right so you can think of it as this low resolution compressed version of all the information that's out there so you can imagine when you're thinking about what kinds of applications is it appropriate for you have to make sure it's applications that you're okay with a lossy compression and then a reconstruction that might not be faithful to the original right and there are some cases where that's fine and there's some cases where it's not um another really prominent and kind of hopefully a concern that people are well aware of um is that Ai and machine learning models are highly biased um and so again depending on the applications we're using it for this could be a really uh really significant risk I think it's important to emphasize because there's been good work and like a lot of awareness of this issue for many many years there's like the weapons of mass destruction book that maybe many people have heard of so like it's been very much in the ethos that statistical models are highly biased and yet we've made very little progress on it so this is a really hard problem like those models have gotten better and better and better and still if you ask this was Dolly II or one of the image generation models for a surgeon you get some white guys if you ask for a secretary you get some kind of suggestive looking women and this isn't the kind of thing you want to have at the heart of an HR System you don't want to have it at the heart of a customer service system um I don't have the slide on it but there is a I saw a recent result that the models also can do some profiling of the person they're talking to and treat them differently so for example more likely to give an uneducated user incorrect information right and that's really not something you would want in a customer facing application and those kinds of concerns get worse as the models get bigger right so the kind of scaling we would do to make the models perform better might also mean that some of these risks get more severe um so a really important one I want to bring up is the the issue of reward hacking um this is specific to this reinforcement learning component but I think is a really really big risk so what I mean by reward hacking is basically you define something you want the model to optimize for you say something like I'm going to have a model that's a customer service agent and it's going to have some chat with a person um and I want to like you know maximize the the customer's rating at the end of that review you know how satisfied with the discussion was your problem solved right but it might find some kind of odd ways of of maximizing that reward which seems like the right thing to measure and if you had humans doing this that's the right thing to measure um when you have models doing it they might find some kind of weird loopholes right so there is actually I'm going to try to pull up this video so there's a nice illustration in this um from these agents that were trained by openai this was several years ago the underlying technology is very similar to what we're using um and these were teams of agents that were supposed to play a hide and seek game so there's like one team trying to hide from the others and the other's trying to find them it's very cute they have like cute little faces and their eyes beam when they find each other and they can like move blocks around and like build walls and things like that um but in the game engine they use there's basically a bug in the physics engine and the models The Seekers quickly find this and exploit this so that they can Ricochet off the walls and fly and like look at The Seekers or look at the hiders from like flying overhead and things like that right um they could exploit it so they could like throw blocks out of the world and make them disappear in a way that weren't supposed to um so this is very much like a cute illustration of something that you can imagine going very badly in the real world so once you imagine that you have a generative AI system that you're giving some even limited access to for example an API because you want it to make be able to take actions on behalf you don't just want a chat bot you might want it to be able to do things like you know cancel and Order when the customer reaches out and says I'd like to cancel this order or even do a very simple thing like send an email in order to confirm a receipt or something so you have some limited API access and unless you're entirely sure that is completely bug free which as far as I know we're never confident of there's a good chance the models will be able to find and exploit those bugs um and maybe you'll catch it but maybe only after some significant damage right um and so actually one of the I think the kind of high level and I'm emphasizing this because like I said I think it uh piggybacks nicely off the panel um the kind of interpretability aspect um is really huge here so we have this very very exciting technology but it is uh the kind of the excitement is there on the the scientific side too like we're a bit surprised at how good it is and we don't actually understand how it's so good um so I think Sydney was a really good illustration of that kind of thing like people spent a lot of time smart people worked on that system you put it out in the wild and it starts doing things that we didn't expect it to do um and so I I think a really good analogy I I hate to make analogies between these models and the human brain because there are of course many differences um but from like a scientific understanding I think that's a nice analogy which is basically like we have no idea how the brain works right like we don't know how like what is the software people are running to get from inputs to outputs um and so we don't feel super comfortable saying that we could precisely control a person right like we don't actually know what's happening under the hood and so the kind of level of understanding we have about these models is somewhat analogous like the way we wouldn't say we understand brains because they're just neurons right like that's a deeply unsatisfying explanation for how we get this Advanced Behavior that's similar that's kind of the level of an understanding we have about these models um so as we're um as we're trying to deploy them like we kind of want to proceed with that level of caution I'm I'm quite an optimist I think that in the next five definitely 10 years we're gonna have a really deep understanding of how they work right but that kind of understanding might really influence how we want to use them how we don't want to use them what kinds of safeguards are in place what kinds of regulations are in place um so I think that's just really what I want to emphasize and kind of the technical side um that I think it's a very very exciting opportunity it's also moving incredibly fast and there's like a very good chance that three years from now we're going to look back on the stuff we were saying now and be like wow were we wrong about XYZ right because that's kind of how this field Works um so I think it's it's not too soon to be thinking about business opportunities for this and we should actually be total completely exploring them um but there's that kind of healthy caution to have in there right as we're thinking about this and just being prepared for those kinds of risks being prepared to be super flexible and mobile as our understanding advances along with our the technology itself and I am exactly at 20 minutes so I will stop there [Applause]

Channel: MIT AI ML Club

Views: 103,626

Keywords: GenAI, Generative AI, AI, Artificial Intelligence

Id: f5Cm68GzEDE

Length: 20min 7sec (1207 seconds)

Published: Tue Mar 14 2023