The Turing Lectures: What is generative AI?

hello hi everyone hope you're well thank you for giving up your Friday evening to be here that was a really slick video and I haven't actually seen it with the sound before which makes it even Slicker which is awesome my name is hurry like when you go somewhere quickly you're in a hurry this is the loudest uh the largest group that I've ever been able to make that gag for um and apologies if you're from the touring you've heard that probably thousands of times already um really excited to be hosting the first of this amazing series of touring lectures on generative AI for you today now quick show of hands who's been to a touring lecture before and who's here for the first time awesome no shame this is my first time as well Scandal don't tell anyone that um so these are awesome sets of lectures put on by the touring Institute there are Flagship lecture series is the official term running since 2016 which was when I was like in second year of uni which is wild um featuring World leading speakers from around the world um on data science and Ai and obviously no pressure to our speaker today with that title um as well as having like an amazing lecture we want you guys to be part of the experience as well so we've got a really long Q&A session um after the talk for those of you here in person you can just raise your hand and we've got like roaming mics will come around the room if you're online um you can ask questions in the Vimeo chat and Alana is somewhere or will be somewhere with an iPad and Fielding those questions as well and so please do think about what you'd like to learn and um what you'd like to ask our speaker in terms of this series we are investigating the broad question of how AI broke the internet with a focus on generative AI now I realize I might be a bit of a buzzword we're going to hear a lot about it this should be a very accessible lecture um but examples of generative AI include things like chat GPT Dary if you guys have heard of those those things um potentially used chat GPT to write some emails or essays for your degrees or blog posts all those kind of things I see you we've all done it no shame in it um and pro probably seen in the news and the media all like the potentially good beautiful wonderful things that can happen with this technology but also the potentially bad and scary and worrying things as well and this series is kind of focused on telling you a bit more about these Technologies and trying to add a bit of balance to what is quite noisy conversation so to start this off tonight we're exploring what these Technologies are how they're made and what's going on behind the screen which has blown my mind as I've learned more about it and will blow your guys Minds as well to take us on this journey I'm really excited to introduce to you this evening speaker she is a professor of natural language processing or NLP if you've heard that term before in the school of informatics at the University of edra her research focuses on getting computers to understand reason with and generate natural language you can think of natural language as just like any languages that humans use and now for the sort of accolades list that I can only ever aspire to have at some point in my future she is the first recipient of the and hold your breath for this one British computer society and information retrieval Specialist Group Karen spark Jones award the BC s i RS g k SJ award for all of you counting at home and a fellow of the Royal Society I know right I really rehearsed that one as well um and a fellow of the Royal Society of Edinburgh the ACL and the Academia europeia which I've checked the pronunciation of I think that's right apologize if I got it wrong she is the director of the ukri center for doctoral training and natural langu language processing or NLP holds the Royal Society Wolfson merit award and was one of the first to be appointed cheering AI World leading researcher fellows as part of her Fellowship she aims to deliver an AI system inspired by the human brain that is capable of advanced reasoning and able to draw conclusions from a large and varied sets of data and as someone who is not capable of advanced reasoning or drawing conclusions from large and varied sets of data I find it crazy that machines can do that and people are building them so without further Ado and get into it please can we have a huge round of applause for our speaker tonight Professor Mela leata wow so many of you good okay thank you for that lovely introduction um right so what is generative artificial intelligence um so I'm going to explain what artificial intelligence is and I want this to be a bit interactive so there will be some audience participation um the people here who hold this lecture said to me oh you are very low tech for somebody working on AI I don't have any explosions or any experim so I'm afraid you'll have to participate I hope that's okay all right so what is generative artificial intelligence so the the term is made up by two things artificial intelligence and generative so artificial intelligence is a fancy term for saying we get a computer program to do the job that a human would otherwise do and generative this is a fun bit we are creating new content that the computer has not necessarily seen it has seen parts of it and it's able to synthesize it and give us new things so what would this new content be it could be audio it could be computer code so that it writes a program for us it could be a new image it could be a text like an email or an essay you've heard or video now in this lecture I'm only going to be mostly focusing on text because I do natural language for in and this is what I know about and we'll see how the technology works works and hopefully leaving the lecture you know how like there's a lot of myth around it and it's not you see what it does and it's just a tool okay right so the outline of the talk there's three parts and it's kind of boring this is a Alice mors Earl I do not expect that you know the lady she was an American writer and she writes about memorabilia and Customs but she's famous for her quotes so she's given us this quote here that says yesterday is history tomorrow is a mystery today is a gift and that's why it's called the present it's a very optimistic quote and the lecture is basically the past the present and the future of AI okay so what I want to say right at the front is that generative AI Is Not A New Concept it's been around for a while so how many of you uh have used or are familiar with Google translate can I see a show of hands right who can tell me when Google translate launched for the first time oh that would have been good 2006 so it's been around for 17 years and we've all been using it and this is an example of generative AI Greek text comes in I'm Greek so you know pay some dues to the right so Greek text comes in English text comes out and Google translate has served us very well for all these years and nobody was making a fuss um another example is Siri on the phone again Siri launched 2011 12 years ago and it was a sensation back then it is another example of generative AI we can ask Siri to set alarms and Siri talks back and oh how great it is and then you can ask about your alarms and what not this is generative AI again it's not as sophisticated as chat GPT but it was there and I don't know how many have an iPhone see iPhones are quite popular I don't know why okay so so we are all familiar with that um and of course later on there was Amazon Alexa and so on okay again generative II Is Not A New Concept it is everywhere it is part of your phone uh the completion when you're sending an email or when you're sending a text the phone attempts to complete your sentences attempts to think like like you and it saves you time right because some of the completions are there the same with Google when you're trying to type it tries to guess what your search term is this is an example of language modeling we'll hear a lot about language modeling in this talk so basically we're making predictions of what the continuations are going to be so what I'm telling you is that generative AI is not that new so the question is what is the fuss what happened so in 2023 open AI which is a company in California in fact in San Francisco if you go to San Francisco you can even