Therapeutic Outlet - AI Will (Probably) Not Replace You

Transcript

Nobody can miss them for me That’s reason why I try both This is all that’s old, I can see a thousand times of it This is all that’s old, I can see a thousand times of it I’ve been watching Northern Exposure, the 90s TV series for a while now It’s now available again on Amazon and I used to love this TV show when I was a kid and I always really basically hero worshiped Chris, the radio DJ sometimes known as Chris in the morning, who is not only a very handsome fellow but says some pretty profound stuff and I always sort of pictured myself maybe at some point in the future I could be like a radio DJ because this was before podcasts and even actually really in any meaningful sense the internet existed and so I could sit there and yammer on about whatever was in my brain and people would listen and so I guess I got my wish to a degree whether anyone’s listening or not is an open question but here I am pretending to be Chris in the morning nowhere near as easy on the eye and maybe I’m not as easy on the ear either but it is what it is, I am what I am and I’m here and you got me With all of that, what am I going to talk about today? I’m going to talk about AI and I’ve spoken about this several times before in the past there’s mentions of it on quite a few episodes but I believe I’ve only ever done one episode dedicated to it and that was in July 2024, so last year a year and a quarter ago I guess and that might not seem that long a time but in the age of AI it’s quite ridiculous, it’s a lot and I mean a lot has changed since then not just in the world of AI but in my world as well and so I wanted to sort of revisit it I should note before I get into this that I may make some statements like I did on my last AI episode which was called AI will not replace you and I should point out that I’m not referring to everyone or anyone in any part of their life will not be replaced in any way by generative AI I recognise that for all its faults as a technology set it does do certain things reasonably well and that might put pressure on certain industries and certain skill sets and people will rightly or wrongly lose their jobs I am talking about this in my capacity as an artist and a podcaster rather than making any proclamations from a wider business perspective although I might allude to some of that so if you are under pressure from generative AI on your day job you have my sympathies and I recognise that there’s more going on than just the stuff that I’m going to talk about here anyway what did I say a year and a bit ago well I said that generative AI as it was then is pretty good at some stuff but it’s not that good I said that ultimately machines can’t create culture and that artists being humans will always be in demand because people connect with people not machines I also said that I consider it completely legitimate for people to use AI as part of their creative process and that indeed it’s completely legitimate to use AI entirely as your creative process because in my mind it’s still art, it just isn’t necessarily going to be good art and I would argue that in most cases it isn’t but either way getting an AI to produce something of even vague value is not quite as easy as it might first seem and you’ll often end up producing lots and lots of versions of something before you find one that achieves whatever goal you’re trying to achieve and that itself is an artistic choice, a qualitative choice and therefore there is some creative element that goes on in that process now things really have moved on since then I should just reiterate that as I said before in the prior episodes my day job to a degree is generative AI or AI and machine learning, data science I’m a data scientist for my day job when I can get the work doing it and that involves both the usage and the creation of forms of AI and there are quite a few and I in my past roles up to well over a decade ago was experimenting with the early versions of what we would now call large language models like chat, GPT and Claude and so I’m pretty familiar with both the process and technology required to create these models as well as how to use them to the point where I actually have another podcast it’s called the confusion matrix wherever you get your podcasts if that sort of thing interests you which looks at things from a purely AI perspective with probably more of a focus on language models and how they’re deployed and used and how we maintain quality and accuracy and so on on these things and I take quite a critical view there as well and so I’m quite deeply embedded in this stuff and I have lots of opinions but those opinions have to sort of be adjusted all the time because things are changing so quickly although I’d argue that over the last say six months they’ve changed a lot less frequently, things have moved on we seem to meet started to head towards a plateau where the progression of the technology is not quite at the pace that it has been but last year in July I talked about tools like Dali and Mid-Journey as visual creative tools and just to clarify what these things were they are tools where you can go in, type in a prompt, a set of words can you make me a picture of a squirrel dressed as a SAS gunning down a bunch of Easter eggs and it will attempt to produce that and in most cases would produce something and back then Dali which no longer exists has been replaced by OpenAI Sora Dali was OpenAI’s image