Large Language Models: A Path to Artificial General Intelligence?

Key Takeaways

so yeah before I begin um so I'm uh author of The generative deep learning book as being said but I'm also co-founder of a company called applied data science Partners we're based in London um and we are hiring so if anyone on the call is looking for work in London uh and is a data scientist or data engineer AI engineer um anything technically related to AI then please do send your CV over um we'd be happy to kind of take that into a consideration when we're hiring and so just send it to myself at David adsp.ai and um we can get back to you so uh yeah large language models so I would start this presentation by saying um you know July 1969 was obviously one of the founding moments in our history as humankind it showed that we were able to for the first time dominate our physical world like we'd never been able to do before by Landing a human on the moon showing absolute control over our understanding of how the physical laws of the universe work and so it's around the same time actually that people were asking very big questions about intelligence what it means to be intelligence whether we can create systems to be intelligent and a lot of this was obviously being driven by things like the space race and the the idea that well if we can land a man on the moon then think what we can do with the agents and robots uh in intelligent systems on our end planet and it seemed very very close it seemed only years away that we would be able to develop systems that were as capable as you know the rockets that were putting people on on different um bodies in space that we we were just around the corner from a revolution in AI as well but actually there was nothing really farther from the truth we were still very very far away from the AI that we know and love today um and it was only really recently in you know maybe 2016 2017 that people were really talking about AI that was uh almost as a new frontier of what we can achieve technologically speaking and one of the founded foundational moments was the alphago paper and alphago system that was developed by Google deepmind here in London and this was a system that was remarkable for a few reasons firstly it mastered the game of Go which has I believe more board configurations than atoms in the universe showing that not only can we um build systems that we're very capable in in simpler games like chess but that actually even the unimaginable number like the number of atoms in the universe is able to be understood and operated over by an AI based system and it was called alien at the time by Lisa doll the opponent of alphago because he was seeing that this system was able to create ideas that he almost saw himself as human and that it was almost impossible to believe that a human could come up with such creative and beautiful ideas over the board um and then you know after this we sort of started asking ourselves well how close are we to AGI artificial general intelligence are we just around the corner from it because now we've got Alphas go and it's only a matter of time before that turns into AGI or would be quite far away and I think people realized we're quite far away still because these systems are very narrow they don't allow you to adapt the rules of the game and still deploy the same model and expect the same performance they they're not zero shot Learners you can't just tell them to do a new task and they are good at it first time around um and so there needed to be a new injection of something and we didn't know what that thing was really back in 2016. but you know when we start thinking about things like AGI we start thinking about things like chat Bots such as portrayed in the film her um and it was actually language and natural language that was the key to unlocking a lot of the ideas that we see today around systems that have been talked about as close to AGI and specifically specifically the the model that we are obviously talking about here is is the gbt model that has come out of open AI where the T here stands for Transformer and it's talked about heavily in my my book um but also is a general way of modeling uh natural language as a sequence of tokens that is more capable and more powerful than simpler ideas such as uh recurrent neural networks and lstms gius anything which kind of treats the sequence as a uh something that requires a hidden state that you need to update and maintain but instead of that hidden State Transformers adopt a different idea called attention and it's this attention mechanism that allows the system to at inference time basically pay more attention to certain preceding words in the sequence than others that gives it this incredible capability it's actually I think a much simpler idea than the likes of lstms um if you see the diagram of the lstm cell in my book you'll soon see there's a lot of complexity in there that seems kind of ridiculous now that we're looking at attention mechanisms are actually a lot a lot simpler to understand and actually a lot more human I think the idea that we're sort of at the point at which we need to decide the next word we decide previous words you know how much attention do we want to pay to them in order to uh to detect and predict the next word so what's amazing is that at scale this is enough to produce a model that shows signs of general intelligence I think that's that's true to say um and when I say at scale I really do mean it sort of an exponentially increasing scale um the early GPT models here were 100 million parameters big in uh 2018 uh maybe breaking the billion parameter by 2019 but now uh with the likes of GPT 3 175 billion uh 3.5 which is rumored to be more the par model of Google which is a half a trillion and now gbt4 which is rumored to be of the one trillion plus Mark although actually recent reports say actually it may be an ensemble of smaller models nobody really knows except those in open AI um with these kind of this kind of scale what we're seeing is that actually you just throw more compute and more data at these models and they do just get progressively better and better um and by better you know what we mean is that they are exhibiting what are called emerging capabilities so these are capabilities uh that we wouldn't expect to necessarily uh see as part of uh the model as it's been trained because it simply trying to predict the next word but actually there are certain characteristics that come out of doing that process that are quite amazing so things like information synthesis um being able to actually compose an image of maybe an ASCII art or something um understand code reasoning theory of Mind embodied interaction these are all things that sure are embedded within text and within human written text but we're not explicitly telling the llm to learn how to do reasoning it seems that just by understanding and having a model of natural language that these things