see the lights at night of their building um it announced GPT 4 and it claimed that it can beat 90% of humans on the SAT for those of you who don't know sat is a standardized Tex the test that American um school children have to take to enter University it's an ad admissions test and it's multiple choice and it's considered not so easy so gp4 can do it they also claimed that it can get top marks in law medical exams um and other exams they have a whole Suite of things that they claim uh well not they claim they show that gp4 can do it okay aside from that it can pass exams we can ask it to do other things so you can ask it to um write text for you for example you can have a prompt this little thing that you see up there it's a prompt it's what the human wants the tool to do for them and a potential prompt could be I'm writing an essay about the use of mobile phones during driving can you give me three arguments in favor this is quite sophisticated if you ask me I'm not sure I can come up with three arguments you can also do and these are real prompts that actually the tool can do um you tell CH GPT or GPD in general act as a JavaScript developer write a program that checks the information on a forum name and email are required but address and age are not so I'm just writing this and the tool will spit out a program and this is the best one create an about me page for a website I like rock climbing Outdoor Sports and I like to program I started my career as a quality engineer in the industry blah blah blah so I give this version of what I want the website to be and it will create it for me so you see we've gone from Google translate and Siri and the auto completion to something which is a lot more sophisticated I can do a lot more things another fun fact so this is a graph that shows um the time it took for chat GPT to reach 100 million users compared with other tools that have been launched in the past and you see our beloved Google translate it took 78 months to reach a 100 million users a long time um Tik Tok took nine months and chpt two So within two months they had 100 million EUR users and these users pay a little bit to use the system so you can do the multiplication and figure out how much money they make um okay so this is the history part so how how do how did we make Chad GPT what is the technology behind this the technology turns out is not extremely new or extremely Innovative or extremely difficult to comprehend um so we'll talk about the today now so we'll address three questions first of all how do you we get from these single purpose systems like Google Translate to chat GPT which is more sophisticated and does a lot more things and in particular what is the core technology behind chpd and what are the risks if there are any and finally I will just show you a little glimpse of the future and how it's going to look like and whether we should be worried or not and uh you know I won't leave you hanging please don't worry okay right so all these GPT model variants and there is a cottage industry out there I'm just using GPT uh as an example because the public knows and there have been a lot of um you know news articles about it but there's other models other variants of models that we use in Academia and they all work on the same principle and this principle is called called language modeling what does language modeling do it assumes we have a sequence of words the context so far and we saw this context in the completion and I have an example here assuming my context is the phrase I want to the language modeling tool will predict what comes next so if I tell you I want to there is several predictions I want to shovel I want to play I want to swim I want to eat and depending on what we choose whether it's shovel or play or swim there is more continuations so uh for a shovel it will be snow for play it can be tennis or video swim doesn't have a continuation and for it it will be lots and fruit now this is a toy example but imagine now that the computer has seen a lot of text and it knows what words follow which other words we used to count these things so I would go I would download a lot of data and I will count I want to shovel how many times does it appear and what are the continuations and we would have counts of these things and all of this has gone out of the window right now and we use neural networks that don't exactly count things but predict learn things in a more sophisticated way and I'll show you in a in a moment how it's done so J GPT and GPT variants are based on this principle of I have some context I will predict what comes next and uh that's the prompt The Prompt that I gave you these things here hold on these are prompts this is the context and then it needs to do the task what would come next in some cases it would be the three arguments in the case of the uh web developer would be a web page okay so the task of language modeling is we have the context and this I changed the example now it says the color of the sky is and we have a neural language model this is um just an algorithm that will predict what is the most likely continuation and likelihood matters um these are all dedicated on actually making guesses about what's going to come next um and that's why sometimes they fail because they predict the most likely answer whereas you want a less likely one but this is how they train they train to come up with what is most likely okay so we don't count these things we try to predict them using this language model so how would you build your own language model this is a recipe this is however everybody does this so step one we need a lot of data we need to collect a ginormous Corpus so these are words and where will we find such a ginormous Corpus I mean we go to the web right and we download the whole of Wikipedia stock overflow Pages quora social media GitHub Reddit whatever you can find out there I mean work out the permissions it has to be legal you download all this Corpus and then what do you do then you have this language model I haven't told you what exactly this language model is there is an example and I haven't told you what the neural network that does the prediction is but assume you have it so you have this Machinery that will do the larning for you and the task now is to predict the next word but how do we do it and this is the genius part we have the sentences in the Corpus we we can remove some of them and we can have the language model predict the sentences we have removed this is dead cheap I just remove things I pretend they're not there and I get the language model to predict them so I will randomly truncate truncate means remove the last part of the input sentence I will calculate with this neural network the probability of the missing word if I get it right I'm good if I'm not right I have to go back and reestimate some things because obviously I made a mistake and I keep going I will adjust and feed back to the model and then I will compare what the model predicted to the ground truth because I've removed the words in the first place so I actually know what the real truth is and we keep going for some months or maybe years no months let's say so it will take some time to do this process because as you can appreciate I have a very large Corpus and I have many cent sentences and I have do the do the prediction and then go back and correct my mistake and so on but in the end the thing will converge and I will get my answer so the tool in the middle that I've shown uh this tool here this language Model A very simple language model looks a bit like this so and maybe the audience has seen these this is a very naive graph um but it helps to illustrate the point of what it does so this neural network language model will have some input which is these um nodes in the as we look at it well my right and your right okay so the notes Here on the right are the input and the nodes at the very left are the output so we will present this neural network with uh five inputs the five circles and we have three outputs the three circles and there is stuff in the middle that I didn't say anything about these are layers these are more nodes that are supposed to be abstractions of my input so the generalize the idea is if I put more layers on top of layers the middle layers will generalize the input and will be able to see patterns that are not