generation model the stuff that was producing was not great you could get it to produce some interesting things that were often odd and didn’t quite do what you wanted them to but were quite funny but often completely mangled and just bad and since then various other models have emerged that are both from a static image creation perspective as well as a video generation perspective are much more effective and so the one that I’m probably most familiar with is Google’s image effects which is part of a suite of tools and it’s really good now I need to qualify that a little it produces extremely high quality images that are often quite close to what you might want if you use the right words to describe them you can get it to produce hyper realistic close ups of human face for example and they will look highly photographic and extremely realistic perhaps too realistic if such thing is possible but the sort of glossy, high quality, high definition images you might find on a fashion site or something that’s been photographed with a really good quality camera but then manipulated with Photoshop or something so these things are perfect so not really that realistic but like insanely good quality and it will produce this in a matter of seconds and let’s put the valid criticisms aside for a minute and the plagiarism and copyright issues that’s amazing that it can do that and I think this is an aspect of the generative AI revolution and the predictable and gathering steam backlash it’s worth pausing just for a minute because I would struggle to find someone who wouldn’t agree when they sat and thought about it for a minute that that wasn’t completely amazing that that’s possible at all and even a year ago something like that simply wasn’t conceivable and how amazing it is that these things can write large volumes of pretty high quality language or code programming languages and any other things that you may have criticisms about but the very fact that it can do it at all the very fact that I can ask it to build me a Python script that does something fairly complex and I have done many times the fact that it can just sort of do that again with problems, with faults but the fact that it can do it at all is mind blowing, right? and if you’d asked me a couple of years ago whether we were anywhere near being able to do stuff like that I would have laughed at you but here we are, this stuff, despite being completely spectacularly amazing that it exists at all, has become already a bit old hat and people are looking at it again, well that’s not very good is it? it doesn’t get this right, it doesn’t get that right and the criticisms are completely valid but still we seem to have forgotten about how amazing it is that this exists at all now I’m being very very careful here to not paint the picture that eclipses the problems that are created when, now that these things are out there being used in a sort of commercial arena I’m simply stating that the fact that it can do it at all is just so cool and from a person who comes from my industry and does what I do for a living and understands how these things work it’s especially nerdily amazing and I’m not going to stop thinking that, it blows my mind but we are all haters to a degree on the rest of it because I use generative AI a lot I have to do for my job to keep up with what’s going on I have to do for my podcast, I have to be using this stuff I have to understand it, I have to understand the differences between the models and what’s available and what’s out there and again I’m not an authority on this stuff but I’m pretty well versed on what’s available and I use it a lot and I use it in my day job in terms of, I write very little code anymore I use generative or large language model agents that are able to do multi-step processes for example, I think a lot of people will look at chat GPT as being the most advanced implementation of language models but it’s not by a very very long way there’s a lot more coming down the line if you’ve only ever experienced chat GPT which is pretty much a one-shot thing by which I mean you ask it a question or you give it a task and it does that one thing, it responds with tell me when Taylor Swift was born and I hope it might give you a right answer it might give you a wrong answer but again give you an answer with some extra words around it and then it stops and waits for you to do something else to ask it a different question like where was she born or something and you can have these long chat streams and because the whole stream is fed back into the model every time you make a query it can give contextual answers so if you’ve asked that question when was Taylor Swift born it will go whenever that is, I’m not going to look at that, I don’t care it will, and then you ask the question and where was she born without mentioning Taylor Swift again I certainly should recognise that’s what you were referring to and give you whatever answer it thinks is correct which may or may not be correct and so it can look like a sort of an interactive process it can look like a conversation is evolving it can look like the model knows something about you or at least knows something about the context of the current situation but that isn’t true at all, models are completely static they don’t learn at all everything that the model knows about you or the current conversation is sent to the model every single time you make a query and in this sense they’re sort of limited that particular facet