pop out as a sort of happy byproduct if you like of the the training process um I just wanted to highlight here this is a system that we've built as a demo in-house at applied data science partners called um the interactive protective game so feel free to go on and play this it's just a simple um web app that you can post your API key into from open Ai and it will generate you a little uh whodunic mystery on the Fly and you can ask questions to all of the different um characters in the game um so it's like a it's a natural language driven RPG if you like where you play the role of the detective um so of course with all of these the the game is completely open there is no barriers on what you can ask the people in the game you can get them to talk to each other if you want um you can ask them what their relationship is with other players in the game and we know that this is all possible because of these allergic capabilities um you know gpt4 has the ability and GPT 3.5 actually to withhold in its memory in its context you know the the who did it so who is the culprit and maintain that state all the way through to the end of the game um so yeah it's a cool little example that sort of demonstrates some of these emerging capabilities I just want to highlight a few more examples that I've been quite excited about recently um just very quickly so Cicero is the first of these I think this was uh one of the early signals that we were getting that there's more to these land Rich models than just just language um this is a system that was developed by meta to play a game called diplomacy and anyone who's played diplomacy will know it's largely a text driven game you have to talk very uh deeply with your counterparts that you're playing with both your allies and your enemies in order to negotiate cooperate and coordinate um so it's a bit like Risk as a game but there's a lot more discussion and you have to submit your actions all at the same time and then the negotiation obviously plays a big part in forming friendships and and um yeah potentially bluffing and what they showed is that they can train a large language model to do this very very well um Cicero is an AI agent that plays the game of diplomacy and it comes in the top 10 of all players even though amongst those that are experts so an absolutely amazing example of how you can embed a large language model into a into a wider system so more of an agent-based turn-taking action taking agents and that's that was the first time that I thought actually there's more to this large language model stuff than just being good at producing uh blog posts for example second example is from a paper called generative agents interactive simulacra of human behavior and for anyone who's played a 2d Sim game this will be sort of very familiar to you this is a basically a 2d Sim game that they'd uh mocked up for the llm to play and the idea is that every character in this game is operated by a large language model so all of the little agents walk around bumping into each other and have the ability to converse with agents and share information so each agent has its own memory store it has its own ability to use tools within the environment and through interactions the researchers are able to demonstrate that information and knowledge could flood across the network just through conversation and in fact quite complex and detailed ideas were being shared amongst Asians and even exhibited things like misinformation um where where you know one agent would say something to another agent and then that was slightly misremembered and it would be passed on incorrectly um and it was quite amazing to see how quite complex interactive Behavior could could emerge from the system so for example one of the players decided to I used to party and then they would go around kind of inviting everyone to their party um and even another of the agents invited someone else on a date out to that party and then that information spread through the network and and all it was like a little Society almost that was being created before our eyes um and they all took on different roles so some were artists some were authors they formed opinions and you know had arguments one of the things that you would expect in a human society were were being shown to develop just by allowing agents in this world to be driven by large language models so again a real indicative and pivotal moment I think that showed the power of training things like GPT models um now a really interesting paper that you might have seen recently is this one called um uh Sparks of artificial general intelligence that came out a few months ago and it was showing basically how gpt4 is already developing and exhibiting some forms of AGI that even more than that if you give it the access to tools and other interfaces that it can actually learn and understand how to use them to even enhance its ability that's inherent within the llm and a good example of this is is large calculations right so GPT is not at all designed to do large multiplications in its head so to speak however if you give it access to a calculator and say look instead of just doing this within your own language model why didn't you ask to use the calculator if you feel like you need to use it and it does so it just knows out of the box that a calculator in this tool would be a good thing to use to solve this particular problem um and so by providing these tools the idea is that we can augment the large language model and we use the llm almost like as a general intelligence engine in the same way that we would use a human as a general intelligence engine and we wouldn't expect it to do things that are anti-human like be able to you know multiply huge numbers together or run a python prompt these are things that you and I can't do without a computer so it's almost you know augmenting the llm to be able to give it these tools in order for it to solve more complex problems um and this yeah leads on to the idea of llm autonomous agents this is a really good diagram I think that explains what we mean by that it's where the llm is just part of a wider system and that it isn't everything but it is almost the brain the central points through which a lot of the other tools and memory stores and task lists and prioritization lists um can operate so for example you might set the llm a general task say to um so to go out and do some marketing about a particular company company it would then break that down into a series of subtasks just using the llm and save that to a memory store and then it could spawn uh separate copies of itself or just continue with the same copy to um either in parallel or sequentially solve those subtasks and of course there's a tree uh structure to all of this so those subtasks might be broken down again some might be dead ends that don't lead to anything in which case that's that information is saved to the database in its memory um it might decide