[00:00:08.880] - Gordon
Hello. My name's Gordon from Drayson Design.
[00:00:11.220] - Taylor
And I'm Taylor from The Creative Tinker.
[00:00:13.440] - Gordon
And welcome to "This week with Taylor and Gordon," a weekly podcast and vidcast where we talk about things that have affected our businesses over the last seven days. Welcome back. It's been a busy week.
[00:00:25.360] - Taylor
It has.
[00:00:25.970] - Gordon
It's always a busy week. And we've had some good weather this week as well, which is a little bit unusual. I remember we went out a couple of days ago and it was really bright and sunny. Yeah, really nice. So let's see, is there any news first of all?
[00:00:44.550] - Taylor
Not from me.
[00:00:45.270] - Gordon
Not from me either. There's been lots going on, but nothing of interest to our listeners, I think. So let's get straight into our first and only topic today. We are going to have a discussion today about the state of AI. It's been talked about all over the place, and we thought we'd just give our little take on it. And for those who don't know what we're talking about, we'll give you a little bit of an idea of AI and how it's progressing at the moment, because it is getting really, really big and expanding so quickly.
[00:01:24.700] - Taylor
In lots of different ways. There's AIs for nearly everything now, and people have probably used AI to build the AI tools. But we started off with images and text to speech, things like that. And then obviously, we've moved into questions and answers type thing with chat GPT where you can pretty much ask it anything and it has an answer for it, whether it be maths question, whether it be code, whether it be asking just information like you would a search engine. So yeah, there's lots of different use cases and things that it's been used for.
[00:02:08.250] - Gordon
So let's break the AI, the big AI word, and break it down into different points, and we'll talk about each one as we go. I should mention that the reason that we decided to talk about this is because of a new AI that we have found for the software that we use to do the transcriptions for all of our episodes. So we basically upload the video and the software transcribes everything. We then have a real person who goes in and checks it all. And some of the mistakes and errors are quite funny, aren't they?
[00:02:50.780] - Taylor
Yeah, the translations.
[00:02:52.350] - Gordon
What they think we've said is very funny. And they've got a new feature which is called "AI Assist." And one of the bits that they can do now is chapter markers. So what it'll do is it'll take the whole of the transcription, and it will look to see what we're talking about, and it will create chapter markers in our videos for that topic. And it will give you time codes. And we've actually added that to all the episodes going back to Episode 100, because that was like 16, 15 episodes and that was all I really wanted to go back to because I could spend a ton of time on it. But I think it's the more recent episodes that people are going to want to look at that chapter feature. And if you go on to YouTube, there'll be a little chapter that will show you what you're listening to at that point, in the part of the episode. And that's a really clever feature. And if you think about what it's doing there, what it's doing is it's basically just analysing the whole text and working out what we mean. It's looking at the meanings behind everything. So that's how we got into this topic for today. So let's break it down into... Let's see, we'll break it down into the AI that we discovered and the point at which we discovered them. So let's start with AI text. So it's things like Jasper, things like AI writing, where it will listen to a prompt that we give it. Everything is to do with prompts. Prompts is what AI live on. So the first bit that we found was the text writing prompts. So basically, you would say, "Give me 10 bullet pointed ideas for blog posts on the topic of building a website," for example. And it would then go away and it would come back and it would list out 10 points. And then you could go and you could use another prompt and say, "Write me a blog post all about..." And then you could give it one of the prompts it's already given you, and it would then write you a paragraph, two paragraphs. And as we've gone on and on, it's going to be more and more paragraphs till you get to a long form article, which is where we are now. Let's just look at that first of all. It needs to be trained, right?
[00:05:32.340] - Taylor
Yes.
[00:05:33.270] - Gordon
So the way it's been trained is on models. And originally, it was being trained on data up to 2019, which is when it started to be becoming slowly a niche thing before it became mainstream. So lots of techy people were finding it. And 2019 was where the data was up to. So do you understand how the models work, or how it's listening to the data that we can explain to people?
[00:06:06.110] - Taylor
Well, they get fed into... Obviously, you need very very powerful computers to obviously analyse them. And I watched some videos on someone doing games and things like that. And basically, you basically just keep running simulations and things like that, or in terms of a game, you're running simulations.
[00:06:28.040] - Gordon
And it's learning.