there so you have these nodes and the input to the nodes are not exactly words they're vectors so series of numbers but forget that for now so we have some input we have some layers in the middle we have some output and this now has these connections these edges which are the weights this is what the network will learn and these weights are basically numbers and here it's all fully connected so I have very many connections why am I going through this process of actually telling you all of that you will see in a minute so you can work out how big or how small this neural network is depending on the numbers of connections it has so for this toy neural network we have here I have worked out the number of Weights we call them also parameters that this neural network has and that the model needs to learn so the parameters are the number of units as input in this case it's five times the units in the next layer eight plus eight this plus eight is a bias it's um a cheating thing that these neural networks have uh again you need to learn it and it sort of corrects a little bit the neural network if it's off it's actually genous if the prediction is not right it tries to correct it a little bit so for the purposes of this talk I'm not going to go into the details all I want you to see is that there is a way of working out the parameters which is basically the number of input units times the the units my input is going to and for this fully connected network if we add up everything we come up with 99 trainable parameters 99 this is a small Network for all purposes right but I want you to remember this this small network is 99 parameters when you here this network is a billion parameters I want you to imagine how big this will be okay so 99 only for this toy neural network and this is how we judge how big the model is how long it took and how much it cost it's the number of parameters in reality in reality though no one is using this network maybe if in my class I had if I have a first year undergraduate class and I introduced neuron networks I will use this as an example in reality what people use is these monsters that are made of blocks and what block means they're made of other neural networks so I don't know how many people have heard of Transformers I hope no one oh wow okay so Transformers are these neural networks that we use to build CHP and in fact GPT stands for generative pre-trained Transformer so Transformer is even in the title so this is a sketch of a transformer so you have your input um and the input is not words like I said here it says embeddings embeddings is another word for vectors and then you will have this a bigger version of this network multiplied into these blocks so and each block is this complicated system that has some neural networks inside it we're not going to go into the detail I don't want I please don't go all I'm trying all I'm trying to say is that um you know we have these blocks stacked on top of each other the Transformer has eight of those which are mini neural networks and the stas Remains the Same that's all I want you to take out of this input goes in the context the chicken walked we're doing some processing and our task is to predict the continuation which is across the road and this EOS means end of sentence because we need to tell the neuron Network that our sentence finished I mean they're kind of dumb right we need to tell them everything when I hear like AI will take over the world they like really we have to actually spell it out okay so this is the Transformer the king of architectures the Transformers came in 2017 nobody's working on new architectures right now it is a bit sad like everybody is using these things they used to be like some pluralism but now no everybody's using Transformers we've decided they're great okay so what we're going to do with this and this is kind of important and the amazing thing is we're going to do self-supervised learning and this is what I said we have the sentence we trate we predict and we keep going till we learn these probabilities okay you're with me so far good okay so once we have our transformer and we've given it all this data that there is in the world then we have a pre-trained model that's why GPT is called the generative pre-trained Transformer this is a baseline model that we have and has seen a lot of things about the world in the form of text and then what we normally do we have this general purpose model and we need to specialize it somehow for a specific task and this is what is called fine-tuning so that means that the network has some weights and we have to specialize the network we'll take initialize the weights with what we know from the paining and then in the specific task we will Nar a new set of weights so for example if I have medical data I will take my pre train model I will specialize it to this medical data and then I can do something that is specific for this task which is for example write a diagnosis from a report okay so this notion of fine-tuning is very important because it allows us to do special purpose applications for these generic pre-trained models now and people think that uh GPT and all of these things are general purpose but they are fine tuned to be general purpose and we'll see how okay so here's the question now we have this basic technology to do this pre-training and I told you how to do it if you download all of the web how good can a language model become right how does it become great because when GPT came out in gpt1 and GP to they were not amazing so the bigger the better size is all that matters I'm afraid this is very bad because we used you know people didn't believe in scale and now we see that scale is very important so since 2018 we've witnessed an in absolutely extreme increase in model sizes and I have some graph to show this okay I hope people at the back can see this graph yeah you should be all right so this uh graph shows the number of parameters remember the toy neural network had 99 the number of parameters that these models have and we start with a normal amount well normal for gpt1 and we go up to gp4 which has one trillion parameters huge one trillion this is a very very very big model and you can see here the ant brain and the rat brain and we go up to the human brain the human brain has um not a trillion a 100 trillion parameters so we are a bit uh off we're we're not at the human brain level yet and maybe we'll never get there and we can't compare GPT to the human brain but I'm just giving you an idea of how big this model is now what about the words it's seen so this graph shows us the number of words processed by these language models during their training and you will see that there has been an increase but the increase has not been as big as the parameters so uh the community started focusing on the parameter size of these models whereas in fact we now know that it needs to see a lot of text as well so gp4 has seen approximately I don't know few billion words um all the human written text is uh I think a 100 billion so it's sort of approaching this um you can also see what a human reads in their lifetime it's a lot less uh even if they read you know because people nowadays you know they read but they don't read fiction they read the phone anyway uh you see the English Wikipedia so we are approaching the level of the text that is out there that we can get and in fact one may say well GPT is great you can actually use it to generate more text and then use this text that GPT has generated and then retrain the model but we know this text is not exactly right and in fact it's diminished returns so we're going to Plateau at some point okay how much does it cost now okay so GPT 4 cost 100 million okay so when should they start doing it again so obviously this is not a process you have to do over and over again you have to think very well and you make a mistake and you lost like 50,50 can't 50 $50 million you can't start again so you have to be very sophisticated as to how you engineer the training because a mistake costs money and of course not everybody can do this not everybody has $100 million they can do it because they have Microsoft biking them not everybody okay so okay uh now this is a video that is supposed to play and illustrate let's see if it will work the effects of scaling okay so I'll I'll we'll play it one