at some point to use the web and do research through web browsers or it might decide to use interactively apis as long as you give it access to these tools it's quite capable of being able to go off and do different things um so Auto GPT is a good example of a autonomous AGI or AI agent I should say um it's one of the fastest growing open source repositories of all time I think it's now got more stars than um than like Pi torch or tensorflow it's an absolutely remarkable um example of what the open source Community can build very very quickly uh and my last one is Voyager so this is a really good example as well of a uh an agent that is able to do something in a virtual world autonomously um Voyager is an agent that effectively can continuously explore the Minecraft world and any skills that it learns first of all by randomly exploring but then of course by thinking more deeply about what it needs to do are saved to a skills database and what I mean by that is the agent has the ability to write python code right to interact with the Minecraft world so as soon as it finds a piece of python code that's written is actually doing something useful it just saves that python code to a database and so the next time it needs to use that ability instead of having to use the llm to write the code it can just use the llm to say do I have any skills in my database that will allow me to do this and if it says yeah you've got a skill here that allows you to turn say a block of wood into a pickaxe then it will do that and it's it's about this idea that the llm is just the general intelligence engine that is tapped into a load of different other systems and memory stores that allow it to over time get better and incrementally more powerful within the environment that it's based so I really highly impressive example I'd highly recommend that you check out the paper for that one particularly um and so I just wanted to finish up really just by kind of summarizing my thinking around llm agents and where they might go in future there are opportunities and risks I think on both sides of the equation there's no doubt about that on the opportunity side we're talking about being able to build agents that can solve problems uh very very quickly and at a global scale and even do things like scientific research although there's a lot I think there's a few steps that we need to put in place before that's possible I think the second thing I'd like to mention as well is the ability for this to turbocharge education um imagine if you were reading my book but alongside that there was a large language model that had been fine-tuned or even just placed into the context window of that llm my entire book so that instead of having to jump onto this call and ask me questions you could just ask the agent if there's anything that isn't understandable and that sort of thing is is possible today and I think there's products being built you know as we speak to really push that kind of application into the world um and yeah obviously like enhanced creativity and entertainment is a big area for generative AI both on the text or image space but also texted music text to video um we're talking about already sort of Hollywood movies being able to be produced in five years time just through AI I think it's a really exciting time for the entertainment industry to see where they might take this technology on the risk side I think a lot of these come with sort of equal and opposite risks many a lot of the time so one of the big risks that people talk about is the risk of huge mass unemployment there's a lot of reports from like McKinsey and those kind of big consultancies saying large parts of the workforce won't be needed as a result of AI I think I hold a slightly more optimistic view than this um and I think I sort of see AI more as a tool than as a thing that's actually going to fully automate and replace a person's job although you know as we've seen with you know other Technologies there will be changes I think to the worst Workforce albeit not massive unemployment um people talk about the control problem as well is there a risk of a iron AGI running away and us not being in control again I think it's an unlikely scenario a very unlikely scenario actually but um perhaps we can discuss that in more detail afterwards and then there's obviously big questions around misinformation alignment um so this is where we're talking about the idea that yes we can build very powerful AIS but do we know that they're aligned to our goals and our general um ethics and you know what we would consider to be good behavior in the world um and lastly dependency you know do we really want to become dependent on AI agents or is there not something about the human condition that actually demands that we have a role to play in this world and we don't just arm off all of our responsibility to technology um so yeah there's huge questions on both sides of this argument um and you know I I'd sort of see myself as somewhere in the middle of this I'm not particularly strongly opinionated one way or the other I'm not an AI doom doom and gloomer but also um I don't think it will be a perfect Utopia um so just to finish up uh my last slide is one of my favorite quotes from JFK um he said this in 1962 um this is seven years before the moon landing um and he was talking about why we're doing this like why are we bothering to to build these systems that can can go to the moon and he encapsulated in this quote I think something really important which is that humans love to solve hard problems we don't do things that are easy we do things that are hard and I think AGI we are effectively at the same point I think with AGI as sitting here today it is almost like 1962 where we're sort of seeing that this stuff might end up being possible but that we're going to have to face a ton of complexity and challenges before we see this kind of Technology actually making a tangible and positive difference in the world um and so I can't wait to see before the end of this decade whether we reach true AGI and I think it's possible but we certainly need to be all aligned ourselves to out what kind of future we want to create um so thank you thanks for your time listening and happy to take any questions from any of you about that or about anything in the book that was so amazing thank you so much um I have a lot of questions about your slides and also about your book but I would like to open up the floor too uh our book club members if anyone has questions please feel free to ask in the chat or just unmute yourself and ask directly one question I had is what is Agi in your presentation uh like yeah what what do you what do you think it is because if you have a response you can go ahead but I can also provide more context no it's a great question I I can I can give you my thoughts on that straight away so when people talk about first of all AI what I what I mean by AI is just assistant