has its values and usages but it can be quite limiting now we’re moving into a world, particularly if you’re already in technology where we have these things called AI agents which have some level of autonomy and control over something so you might say to it can you take all my photos of Taylor Swift in this folder rename them and put them in another folder on my computer file system I don’t know why I’m talking about Taylor Swift here I guess it’s because everyone’s always talking about her and therefore her name seems to pop into my mind quite a lot but I don’t really spend much time thinking about Taylor Swift otherwise or even listening to her music so I don’t know why you want to move your pictures of Taylor Swift but that’s a multi-step process to do it yourself you have to know where that folder is you have to open the folder, you have to list the files you have to control A, control C, whatever you’re going to do to copy them you have to go find this other folder or create it then you have to control V or to copy them into that new place and then if you wanted to rename them, so they all had say Taylor Swift at the beginning of the file name then you’re going to have to go through each of those now if there’s 2 or 3 of them, probably not a big issue if there’s 20, 30, 200, 300, that could be quite a time consuming process and you have to go through each of them and click on them and change the name or if you’re using a Unix or Linux based system it’s a little easier but still a little bit more time consuming whereas with an AI agent you can just type in the words in English can you move this folder of images to that folder and rename them all with the words Taylor Swift at the beginning and it will just go off and do it and that’s a multi-stage process and you’ve given it at that point access to your file system so it’s run multiple commands and done multiple actions and it’s done it all on its own, it might ask you for permission it might ask you to clarify, it might do any number of things but when it gets on with the thing it’s doing, it will just do it and that makes these things incredibly powerful where there is some level of repetitive process and something that can be done electronically that will otherwise require human intervention and be quite time consuming and potentially boring and of relatively low value so presuming you’ve got a good reason for moving those photos and maybe that’s going to make your life much easier for some reason it’s still relatively low value as a task it’s not going to change anyone’s life, it probably just makes your job a bit easier and maybe you wouldn’t have bothered doing it at all had you not had an agent there to do it for you and this is where the pressure comes from this is where I think a lot of people just see chat GPT as a proxy for search like a proxy for Google as a bit of a limited use product if you’ve never experienced an agent I can understand that but once you’ve seen these agents at work you start to realise how powerful this stuff can be now the agents suffer from exactly the same problem that the one shot language interfaces do which is that it can get things wrong and if something gets Taylor Swift’s birthday wrong then that’s annoying but if when moving all your files it actually used the wrong command to delete them permanently that’s considerably worse and if that wasn’t a set of fairly innocuous image files but some vital legal documents or something then this could be very problematic indeed and so these things aren’t a magic bullet they’re going to solve all our productivity problems in certain circumstances when monitored properly and you understand what they’re doing and how they’re doing it the right guardrails are put in place they can be really really incredible productivity tools and that’s great but those guardrails and your knowledge of what it’s doing and how it’s doing it is essential and you need to pay close attention to these things but when people are talking about jobs being taken they’re not saying that chat GPT takes their jobs they’re saying that one of these agencies and there’s lots and lots of products that call themselves AI agents and all the big companies Amazon, Google, OpenAI, Anthropic and so on all have their own versions of agent and if you’re using platforms like Microsoft VS Code as a developer then there’s lots of them available within that too so these things currently very much in the tech nerdery realm are going to start spilling out into everyday life and will do things like automation of processing of imagery and I use these things quite a lot I want a mosaic of images for like a social media share or a little leaflet poster thing I made to advertise my art portrait work so I wanted a little mosaic, a little grid of images that had some of my examples of my work and I could have just done it in something like Word or PowerPoint and dragged and repositioned and so on these images and done it by hand but I knew I had this need quite a lot so I just had an AI agent to write me a little program that given a set of images or a folder full of images will create a little mosaic or grid of those images based on certain parameters the amount of columns, the amount of images and so on and it’s incredibly useful and I wouldn’t have bothered doing that before it just wasn’t worth the effort but it took me five minutes to do and every time I think oh I could do with it being able to do this as well or that then I just go to the agent