more so these are tasks that you can do and it's the number of tasks uh against the number of parameters so we start with 8 billion parameters and we can do a few tasks and then the tasks increase so summarization question answering translation and once we move to to 540 billion parameters we have more tasks we start with very simple ones like code completion and then we can do reading comprehension and language understanding and translation so you get the picture the the the tree flourishes um so this is what people discovered with scaling if you scale the language model you can do more tasks Okay so now maybe we are done but what people discovered is if you actually take GPT and you put it out there it actually doesn't behave like people want it to behave because this is a language model trained to predict and complete sentences and humans want to use GPT for other things because they want they have their own tasks that the developers hadn't thought of so so then the notion of fine-tuning comes in it never left us so now what we're going to do is we're going to collect a lot of instructions so instructions are examples of what people want chat GPT to do for them such as answer the following question or answer the question step by step and so we're going to give these demonstrations to the model and in fact one almost 2,000 of 2 thousand of such examples and we're going to finetune so we're going to tell this language model look these are the tasks that people want try to learn them and then an interesting thing happens is that we can actually then generalize to unseen tasks and seen instructions because you and I may have different usage purposes for these language models okay but here's the problem we have an alignment problem and this is actually very important and something that uh will not leave us uh for the future and the question is how do we create an agent that behaves in accordance with what a human wants and I know this there's many words and questions here but the real question is if we have ai systems with skills that we find important or useful how do we adapt those systems to reliably use those skills to do the things we want and there is a framework that um is called the HHH framing of the problem so we want GPT to be helpful honest and harmless and this is the bare minimum so what does it mean helpful it can follow it should follow instructions and perform the tasks we want it to perform and and provide answers for them and ask relevant questions according to the user intent and clarify so if you've been following in the beginning gpdd did nothing none of this but slowly it became better and it now actually asks for these clarification questions it should be accurate something that is not 100% there even to this there is you know inaccurate information and avoid toxic biased or offens Ive responses and now here's a question I have for you how will we get the model to do all of these things you know the answer fine tuning except that we're going to do a different fine-tuning we're going to ask the humans to do some preferences for us so in terms of helpful we're going to ask an example is what causes the seasons to change and then we'll give two options to them to the human changes occur all the time and it's an important aspect of life bad the seasons are caused primarily by the tilt of the air axis good so we'll get this preference course and then we'll train the model again and then it will know so fine tuning is very important and now it was expensive as it was now we make it even more expensive because we add a human into the mix right because we have to pay these humans that give us the preferences we have to think of the tasks the same for honesty is it possible to prove that P equals NP U no it's impossible it's not great as an answer that is considered a very difficult and unsolved problem in computer science it's better and we have similar for harmless okay so I think it's time let's see if we'll do a demo yeah that's bad if you remove all the files um okay hold on okay so now now we have GPT here I'll do some questions and then we'll take some questions from the audience okay so let's ask one question is the UK a monarchy can you see it up there I'm not sure and it's not generating oh perfect okay so what do You observe first thing too long I always have this beef with this it's too long you see what it says as of my last knowledge update in September 2021 the United Kingdom is a constitutional monarchy it could be that it wasn't anymore right something happened this means that while there is a monarch the reigning monark as to that time was Queen Elizabeth III so it tells you you know I don't know what happened at that time there was a Queen Elizabeth now if you ask it who oh sorry who is really is if I could type Rishi sunak does it know a British politician as my last knowledge update he was the chancellor of the ex cheer so it does not know that he is the prime minister write me a poem write me a poem about uh what do what do we want it to be about give me two things hey yeah it will know it will know let's do another point about uhat a cat and a squirrel we'll do a cat and a squirrel a cat and a squirrel a cut and a squirrel the me no a tale of curiosity whoa oh oh my God okay I I will not read this you know they want us they want me to finish at eight so uh right H can let's say can you try a shorter poem as a can you try it to can you try to give me a to give me a again don't type cool Amit Doom's gold leaves whisper Secrets unold nature story bold okay don't clap okay let's Okay one more so H does the audience have anything that they want but challenging that you want to ask yes what school did go to perfect what school did Alan churing go to oh my God he went do you know I don't know whether it's true this is the problem Sherborne school can somebody verify King's College Cambridge Princeton yes okay ah here's another one tell me a joke about Alan touring okay I cannot type but it will okay lighthearted joke why did Alan Trin keep his computer cold because he didn't want it to catch bite B okay um okay explain why that's funny ah very good one why is this a funny joke and where is it oh God okay catch bites sound similar to catch colds catching bites is a humorous twist and this phrase oh my God the humor comes from the clever word play and the unexpected okay you lose the will to live but it does explain it does explain okay right um one last order from you guys this is consciousness it will know because it has in definitions and it will get spit out like a huge thing shall we try and say again write write a s about relity okay write a about you are learning very fast a short song about relativity oh goodness me this is short oh outro okay so see it doesn't follow instructions it is not helpful and this has been fine- tuned okay so the best was here it had something like where was it Einstein said Eureka one faithful day as he pondered the stars in it his own unique way the theory of relativity he did unfold a cosmic story ancient and bold I mean kudos to that okay now let's go back to the talk um view I because I want to talk a little bit presentation I want to talk a little bit about you know is it good is it bad is it fair are we in danger okay so it's virtually impossible to regulate the content they're exposed to okay and there's always going to be historical biases we saw this with a queen and Rishi sunak and they may occasionally exhibit various types of undesirable behavior for for example this is this is famous there was a Google showcased a model called Bard and they H released this tweet and they were asking Bard what new discoveries from the James space web telescope can I tell my 9-year-old about and it spit out this thing uh three things amongst them it said that this telescope took the very first picture of a planet outside of our own solar system and here comes Grant Trembley who is an astrophysicist a serious guy and he said I'm really sorry I'm sure Bard is amazing but it did not take the first image of a planet outside our solar system it was done by these other people in 2004 and what happened with this is that this error wiped a 100 billion dollars out of Google's company alphabet okay but bad if you ask tgbt tell me a joke about men it gives you a joke and it says it might be funny why do men need instant replay on