that can solve problems now typically AI has been narrow so it's been a system that can solve one problem very very well AGI I think there's no there's no kind of one definition that I can give but what I would say is that there is a sliding scale from an AI system that can solve one problem to an AI system that can solve any problem now at some point along that line I would say AGI occurs and I don't have a kind of in my head an idea of exactly where that line is drawn but certainly gpt4 is further along that line than any system I've ever seen however and I don't mean this in a kind of flippant way but it's not going to make me a cup of tea and what you know what I mean by that is I think gpt4 is not at the moment tasked to make physical changes in the world and that is for me a major component of what I would consider to be true AGI um it's going to be a I think um an overlap between Ai and Robotics I think and I do see it as really important in future that any system that we consider to be truly General intelligent will need to be able to make very significant changes in the real world and I think this is where a lot of the danger comes in as soon as we start giving systems that are intelligent access to tools that can make physical changes in the world so not just kind of changes to a string of text but actual you know physical things that we can see then that's where we need to start talking very very seriously about regulation and control um so yeah we're certainly further along the um let's say the the kind of intangible problem solving where it's not actually making a change in the physical world but it's like solving uh something within the text string within the context window but that we're actually behind on a lot of other things like Robotics and we're seeing this with self-driving cars right the the systems are trying to do something in the real world and it's extreme extremely difficult to reach that kind of level five self-driving okay so it's it seems to me it's like your definition is like the set of capabilities that a human can do it's like once in AI can do all of those things then it can solve all the problems we can it'll be like AGI yeah exactly that's it and I think like we don't give ourselves enough credit sometimes as humans for how much we can do you know yeah we're pretty dope yeah we're awesome and and you know the AI systems are great but we need to also give ourselves credit for what we can achieve um over and above these systems I think all right thank you thanks for the talk your slides are amazing um you talked about your definition of AGI I mean would you say that Shane legg's definition um Intelligence being the measure of an agent's ability to achieve goals in a wide range of environments but is it something that you also like um agree with and uh also um should we achieve this AGI um is it should the jail to be embodied in humanoids only is that when we can say that now we've achieved AGI or um should embodied in any other thing let's say drones or they talked about cars self-driving um there's art and also my other question is on an alignment um people from around the world have different values right you and I probably value different things um on alignment we disagree on a lot so are you in favor of like say having like a body a unpodian stuff that regulates and says hey this uh whatever these are the values everyone agrees to so if you're building um AI systems we should achieve X Y and Z that everyone agrees to that's online yeah cheers yeah thanks thanks for the questions um so I'll take those one by one so yeah I'm broadly in line with that definition um I don't particularly you know like the semantics I think are part of important but also like I think generally people have an inherent understanding of what AGI will feel like when we see it um so yeah the only thing I maybe would change that definition is I think it's important that anything that we truly call AGI is robust to changes in the environment so it's not just about being very very good at a diverse range of environments when you've been trained on a diverse range of environments it's about the AGI system being able to adapt quickly to new changes whether and new tasks as well um and I think llms are particularly good at this because everything is done within the context window so you can introduce an entirely new task in the context window and because it's sort of I guess the way of thinking about it is it's mapping any any activity that you can give it into something consistent which is a stream of tokens and so the stream of tokens is the environment in which it operates that that is its World um that's why they're very very powerful General learners um so yeah that would be that would be what I add to that um and then I think your second point there was around embodiment yeah so I I don't necessarily think humanoid but I do think um we need to be extremely cautious if we're allowing AI systems to make physical changes to the world and what I mean by that is giving them access to tools that can um that can cause us harm like in a physical way not just you know connecting them to tools that um you know make API calls to another database or something but I'm talking about like you know the ability for AGI systems to have access to like you said in the case of putting them into a drone like that for me is extraordinarily dangerous and a bad idea and I think this is where your third Point comes in which is the sort of Regulation and ethics of this and yeah I do think that there will need to be some sort of body or at least governments that is agreed upon unilaterally that kind of talks about you know what we do want these systems to do and what we don't want them to do um in the same way I suppose you know we have things like human rights conventions and we have sort of agreements about you know War and what what what war looks like from a legal perspective between countries and whilst no one obviously wants anything like this to happen there needs to be some sort of broad agreement as to what's um almost like what are the rules what are we trying to all agree that this is at least a set of standards that we can all agree on so I think it would be you know these would be fairly broad I don't think they would be particularly um um you know detailed and you're right that everyone has a different sort of sense of uh you know what is right and what is wrong in the world and Alignment is going to be tricky between countries but I do think there are a sort of ultimately a set of standards that we can all agree on that would prevent you know catastrophic failings and failures in future um I don't think we're my personal viewers were not close to the kind of catastrophe that some would have us believe that we are close to I think there's a lot of pieces missing we've effectively built a system that can learn very well what the next token is in the sequence and that's wonderful and