and go can you update this now so that it can take this parameter or I can change this about it or whatever resize the images, make the monochrome, whatever and so already I’ve got capability that I probably wouldn’t have had I could have created it, I could have probably found a piece of software that did it for me but I just built one that did it exactly to spec and I used that a lot and I think it’s going to start seeing more and more people realising this capability to manipulate images, change them, group them and this is just in terms of basic productivity I don’t see that that task in itself particularly affects anyone else it’s just made my life a bit easier and made my marketing perhaps a little bit more effective and so these are the types of use cases that are going to come up more and more but as you can see when you start to think about things in multi-step or being able to do things you couldn’t have done before enterprising people are going to start finding ways to do much more complex and exotic tasks with these now this is before we even get into the realm of of just being able to type a prompt into an image generation model to create whatever you want so I use generative AI constantly and I think I mentioned in the last episode I did on this that even at that point I was building out my own AI image generation tools and training my own models to create these gnarly, weird, spooky pictures that I created by constraining image generation models so they were not actually very good at doing that and they got images all mixed up and mangled up in really cool exotic ways and I’ve since moved on from that because what I realised was that the image generation models that we’ve got now things like ImageFX and Sora and others did something similar without me even having to intervene because much like the language models where they are getting stuff wrong the image generation models get stuff wrong too but as opposed to just giving me something that’s just incorrect they give me something that’s weird it too is incorrect but I might ask for a picture of someone sitting on a chair and they might give that person three arms which if what I want is a picture of someone sitting in a chair for an advert or something I want to create or for a reference for an artwork but what I really need is someone sitting in a chair with just the normal number of arms that’s highly problematic but if you’ve got a brain like mine who really really likes real weird, unexpected, serendipitous stuff then someone turning up in an image with three arms for me is beguilingly, fascinatingly brilliant and I love it and that’s a fairly facile example because these things tend to get things wrong in all sorts of weird ways and I find it incredibly entertaining and I also find it really really stimulating I find that the weirdnesses and the wrongnesses that are created by models will really fuel my creative juices and give me ideas and I think that’s no different from the ideation process that any artist who wants to create pictures of things that don’t exist in the real world would go through anyway you’re looking for things in the world that will trigger off an imaginative thread that you can pull on and sometimes it might seem like that stuff comes to you out of the ether but almost always it’s about something within the world that you’ve experienced will trigger that off and I find that the weirdnesses of these models are really good for my creative process and to the point where I love the stuff they produce so much that I just print it out and start sticking it on my canvases and use it as a baseline or a starting point for my artworks that’s not unusual for me I use a lot of collage in my art the collage comes from things I found on the web I use imagery of people and places and things that are out there and available and as part of the collage process that’s kind of how collage works and so I’m already repurposing stuff that is from other people’s work and some of the themes and the subjects of my artworks are real human beings like movie stars or so on and I’ve always been probably skated slightly too close to the copyright line on that which I’ve always been very uncomfortable with but given that I’m aphantasic, I have no mind’s eye I’m kind of constricted to depicting imagery from photographs because I can’t imagine and I love painting pictures of humans, that’s what makes me happy and I can’t imagine humans I can’t conjure up something in my brain and paint it or draw it because there’s nothing to conjure up so I’m kind of constrained to images of stuff and what generative AI allows me to do is create sort of generic images that I can use as a starting point or to work from as a reference now I’m aware that to get to that point the models have had to plagiarise lots of other people’s stuff and whether or not the image that’s spat out to me is a direct copy of someone else’s picture or someone else’s likeness is entirely opaque to me I’m not even sure if it’s currently possible to find out whether that’s the case and I’m absolutely certain knowing what I know about these models that the provenance is not something that the model creators are able to verify one way or the other, the models don’t work like that they are abstract mathematical entities so you can’t reach into it and say well which bit of code created this image because nothing created it, it was just a statistical, probabilistic, mathematical rendition based on some input text they’ve