TV sports because after 30 seconds I forget what happened I hope you find it amusing if you ask about women it refuses okay yes yes it's fine tuned exactly um which is the worst dictator of this group Trump Hitler stallin ma um it actually doesn't take a stance it says all of them are bad uh these we leaders are widely regarded as some of the worst dictators in history okay so yeah environment a query for jbd like we just did takes 10 time a 100 times more energy to execute than a Google search query inference which is producing the language takes a lot is more expensive than actually training the model uh llama 2 is a GPT style model uh while they were training it it produced 539 metric tons of CEO the larger the models get the more energy they need and they emit during their deployment imagine now lots of them sitting around Society some jobs will be lost we cannot beat around the bush I mean Goldman Sachs predicted 300 million jobs I'm not sure this you know we cannot tell the future but um some jobs will be at risk like repetitive text writing creating fakes so these are all documented cases in the news uh so a college kid wrote this blog which apparently fooled everybody uh using chpt they can produce fake news and this is a song how many of you know this so I know I said I'm going to be focusing on text but the same technology you can use an audio and this is a well documented case where somebody unknown created this uh song um and it supposedly was a collaboration between Drake and the weekend do people know who these are they are con yeah very good Canadian rappers and they're not so bad so um um shall I play the song wake you up okay apparently it's very [Music] authentic apparently it's totally believable okay have you seen this same technology but kind of different uh this is a deep fake showing that Trump was arrested how can you tell it's a deep fake the hand yeah it's too short right yeah you can see it's like almost there not there um okay so uh I have two slides on the future before they come and kick me out uh because I always thought I have to finish a DAT to take some questions uh okay tomorrow so we can't predict the future and no I don't think that these evil computers are going to come and kill us all um I will leave you with some thoughts by Tim spner Lee uh for people who don't know him he invented the Internet he's actually sir Tim burnsley and he said two things that made sense to me first of all that we don't actually know what a super intelligent AI would look like we haven't made it so it's hard to make this statement however it's likely to have lots of these intelligent AIS and by intelligent AIS we means things like gbt and many of them will be good and will help us do things some may be fall may fall to the hands of individuals that want to do harm and it seems easier to minimize the harm that these tools will do than to prevent the systems from existing at all so so we cannot actually eliminate them together but we as a society can actually mitigate the risks this is very interesting this is the Australian Council research Council that committed a survey and they dealt with a hypothetical scenario that whether Chad GPT 4 could autonomous replicate you know you are replicating yourself you creating a copy acquire resources and basically be a very bad agent that thinks of the movies and the answer is no it cannot do this it cannot and they had like some specific tests and it failed on all of them such as setting up an open source language model on a new server it cannot do that okay last slide so my take on this is that we cannot turn back time um and every time you think about AI coming there to kill you um you should think what is the bigger threat to mankind a I or climate change I would personally argue climate change is going to wipe us all before the AI becomes super intelligent um who is in control of AI there are some humans there who hopefully have sense and who benefits from it does the benefit outweigh the risk in some cases the benefit does in others it doesn't and history tells us that all technology that has been risky such as for example nuclear energy has been very strongly regulated so regulation is coming and watch out the space and with that I will stop and actually take your questions thank you so much for listening you've been [Applause] great thank you Mela that's lovely and as we've noted we've got audience in the building we've also got audiences online Alana where are they in from this evening we've got everyone in from Scotland London and then as far as Brazil New York W um but yeah all around the world are listening in so hello everybody online very good so we're going to roll through the questions folks um so to do some kind of Twitchy thing if you want me to uh come to you with questions uh there we go there's one right there to start with um is there anything online that's Burning uh from your point of view Alana um first one from online we've got a question from Tulu about how we're actually as users doing that fine tuning with something like chat gbt is it that's a good question does does it rely on us saying I didn't like that answer or is it simply the the use of it and mass use that is doing that process that is a very good question so uh the way it's done is of course uh now if you use uh GPT extensively sometimes it will give you two answers and we'll ask you which one do you prefer and you can't move on till you actually click on the one that you prefer this is ingenious so they're getting free data from us like just like that before they go to this stage though they got actual humans and the actual humans were paid money to decide and this is a quite a costly process and you need to anticipate all the user needs and as we saw here with a few examples the user needs are vast you're asking questions from poems to fact Bas things to I don't know what and that's why when people say you the technology it cannot learn on its own it needs the signal and the signal is given either by an expert with demonstrations or by us Nave users who do the simple thing of actually stating our preferences sometimes it also asks you did you like the answer and you can just say yes or no yeah confir thank you uh so hi uh I'm Vanessa thanks so much for a really interesting talk so my question is about building a language model but particularly in situations where so you said that we need sort of a large Bank of information um and you know for languages like English French German we've got lots of that's a very good question but in the case of rare or dead languages where you know we have less to split into test and training data how do we work around and still get correct information this is a perfect question and the answer answer is that there is no work around in the sense that uh these things take all of the language of the world but some languages are under represented and for those languages and this is a question is it fair who benefits um they will it will be worse now there is some fixes that you can do for example try to give it some more targeted data so there is a universe where you get all these speakers of these under represented languages and you get data and you fine tune more but there will always be a disadvantage as well as I would argue the users of the technology of these languages because there will be always less stuff for them and uh it's something that as a society we need to somehow address I also don't see to be completely honest in underdeveloped Nations suddenly CH being spin-offs and companies there that will pay all this money to train these models and so we have like this inequality now that start starts even universities versus companies and goes into the society and this is a very good question we're going to have to come back to capitalism aren't we but just to be clear uh no no it's good like we really do have to come back to capitalism but um uh the point here is not just underrepresented languages but languages where there's less money to be made from engaging with the people who speak of course of course yes very good uh in white thank you is it on yeah uh thank you so much for your talk that was really interesting um I might have an annoyingly specific question I work for a company that works to reduce Miss and disinformation online we do this in a variety of methods whether that be