they they are generalist Learners and they can do amazing things in the context window but there's um the technologies that we're using for things like memory skill retrieval um you know the ability to prioritize and Order tasks these are older Technologies these are things we've had for years there's no developments really being made there um you know looking up of information is being done by a cosine similarities still within most Vector databases so yeah we've got one new technology on our hands and it's important to remember that's that we've developed one thing but it's it is awesome but it's also like just one part of the puzzle um so yeah I hope that answers your questions probably yeah yeah thanks awesome we got another question from the chat what are your insights on the ethical implications open source and slash diffusion models for example I'm thinking about stable diffusion that have been mostly ignored by the community even though it tried to solve them previous versions ethical concerns yeah so so kind of summarize some I guess still man the side for both both sides of this argument you know the open source Community would say look in order to solve these you know problems that we're having with things like um you know training data not being particularly open uh and the fact that any problems with the model are being owned by a very very small group of people and trying to be solved by a small group of people why don't we just release them into the world and we'll see we'll be able to tackle the problems a lot more efficiently if there's more eyes on it um and that's the one side and then I guess the opposite view to that would be well if you don't really want these models getting into the hands of very powerful people with uh nefarious actions and nefarious um morals so yeah I I kind of my own view on this is I suppose I fall slightly down more on side of the open sourcing of things because I think technology technology especially AI technology we're not talking about something like nuclear weapons where the actual starting to get you know to get actually started with a nuclear weapon is actually there's a huge kind of barrier to entry for most people so I feel AI technology is something that we can't we can't close off to people people will find ways to learn about AI technology and how to build models like this um whether or not they're behind open AI stores or not I mean people move jobs people like talk to other people people write blog posts it's just not the way that Tech has ever been done I think to to sort of say look we've got this new software and we're not going to tell anyone how it works I think the open source Community is an incredibly quick and um they will find ways to to produce similar Technologies and in some ways if not better we're starting to see this with large anguish models where they were initially I guess open and open AI were very clear about how they were building their models and then suddenly it all went behind closed doors with um the gpt3 where it's not really being released and then even more on 3.5 and 4 the whole trading process and methodology wasn't even released um and yet now we have commercial models like Falcon and uh MPT and the Llama sort of lineage of models that are not quite as capable but certainly catching up to the likes of the most powerful gpt4 models so um yeah I don't I don't think the argument of sort of just trying to keep it in a box is um particularly strong one so okay I I had a question about like you know making stuff so as you know like open AI just released like the function calling stuff and but Lang chain does this and we're talking about building autonomous agents I myself am like a Founder so I'm choosing a tech stack and betting on it right so what's your view on like that situation because the the the the fear is like the library is gonna break down or like the community is going to dissolve or something right so I just love your thoughts on that yeah another great question um I'm a Founder as well so we're sort of facing the same things as our clients and you know where they're coming to us and saying like do I bet on open AI or is it Microsoft Azure and everything's going to be just in there soon um like do I build do I host everything in the house because I don't want the data moving anywhere uh like the big questions for companies what I would say is first of all just build everything as modular as possible so if the worst comes to the worst and the community breaks down I don't think it will by the way the Lang chain I think there's a there's a real use case for it um then like yeah you'd have to Pivot away from it and find other ways to kind of you know leverage the same technology within commercial offerings uh proprietary offerings um but you know a I don't think that's going to happen there's a very thriving community and I think you know they're talking about stuff like you know investment and like yeah almost like setting up as a proper company and stuff so um yeah there's no way that I can see that happening um but yeah I do think like Microsoft and open AI seem to be fairly ahead um although we're hearing from deepmind they're going to release like Gemini model next week and that's going to be even more awesome than chat GPT and so yeah when I hear this stuff I just think like build everything modular don't try and like build anything around a particular engine but just think to yourself how can I modularize this bit of the solution so that ultimately it's just an impact an output machine and I can click in any llm that I want in future um yeah I mean that that's that's my only advice I would say yeah makes sense thanks foreign since we all have read the book or a book club members so let's uh uh change gear a little bit to talk about your book um but for viewers who have not read the book would you mind give a brief introduction about the book and tell them why they should read the book yeah sure uh here it is this is the book um yeah background but you get the gist it's a book um it's really it's quite quite meaty it's much meaty than the first edition um so the first edition I wrote in 2019 before the world I've heard of llms and like it wasn't a big concept like people were talking about it but it was still very Niche um and I came to the second edition like last year when I was starting to write uh about this time last year actually and so again this is before to actually BT this is before uh Dali 2 had just been released actually but it was before stable diffusion um and I thought to myself like all right that's cool there's a few new things to write about and maybe I can sort of get away with just you know updating a few chapters and um yeah just freshening it up with new examples and then like everything exploded and suddenly the whole world was talking about generative Ai and I think I realized I needed to rewrite the whole book um there's just so many topics and so many