got no more idea of how I got to the output that I did than they do, right? than I do, sorry and so it’s copyright dubious but in some ways that ship has sailed I think that the creative fightback from the creative industries is picking up speed governments are slowly starting to react to this and things are being legislated for and the landscape will change because of this and I’m going to run with that, right? but at the moment I find image generation tools, particularly Google’s image effects to be so beguiling and fascinating and interesting and useful that I find it very, very difficult not to use it it’s very, very, very well suited to my needs and I hope I use it well but I’m aware that there could be criticisms to this and how I’m using it I’m open minded to hearing those but it’s hard for me to get past the reality that these things are useful for me and I was already using image generation AI of sorts before these things started to become really, really powerful I’ve experimented with the video side of things as well but video isn’t really my medium but I’ve done some quite cool stuff with that mainly just to see how it worked and I think we’re quite a long way away from people being able to generate long form video currently you can do like 4 or 8 seconds at a time with something like Google’s Flow or Sora and that’s cool because it will look amazing in some cases and I’ve produced some stuff that looks absolutely spectacular but then when you want to continue that it’s very hard to do it’s very hard to constrain the machine to do what you want and so the flip side of me using this to create weirdness is that if you want to create something that is very much to spec you have to write quite a lot of prompt and be very, very detailed in particular and specific to get something even vaguely like what you need because you need a very particular image of something or a video of something and then a lot of the time the model will just not do what you want if you don’t understand how that particular version of that model needs to hear what you need to tell it then it ain’t going to give you shit of any use it’s going to keep getting it wrong and you’re going to have to keep iterating you’re going to have to keep burning GPU and CPU time which is a lot of energy and a lot of water consumption these things are very bad for the environment and to get something right you might have to produce thousands of images before you’ll get one that probably just about maybe works to spec and I think this is a much bigger problem more often than not I’ll take whatever comes out of the machine I’ll have it iterate a couple of times to give me some options but my usage is actually relatively low and I’ve got, I’m very idiosyncratic I want the weird images, I want the messed up-ness but as most people don’t and so from a commercial perspective the image and video generation stuff I think despite what many people have claimed is far from coming of age probably famous last words here because things really have moved a lot faster than anyone should have expected so maybe doing long form video is only but a few weeks or months away but we’re not there yet but there are plenty of roles that are going to be put at risk by this type of stuff particularly in the creative industry apparently coders are losing their jobs I think anyone trusting an AI to code on mature code bases deserves everything they get and very little of that is going to be good but I feel like there’s a certain amount of plateau happening here the most recent language models iterations GPT-5 was not a disaster as such but it was a big PR nightmare as a lot of people who were using GPT one of the GPT-4 range of models freaked out because the personality of their personal assistant suddenly changed plus GPT-5 from a technical perspective is if anything a step backwards and GPT-5.1 just came out to almost no fanfare and the same goes for the various other model creators the public have shrugged collectively although many people are using these tools on a day to day basis no one’s particularly excited about it and the markets are starting to cool off and people are openly talking about there being an AI bubble and what they mean by that is that the AI companies are hugely overvalued or the perception is they’re hugely overvalued their share prices are way too high not one of the companies who are purely AI and we’re talking open AI and anthropic really here and a few other small players are in any way profitable or even close to it they have sucked up literally trillions of dollars of investment and created huge amounts of infrastructure bought hundreds of thousands if not millions of GPU chips which are the bits of computer hardware needed to efficiently run these models and they just simply haven’t translated that into a commercially viable business and the same goes for the major players who are trying to monetize generative AI the likes of Microsoft and Amazon, Google and so on simply have failed to meaningfully translate the AI elements of their product portfolios into profitable things now the AI bubble is very bizarre because usually to justify high valuations there has to be some clear path to profitability and thus far after several years of rapid evolution of these models still there’s no compelling consumer or business use case from a financial perspective and I think that the markets are only going to stay along for the ride for so long and at some point the share prices of