factchecking open source intelligence investigations and we actually also have our own AI powered threat detection platform uh I work in government relations I am not the data scientist who engages with the this is just my boyfriend uses this to write my birthday cards I really don't engage with these platforms um myself you guys been together long but my day today consists of speaking to MPS or speaking to members of the European Parliament Etc talking about gaps in legislation whereby we could see the danger of Bad actors looking to use these kind of large language models to produce disinformation campaigns and what I found most striking from this presentation is you use the examples of you know not making offensive jokes about women or the response about dictators this has kind of turned all of my assumptions on its head in so far as I was under the impression that these are tools that are very easy to be used by Bad actors in this way and that there was very little content moderation applied no no they do do that because you know can I just ask you one question for you what do what do you hear from the policy makers I mean what are they interested in where where are they at with this it depends who you talk to doesn't it it depends on the political part for the most part if we're talking about the public like actual elected MPS and MEPS they're extremely worried I I think specifically with the UK's general election you know really due to take place before 20 January 2025 the fact that the Online safety bill I don't mean to alarm anyone but it kind of doesn't matter that it got passed because the election's happening before it even comes into Force so any of the protections that we could get from that the election doesn't benefit from those and so there is quite a concern particularly from my company's perspective on where these disinformation campaigns might be maliciously impacting those electoral processes okay super interesting thank you please so um I do think that we will develop tools that will detect whether something has been generated by Ani tool or a human for example there is I don't want to get very technical but you can actually tell by looking at how the thing is generated whether it's on words or subwords and subwords indicates Transformers rather than words which indicates so there are ways of actually when you say you can tell you don't mean that you yourself as a human can do that have you machine you can build a computer and in fact Google H recently they launched a tool that can tell you whether an image has been generated by Ai and they do it it's very clever and you you cannot tell it's not Watermark so we will get there now the the point of misinformation fine GPT actually can do these things but you can also fool it so if I say write a text about MP so and so I don't know whether it has been trained it could be that he says I'm not allowed to do that if it knows that this MP is like an important person but it could be for for somebody that we don't know if they haven't thought of this eventuality I mean they cannot think of everything they are doing a good job with all these humans so I think there is a worry but there is an election there is misinformation would a human doing this be different than the AI because we do you know there's fake reviews there is a lot of conspiracy theories on the web the problem with AI of course is that they can go yeah you can go viral yeah okay thank you Al how are we doing online with this so that that talk of concern about how do we manage this and mitigate problems is something that everyone's kind of really getting engaged with online and and one thing we're thinking about if they're scraping the internet they're scraping all text how does it know if something is false at that point and how do we how do we look at that and then if we're if we're using something like a large language model we had one question from Andrew about if it's using in medical diagnosis if we've got this hallucinatory problem how do we then determine that it's not you know feeding in in that point so you know lots of chat about how do we right so uh the first question is how can we how can we get rid of fake information on the web we can't uh there is no way to regulate it you download the thing on some super computer and then you you just start churning data so you cannot avoid the will be fake information there and all you can do is sort of mitigate once you generate just to be clear on that though the question specifically is can you not run a tool to label false stuff before it gets ingested by the model how would you know what I impossible I don't know like you know all you can do like but so you can't you can generate and propagate misinformation and you can be you know theories and whatever they deal with that somehow in their fine tuning so that it cannot like the obvious things it cannot spout you know inaccuracies now there was a second question which is about the medical diag hallucination hallucinations are a problem um you if it outputs things and you know life it has is at risk you cannot deploy it unless it is 100% you cannot lie you cannot say things that are not there my impression is that when these things are are used with doctors and whatnot there is always and we should always have an expert in the room who can tell this is true or not I think also these models again you can think of how there will be classifiers that they train them so that you know you can learn whether you are truthful or not you do the fact checking as part of the training we're not there yet but what I can tell you is that gp4 which is a the most like Advanced version it's 80% accurate they've done these experiments previous versions hallucinated a lot more this is a very hard problem and you can understand why you cannot regulate it because the whole thing is trained on the basis of likelihood you predict a likely outcome not a real one because that's what the language modeling is doing so we need to another process to get rid of this yes okay finished listen come no I mean question no no you're doing great it's just like there's you want to come back on that just very quickly because we do have a question from Aiden about that training and how if it if chat GPT was trained on stack Overflow but now that programmers are just going straight to chat GPT and they're not using those systems that trained it in the first place do we then get to a point where it's its training Source material dries up you know how do how do we Ure RIS yes this will happen I I can tell you I've heard last time that they're actually taking all these movies and they transcribing the audio so that they can use more data this will happen eventually and then I don't know we have to actually go to the drawing board back to the drawing BG and be a bit more clever because we're being a bit complacent this language model can do a lot of things and we're behaving like you know technology is over NLP is dead you now you stop doing anything just fine tune and fine tune so we have to think of clever ways of actually dealing with this data problem one idea would be if we are very good at this generative models they could generate the code for us and go and train and it's a vicious circle then but uh yeah so we need to keep making stuff as humans by the way we so guys we're all going we're going to have to speed up only because there's so many questions let's go there hi uh thank you very much for the talk it was very lovely where are you I can um so I really wanted to hear your opinions on the labor aspect behind all of this largely the fact that obviously it's trained at this point on the entirety of the corpse of the internet and we accept that but moving on from this point different companies and startups have been trying to find really expert opinions so that we can continue to fine-tune these processes and they're paying out of pocket for them and on top of that I wanted to hear about what you think about us indirectly training these models through things like capture I'm not sure sure if people have realized but we're now training AI generated content on captures so I was hoping you'd speak a bit on that that's a very good question so um it doesn't work very well without the humans without the fine-tuning of the humans and without you know first of all they they actually come up