things that you know were true in 2019 that now it's you know it's been blown out the water um and so the whole book is Rewritten I think from the ground up to be what I hope is a true reflection of the generative Ai and Landscape as I was as I was seeing it unfold before me um so in 2022 but then also yeah into early 2023 and what I've also tried to do is not make it aligned to any one particular model type like you could write a whole book about Transformers and people do you could write a whole book about the fusion models or Gans um but I think what the last few years has taught us is that we don't know where the next big idea is coming from um you know if you talk to people in 2017 or even 2018 2019 they'd have still been talking about like building bigger and bigger Gans um and there's tons of books out there that are written about Gans that now look like well you know we've got diffusion models now um and even the advanced Gans are kind of using ideas from um yeah from other kind of fields like Transformers so I think what this book hopefully does is give you absolutely the groundwork that you need to understand any kind of generative AI model and it doesn't try to push you into one sort of Direction and believe that Transformers are everything organza everything or diffusion models are everything you need to know but to keep an open mind it talks about kinds of model that I haven't really seen written about elsewhere like normalizing flows energy-based models which is what you know is very very keen on exploring um at meta within his like Japan architectures and stuff so like the next big idea might come from energy based models and suddenly we won't be talking about Transformers anymore or we won't be talking about diffusion models and so yeah if you want a general sort of foundational understanding I think I hope I hope that what I've written is the book that you need to get that awesome that's really helpful thank you we got a hand erased earlier uh would you like to ask your question now I think he answered it because I wanted to ask you concentrated level not heavily but I saw compared to your previous book the introduction of uh Transformers so I just wanted to ask like you see the field uh converging towards transfer language and Transformers especially um but it's I guess you answered it talking about just don't uh concentrate on one particular model because you don't know what will happen tomorrow you know something else might come and uh append Transformers and all but all in all I just wanted to ask you what do you see with Transformers at the moment I mean they're doing pretty well in a lot of things even Vision uh supplies with the vision Transformers um what do you see the future of Transformers themselves yeah you're right they are being used in different modalities not just text Vision even video obviously the other day I think the general idea that you can treat a lot of kinds of data as a stream of tokens is quite a powerful idea um which is why Transformers have done so well I think with Transformers the important thing to kind of keep on top of is um like what are the different components that are being clicked together within the Transformer to make them work and a good example of this is attention you can learn about attention from my book and from I know other resources and that will give you a base understanding of what attention is and what it does but then it's like the branch of it or the trunk of a tree you you then need to kind of explore well what other stuff is being done today that builds on the idea of attention so there's there's things like flash attention for example that I used in a lot of the state of the art models these days um that give the attention layer first of all the ability to be much more memory efficient um and also the training process to be more efficient and faster and to maintain you know the ability for it to be very powerful um so you know what I guess you're not going to get from any single book is a complete overview of every single tiny like paper that's ever written about attention um but you know what you can get from the book is what attention is what it can do why we're using it why it's better than recurrent neural networks and so on so that you can then read a paper that is Advanced like the flash attention paper and at least understand what it's talking about when you start talking about Keys queries values um embedding dimensions and all this kind of thing um so yeah I mean where I see Transformers going is that yeah I think they will probably end up getting bigger but also I'm excited about the idea of it being a um a type of model that you can use in a more um sort of ensembled way so I think that it'd be really interesting to see how uh gpt4 when the paper if they ever released the the model information um whether it is just an ensemble of smaller Transformers whether it is this one giant Transformer that tries to learn everything um so yeah I would say just to kind of you know keep on top of these little add-ons if you liked the base model uh Laura it's another one l-o-r-a low rank um adaptation which is again the ability to fine-tune Transformers and on in a much more efficient way um and quantization there's a ton of little kind of sub fields that come off the main branch of Transformers yeah I appreciate your reply thanks you mentioned the Transformers will get bigger and bigger but also in the book you think you said the models will get more and more efficient and smaller so do you think you will get bigger or get smaller yeah yeah great question I I could go either way I've probably given both opinions at some point I think if you look historically what's happened is that the model sizes has kind of stagnated because in 2020 we had tpt3 which is 175 billion parameters and people were already saying oh in 2021 they're going to be a trillion because if you look at that curve from 2018 which is 100 million through to 175 million in 2020 you know like 2023 we should be we should be up in like the multi-trillions if you follow that curve but we're not we're still being talked a lot of models are still bouncing around 175 billion parameter model falcon for example the Open Source One Is 40 billion um you know gpd4 is rumored to be a trillion but might just be an ensemble of smaller models um so when I say they get bigger what I mean is I suppose they'll get more efficient with the number of parameters that they are using um so they might get bigger in capability but they might stay the same size or if not smaller um in actual parameter count and that's because ultimately if you're going to deploy these models they have to be highly um parallelized they have to be really quick at inference um so you can't just keep adding adding more and more parameters without a knock-on effect for things like inference time um so I think yeah they might get better they might get bigger but they also might just sort of become