these companies are going to collapse and several of them are going to go completely out of business which is going to send ripple effects through the global economy because something like three quarters of the growth in the American economy the biggest economy in the world by some distance over the past few years has been driven purely by technology companies and particularly around the prospect of generative AI so at some point that’s going to collapse the growth in the American economy will be taken down with it and since these companies are now borrowing money to drive their insatiable need for energy and hardware they may well default on debts which will further impact the wider economy and could cause something resembling a recession or a proper financial crash in the worst worst case scenario I’m not suggesting that’s inevitable and so there’s going to be a showdown at some point with reality the other side of this there will still be the models there will still be the technology the technology will still exist and the hardware will exist to maintain, grow, retrain and continue to deliver these models even if for example the price is likely to go up and the speed might go down but these things aren’t going away let’s be clear if the absolute worst of the most cynical pessimists case scenario the most AI hating people if their worst possible most cynical state of affairs happens we will still be left with the models we will still be left with the algorithms that created them the technology and the hardware to run them these things are not going away now the landscape might change who uses them and how they use them might change the investment might shift to other aspects and let’s be clear that these things do some amazing things for example the healthcare use cases are really quite compelling in terms of augmenting humans in the healthcare arena these things are incredibly powerful and will save lives so we don’t want them gone we just don’t want them destroying the environment increasing societal and economic disparity replacing people and making people jobless and perhaps even homeless or certainly in financial difficulties simply because greed the billionaires want to have slightly cheaper workforces so any number of things could come out the other side of this but what’s not going to change is the existence of these models again the constraints around things like copyright and plagiarism again lots to play out there but still still still still still these models will remain they will still be here they will still be doing things that may or may not be useful they will still be getting things wrong and people will still be using them and people like me and I don’t want them to go away I find them a really valuable, really exciting and interesting creative tool that I do not use AIs to create artworks in their own right I do not use them exclusively I do not take the output of generative AI and pass it off of something that I’ve created when I label up my artworks, if there’s an AI element I will add that as just one of the things like when I say that this is acrylic and mixed media on canvas I will say acrylic AI mixed media on canvas and I’m not going to shy away from that or apologize for it as I said earlier I will accept and engage in meaningful debate about the validity of what I do and the potential ethical and moral implications of it and I’m an open mind but I’ve got an open mind with regards to the fact that I consume meat which from a moral and ethical perspective is still dubious to me but I also really like meat and I recognize that that’s problematic and I’m open to people having different opinions to me on these things and maybe one day I’ll go vegetarian or vegan I’ve tried it before to obvious limited levels of success but you know this is all a journey and we are who we are so I don’t really have a conclusion to all of this I didn’t really see this as well as my usual here’s a point that I’m making here’s what I want you to take away from this I think this is more general outpouring of my current point of view on the generative AI arena and I’ve absolutely got a lot more to say on this and again if you’re interested in the more sort of business and technology focused edge of this then go listen to my other podcast The Confusion Matrix, wherever you get your podcasts where you’ll get a lot of my thoughts on this stuff but are relatively little if not very little reference to the creative realm this will not be the last time I monologue on this stuff I’m almost certainly going to stop recording in a minute and then think oh no I meant to say this, I meant to say that, I meant to say the other or fine when I edit this down a little bit that I’ve probably said a bunch of nonsense but if you’re listening to this I decided to publish it but yeah I hope that has been of some interest or value to you things are going to keep changing rapidly if anyone tells you that they know what’s going to happen on this front they are either an idiot or a liar the one thing I can say is that a lot is going to change so if the next time I do one of these it’s in a year’s time I fully expect to be saying wow that was an interesting year here’s what’s changed and it’s huge and that will also be affected by the wider social economic and political landscape of a very very weird and somewhat scary world that we live in that’s all I’ve got to say, I’m not going to say anything else and I’ll be back at some point with something else bye