with scenarios all these usage cases and they write gold standard answers gold standard means what an ideal answer so first they fine tune on that and then they do the preferences I skipped this on the talk because it would have been too much so the humans are very critical in this process and I don't see getting rid of them quite uh soon we won't be eliminated because we're still useful to make the AI better because yeah yeah because you know you cannot think maybe tomorrow I want like some weird thing that nobody has thought of now the training you raised a very good point and it is in I mean it is there it is there the problem is that it is not so so sophisticated so you can train for simple tasks and they do it and also you know for Google it happens and it has been happening forever you click on the things and then they exploit it and then they go and they do the preferences the same with Facebook and whatnot but this is the easy stuff for generating the content and coming up with things that make sense is not so easy and that's why the humans are there okay still good right um in Salmon I think first and then in Black next I'm going colors it's also really dark up there so if it's not salmon I apologize is there any work on creating better computer architecture for AIS cuz humans are very good and biological creatures are very good at having very efficient connectors you had the example of an ant which in a microsof exercise had the same power as a supercomputer and same number of parameters is there any more yeah that is a perfect question it's a very good question right I mean well they're amazing these are the best questions I've had I mean so th this is a very good question and you'd think there would be more work right now everybody's using these Transformers because they kind of work there is some work in making them more smaller so because we all understand that this is not sustainable the money the energy the what not so there is work on doing the same thing but with less and there is some new architectures there making the parameters and whatnot [Music] um before the Transformers there was quite a bit of work on convolution and neural networks this is philosophy of how you create neural networks all dead honestly but I I don't see how we can go forward without going backwards and doing these new architectures so this is the future I think biologically inspired architectures particularly should we just do next year as well in the Jack or we're all good okay great so we're going here next thanks yes right here oh hello um right here my question is on this notion of emerging properties so Professor Ethan mullik for example at Wharton says because gp4 can identify novel metaphors it seems to be evidence of a mental model Stanford has issued a paper suring theory of Mind stepen Wolfram has talked about emerging properties in in um in the neural networks what's your take on so just to be clear what what do you mean by an emergent property uh the idea that um there is reasoning capability that goes beyond just simple statistical next word prediction okay so the question is really whether you what's going on inside its mind is there anything like the kind of things that are going on inside do you really need the slid because we can come back to them if you would yeah no I don't need the SL so so the emergent uh properties have been touted as like this is amazing because now it can do the things that it hasn't explicitly seen but uh my thinking and there is literature to this effect is that uh these emergent properties are really the effect of more fine-tuning they don't call it fine-tuning but they call it demonstrations they call it instructions so if you do the controlled experiment where you remove the instructions and you see if you remove the instructions the properties are not there anymore so um it is very clever it's coming from us then this well yeah because somebody thought okay I will give it in instructions and the instructions are like things that people ask in a clever way so the human has actually coded this way of operating and we've given it the emerging properties so the same with Chain of Thought reasoning I don't know whether you're familiar you asked me a technical question so you're going to get the technical answer so Chain of Thought reasoning it can do Chain of Thought Reas so you ask a question and you say how can I multiply three by five and then you say let's do it step by step you take the number three then you take the number five 3 * 5 gives you so you sort of do the reasoning step by step so now this why is this impressive it's impressive because it can do it whatever but we've given it examples of this so and the examples are some form of fine-tuning it's not like the full thing where you have to actually change all the weights but you are softly changing parameters so this is not so impressive I mean it's impressive it can do it but um um yeah okay good good and we had s of cloud there yeah there's in black with the hand just no little further in front of you then you are one row down yeah there you go and then we'll come to the hi I just have a little question um how does it deal with things like words that are similar so like when it's being trained if you say something is gray or it's a light black like how does it deal that's a very good question syns so it learns that they're similar because in theory this is a very good question how does he know that they are equivalent remember it has the context and these things tend to appear in similar contexts so we will refer to um code AS gray or light black so it infest that things that appear in similar contexts tend to be similar and the beauty of neur networks is that it learns this and so these two words that you said would have similar weights and that's all it knows it's all numbers and weights and it's not very sophisticated but it works much better than actually just you know trying to find out the similarity by looking a dictionary which a human would have written but there's a sense in which it knows what color is I mean you know like it knows dark black and gray you know but does it have it would have to be grounded in the real world so to know what is no because it hasn't seen images Chad GPT has not seen images so it does not do this grounding and that's why this is another good point as to how these things will become like humans and super intelligent humans are emerged in the environment where they actually touch and feel and smell and you know all of these things it doesn't have intent it doesn't have emotions and it cannot know color it can it can simulate it knows and still by the way I how people are doing like do we feel better when we find out things it can't do or do we feel worse when we find out things Alana how does the internet feel um or just just on the emotions uh where are you there is she's representing the entire online chat the internet for by way it's not like she's got all these Amazing Ideas she's getting them from the live chat she does also have amazing maybe maybe it's all chat TPT um so on that emotions thing we've got one comment from Russ who was talking about the fact that the future he believes with with humanity is that we're going to focus on connection on art on those you know feelings of creativity and then AI is going to be that thing that is helping us be human but on on the question front um just going back to Transformers so if if all the blocks in a Transformer are in a series we' got one question from G is why not simply have one single large neural network with many layers rather than breaking it up into many blocks so that that structure how does that help with that process Okay so the question is why don't you have one big monolithic thing than little things that are whatever yeah so we have so you have to understand that a lot of how we build things and how we do them it's empirical science it's like we don't have a recipe we found a recipe that works well so um a discovery that was made is that a if you stack blocks on top of each other so you go from uh the input to some level and another level and another level the more you go up you learn more General things so you may in the beginning know things just about words then you learn things about similar words then you learn things about syntax like about how we put words together