more efficient using the number of parises that they've got right and maybe use different types of Hardware even yeah exactly there's a lot of trips now being built that are specifically designed for kind of um the matrix multiplication that goes on within um Transformers for sure very cool I have a kind of a specific question like the chapter 10 a lot of our book club members find it hard to understand advanced games I wonder if you could talk a little bit about like pro-gang Style gang video game and other models mentioned in the chapter just high level um and which of those are the most state-of-the-art models and you mentioned stable diffusion is and perform the state of state of the art games so it doesn't mean like we don't really need to learn about them yeah that's a great question I've had I've had questions about this before as well you know people say why are you even writing about Gans at all um because you know isn't it all about diffusion models now and so on and I think there's two reasons and two answers the question as to why they're in there first of all they certainly add to the history in the the richness of the field um and I think it's important to recognize that still a lot of the state-of-the-art benchmarks are still held by Gans like style Gan XL for example is a absolutely state-of-the-art model for image Generation Um and also the second thing to recognize is that even within so-called non-dan architectures a lot of ideas from Gans are still being used so to give a concrete example the idea of having a discriminator that can try to detect whether something is real or fake is still being used within things like Vision Transformers um there is a component to for example um uh the model type called VQ gown is still driving uh things like Dali 2 um and so like a lot of the a lot of state-of-the-art sure is diffusion models but actually they borrow ideas from dance quite heavily and um so yeah so yeah it's it's important first of all to recognize historically the significance of the Gan but also like the Core Concepts are still very much relevant today um in terms of like that specific chapter I think what I wanted to do is give a little flavor almost like a yeah um like one of those meals that you get where you get a little bit of lots of different things uh of a few different states of the art architectures and to show you that lineage from like we've just done a whole chapter on how to build a gan and now here's what state of the art looks like and to show like it's often small tweaks to the underlying architecture that make a huge difference so even if it's not completely 100 clear um you know how uh mathematically something works um to take for example style gain as an example the ability to kind of inject that style Vector into different layers of the Gan is what gives it the power of being able to be able to you know just adjust the hairs on somebody's head compared to the entire way in which they're looking left or right and it depends at which point you inject that style as to where the the style gets represented is it at the high level or is it at the low level and we're starting to see ideas like that in diffusion models as well where in mid-journey now there's a kind of very subtle and very significant or something like that button where the diffusion process is kind of stopped after a certain point and then the style you want to um to add is injected at that point or you can stop at the process much earlier and inject the style higher up the diffusion chain um so yeah it's it's about sort of recognizing the the links between different ideas uh I think you know it's true to say a lot of the best ideas ever come from a crossover between different fields um it's people talking to each other who've studied Gans really heavily talking to people who studied diffusion really heavily where a whole new idea might come up that no one would have thought of on their own um so yeah it's good to have a general understanding I think of every field uh including guns and diffusion models and everything else that's really helpful thank you we have another uh question from the book club earlier about the music application which I find is fascinating because before I thought music isn't that just text but then when I read read about our chapters like it's totally different it's so much more complex right you mentioned the method of combining uh notes and duration tokens separately you mentioned about how to do fellow phonic music there are two methods great grid tokenization and also the event-based tokenization and also like music is so fascinating you're trading piano roll as picture um yeah so those are really fascinating to me the question is kind of funny uh we have a question about do you play instruments do you think it's important to understand music theory to be able to make music applications yeah great question I do play an instrument I play the cello um so I yeah I've always played the cello my whole life as long as I remember from about eight years old and I think you know when I was working out what I wanted to do as a kid it was always going to be something in AI or something in music um so yeah I chose the AI route because uh yeah it's very very hard work to be a professional musician I have a lot of respect for people who do that um so yeah I love going to concerts and you know orchestral music is one of my favorite things the concertos and so on um yeah but to be given the opportunity to talk and write about the intersection of those two things is really exciting for me which is why I wrote a whole chapter about it because I wanted it to be quite a significant part of the book um and yeah the the ways in which so first of all I think it's important to note music generation lags behind on the scale of uh AI development behind something like llms and text based generation because first of all there's a ton of language out there to train on um so we've already got the the training data there's not as much music I think out there so to be able to be to be pulled into the models um but also I think music is a lot harder and you think about language generation it's it's it's token of tokens of integers effectively that you're looking to to generate most languages certainly English um you know is non-tonal so if you're if you're trying to create um you know English English spoken text or an audio file or something that's a lot more that's a lot easier than um yeah than some of the more tonal languages like Mandarin for example um but then also I I think if we're looking at music itself there's so much Nuance to that sound as to how it makes us feel emotively that you don't necessarily get in text text is more about grammar rules um and when you listen to a piece of music I think you can you can quite quickly understand that why this is so complex to generate because you know whilst there's not as much grammar to learn I think the complexity of music comes