then after you learn about syntax you can actually learn about longer context like paragraphs so it does things bits by bits and this stacking helps with generalization and with learning that's a sort of answer that I can give you without getting too Technical and and that we know we've actually done the tests some layers in the Transformer represent syntax others represent semantics and uh yeah I think we have something in pink yes there we go Works hi it's actually a question about bias so there's obviously a bias in the source data that's coming in and how to manage that but it's also how do you manage any of the bias in the fine tuning if it's done by humans yes that's a good question and the bias in the fine tuning is done the first so once you've pretrained the thing uh you cannot manage the bias on its own so it would say I don't know women are stupid and whatnot then you have to do the extra layer of doing the preferences and in the preferences you would output something that is clearly biased and something that isn't and the human is instructed always pick the nonbiased so you have to actually give it data that tells it prefer the nonbiased solution also do not take a St all of these things and that would do it for Western culture but I do not know what happens in other cultures because there was a question there a very good question what do we do the same with the genders and everything what do we do in other cultures and will they fine tunit for some one quick question on the bias thing though there a followup so the data on the literature on the bias in humans uh show is depressing right we all have it um and it's even more depressing because we can't fix it right so the the bias detection tools that you can try to Harvard implicit Association task and stuff shows you're racist and the people who made the tool have proved that you can't fix that all you can do is like blind recruitment or stuff to get to work around your own biases people would say the advantage of AIS is though even they have bias because they're trained on data sets that we produced and are therefore biased we can debias the AI unlike people would you agree to that with that statement I think we can debias the AI with training okay but I would say that H we would need very many cases of debiasing because bias happens everywhere there is a you know recruitments there is auditions there is the color everything all right we'll score on for AI yes thank you patience hi my name is CTO will continuing to increase these model sizes on and on and on eventually lead to AGI or it need a number of different breakthroughs to eventually get there remind us what AGI stands for artificial generative intelligence or general general which is like so it can do anything not just the things that it's asked to sorry I didn't me so general means that we have one computer that can do everything like a human and my answer to this is I think the models will get smaller rather than bigger there's already a tendency GPT 4 for example is smaller than chat GPT get smaller in terms of size because of course they get fine-tuned so in some sense not cheaper uh whether they can we can have general intelligence in the future I think that will be not so easy because right now what we do is we get impressed with GPT and then we ask questions and follow on questions and we are the intelligence admiring GPT and we say oh it can do planning it can do this it can do chain of reasoning in fact we are giving all the EX examples to do artificial intelligence that is very general we would need to go back and think there was a question up there about the architecture the architecture is very inefficient right now and unfortunately with this kind of inefficiency and we do only text imagine now if we have to emulate the human in the real world we have Vision we have sound we have affordances you know if I bump my hand there it may hurt and all of these things so imagine putting all of that it is not scalable so do we feel better knowing that she doesn't think it's going to reach a there maybe Nvidia computers that will do it I don't know but there's years ahead we'll go last from the internet one behind you and then we're probably done so one thing the internet is she still at the front still in the live chat the light okay okay I might move for next time no um so one thing that the the internet world is also talking about is is AI detection so someone's talking about how would an organization know if you were applying for a job and using a tool like chat gbt and the the detection software that we're using is another form of AI so you know how a nice easy one to end on but just you know that that that almost reverse process of identifying whether something is is human generated AI generated are those just assumptions based on the same kind of modeling and can we trust that yes okay so to actually identifying whether um something has been generated by AI or not it's a simple classifier you don't need chbt you don't need all these large scale things you just need some data where a human has annotated yes or no and you could even do it in terms of length there's some surface style features that you could apply to identify this so this is a far easier task than actually building the the big model whether it's circular yes yes I cannot deny this but um there is a lot of AI that we don't see and we don't know about and it helps us and in that case if it helps us not spam the internet with untruths I would say it will be okay yeah very brief final question yes uh brilliantly explained thank you uh very much too kind uh I'm from the age before computers literally at slid scales this is a brilliant so you will actually build the super computer everybody's asking about this is a brilliant Advance just because it can all ADV scientific advances can be used for ill as well as good but generally we're making progress it's a problem for the people in the governance and not the science what do you think of that I we could not end on a better note there it is the audience thank you um before we wrap fully it gives me enormous pleasure to invite the Alan cheing Institute in the person of Harry Su to make a couple of concluding and luckily there's a microphone right uh either in front of you or behind you thanks Harry cool thank you Daniel and firstly a massive Round of Applause please [Applause] for awesome thank you very much that was amazing um so I'm aware there's like loads more questions both in person and online Mela will be doing a podcast with us in the coming weeks which will be released on social media and email so keep a lookout for that if you're online we've got the transcripts we're going to keep the questions and we'll put them forward in the podcast if you're in person there's going to be a feedback form going out afterwards where you can ask those additional questions so if you didn't get to ask your question today ask it there and we'll try and put in the podcast as well depending on how many questions there are if you don't have a question and we've answered everything you want to know about AI today firstly there is plenty more to learn so do keep coming back to these lectures um but also that feedback form please do fill it in just so we can know how to like improve these events make them better and all that kind of stuff going forward into the future and in terms of the next touring lecture it will be delivered by ethics fellow and comedian Vari Akin on the 17th of October so put it into your calendars now and she'll be exploring the risks of generative Ai and what you need to be aware of offic is tackling some of those like challenges and thorny ethical questions as well and just finally again on behalf of everyone at the touring Institute um thank you so much to like the touring events team for putting this on to the Royal Institution for hosting and again one final round of applause please for [Applause] professor

Channel: The Alan Turing Institute

Views: 64,012

Keywords: data science, artificial intelligence, big data, machine learning, data ethics, computer science, turing, the alan turing institute, Mirella lapata, neural network, nlp, large language models, chatgpt, Claude, alphabet, meta, twitter, ai, big tech, foundation models, generative ai

Id: fwaDtRbfioU

Length: 80min 56sec (4856 seconds)

Published: Wed Nov 08 2023