from the subtlety and how it's interpreted by The Listener um and it seems that quite amazingly two people can kind of listen to the same piece of music and get something completely different out of it so it must be even harder for an AI to understand what makes sense musically and what doesn't make sense um so yeah I know this company is kind of working on this actively this Harman AI That's an offshoot of stability AI they're working actively on the music generation problem obviously Google with music LM that was released recently I know meta have just released a spoken voice I think it's called something like um localizer or something like that really interesting as well that's more to do with human audio generator human spoken generation um but yeah in the book again we just go through the fundamentals of what kind of problem we're trying to solve here it's still a sequence prediction problem but no duration obviously has a big role to play as in text I can speak very very slowly or very very quickly and the meaning is the same but in music it isn't that case it isn't the case and you also have pitch as well so I can speak very very high or very very low and again the meaning is exactly the same but in music basically things are completely different so um there's a whole two extra Dimensions there that you know you don't have to worry about I guess if you're just dealing with with text generation right it's amazing that it's lacking behind um just curious to see what would happen next because the method you mentioned anything is pretty advanced yeah I think um you think about it a lot of music generation is solvable now so like if you're just doing a jingle for an advert or something or um something that's like I've got a drum track and there's maybe a tune over the top of it you know that sort of music I think we can safely say is is is uh not solved but at least on the way to being solved but the question is whether you could get an AI system to write a whole Symphony with multiple movements and ideas from the first movement of that Symphony you know recapitulated in the last movement or to sort of to have different sort of tones and emotions that are captured within that piece of music that mean something I don't know we're not certainly not there yet it's the same idea I guess as you know writing a novel isn't it it's very different to write a piece of copy for your website or a blog post or a you know a tweet than it is to write the entire Harry Potter series and all of the ideas and the consistency across those seven books or Lord of the Rings or whatever um you know these epic works of literature seem completely you know aliens or to certainly to me I could I could never imagine being able to hold that amount of information in my head and produce such a consistent and wonderful work um and it'll be interesting to see whether and if we can get AI systems to do the same or whether we'll always be able to say yeah it's not quite as good you know yeah yeah makes sense uh we got three minutes anyone have one last question so we can ask one last question no okay I'll go ahead so there are a lot of startups in the space you know every Enterprise wanting to have a piece of pie on this but I feel like a lot of the ideas are not valuable what kind of applications or ideas do you think are actually valuable in this space okay I would say okay there's low hanging fruit for sure like the low hanging fruit is can I respond to my customer service inquiries faster can I write copy quicker when a new product comes onto my website um can I talk to my documents in ways that I haven't been able to do before so instead of me having to read um a book or a paper and sort of just sit there and wonder you know why did they mean this why didn't they say this I could just ask the document um so there's kind of low hanging fruit like that I think the next level Beyond this is things like um using AI within education so to perform kind of personalized learning experiences for every student where they can ask an assistant or an agent to help them with a particular thing that they're not understanding um that feels like slightly more valuable and also more complex to me to maintain kind of consistency over time to come up with personalized learning Pathways and plans for you know putting together a test just that was tailored to you um and so that you can you know enhance your understanding I mean I'm trying to learn German at the moment and I'd love something like that to kind of every day just give me exactly the 50 things that are 20 things let's say that I need to um level 20 questions that I can write at the limit of what I'm capable of learning um there's nothing really out there that does that at the moment not not well anyway so that's another angle and then the third you know where is this perhaps going in the future I would say the agent-based llms are I think the most exciting thing I've seen for a long long time in technology um and there's problems with this at the moment I mean even Auto GPT it's got like 150 000 stars or something and yet I've yet to see I've yet to see a real sort of absolutely killer app that someone's built using Auto GPT um and I think the reason for that isn't because the idea isn't good I think the idea is fantastic but it's complex and it's really difficult um to get these agents to kind of maintain consistency over time um you know they are ultimately it's like you know we're here and they're trying to find a trajectory to over here somewhere it's very easy to miss and it's about how do we give them that kind of human oversight of sort of stepping back and saying hang on I'm off track here like I'm not solving this problem at the moment um so there might be ideas something like reinforcement learning that are being built in um also I'm excited to see deepminds Gemini that they've sort of rumored to be incredible and use ideas from alphago um and yeah there's a lot more to come I think but they're the sort of three stages I would I would bucket things into the low-hanging crew and then you know actual practical personalized applications such as in education and then maybe agent-based llms in future love it love it thank you so much we're at time really appreciated our conversation it was such a great learning experience thank you so much for the book and the conversation um yeah you're very welcome thank you I'll post the videos you know I will send you the link right and I I just the last thing I'd say if it's okay um thank you first of all for having me it's been absolute pleasure talking to you all if anyone fancies leaving a review on Amazon I would love you to do so it really helps um first of all just for me to get feedback but also just helps with like other people wanting to understand what the book is and what they can expect I would love it if you could leave a review and I'd be really appreciative of that so thank you yes well do thank you so much bye