[00:06:28.880] - Taylor
Yes. But you're giving them points. So in terms of the game for the AI, it was getting given points if they did something correct. And obviously, as you go on, and so how they are trained is, if it gives you a piece of content back that you've... So you give it a bunch of information and it then spits back random words, random sentences, you then obviously say, "Well, this doesn't make sense." So it goes back and it comes back with a newer version. And so this is how you have iterations of AI and how it improves over time is that you keep giving it feedback, basically, saying, "This is good. This isn't good." And over time, it learns. It goes, "Okay, well, last time I did this, you gave me this type of feedback and said it wasn't good. So I'm not going to give you that. I'm going to change it and give you something else." It still could be not good. But I've at least learned that what I gave you before wasn't good.
[00:07:24.210] - Gordon
And then what it does is it grabs text. Like Guttenberg is the library of out of print and out of copyright books. So you can feed that all into the AI, give that a model. It now has an idea of sentence structure and paragraphs and when to put punctuation in, et cetera. And then you add in a load of other articles and things from newspapers. So it's got data, it's got information. It knows about all sorts of different things. So when you ask it to write something for you it's not copying text from a particular place. It's using all of the information that it's learned to then generate new copy. Now, the problem we've always had with writing articles, for example, is that you have to fact check it, because it doesn't always get its facts correct. I asked it to write me some description of Gordon's Magic, and there was a little bit on the web that it could pick from, but it also had learned about other magicians, and it gave information like, "Gordon's Magic has been going for four years," which isn't true; it's more like 30 years. But you have to fact check it. That's really difficult to say. And make sure that what it's giving you is a good starter point. You don't want to ever copy and paste a whole article and just publish it, because it's not going to do you any favours because it might be giving incorrect information. More importantly, Google is now starting to learn itself what is being created solely by AI and is downgrading that content. So just copying and pasting just to get your number of blog posts up, for example, is not going to do you any favours. You'll end up ruining your credibility. But the AI writing is getting really good. I mean, Jasper, for example, is one of the big ones out there, and it's quite expensive, but it's also very good. You can give it very little information. You can treat it a little bit like a conversation. We'll come back to the conversation part again in the future where you can just ask it to redo it, but focus on a particular point that you want to make. And this is the first introduction I had to AI. The second introduction I had was Dall-e. So Dall-e is the Open AI's text to image AI. And I basically just came across a website and it had this description and video of someone typing in, "make me an image of an avocado chair." That's the one I remember. And it came out with this image that looked just like an avocado that could be used as a chair. And I was thinking, "how is this possible? I don't understand." But again, it's all to do with the models. And the models have been trained on lots of different images as to what something is. And it's got a name, and I just cannot remember what it's called at the moment, this type of process is called... You tell them a bit about what you know about this, and I'll think about it.
[00:11:09.570] - Taylor
So Midjourney is another...
[00:11:11.480] - Gordon
Midjourney is very good.
[00:11:13.190] - Taylor
Image, well, brief to image...
[00:11:15.190] - Gordon
Ooh, stable diffusion. That was the word I was looking for. So that's the process that they're using. We'll come back to Midjourney. It's to do with finding a page of noise. So you've got noises, effectively, either black or white pixels, and it's one or the other, and it's a whole screen full of them, and it's all random. And that's what noise is. And it's generating images based on that original noise pattern and then changing the content into what you're looking for. And every derivative of it is getting clearer and clearer until it ends up with an image that fits your prompt. So yes, Stable Diffusion is where I first found this, and that's what Dall-e uses, I believe. Then you've got things like Midjourney, which have their own models. So keep going with Midjourney because it's really interesting.
[00:12:14.430] - Taylor
Midjourney was one of the first good image to AI, or sorry, text to image AI that took the internet by storm, I guess.
[00:12:27.220] - Gordon
It's got a very good style, hasn't it? That's what...
[00:12:30.250] - Taylor
Yeah, a lot of the images or prompts were quite close to what had been requested, but at a high level.
[00:12:40.250] - Gordon
And it's got a fantasy, mythical type feel to it.
[00:12:45.800] - Taylor
Yeah. So you can pass in different parameters. And obviously, there are brief makers out there to help you build up good briefs to get good results. So you can pass in, "I want it to be this size aspect ratio," for example. "I want it to have this type of style." So some people put like Marilyn Monroe type, the artist that did... Andy Warhol, I think his name is, that type of style of an image. And you can obviously pass in these parameters and you can get back a very realistic, in the style that you've asked for, or a realistic type image.
[00:13:26.890] - Gordon
Yeah. And the more information you give the prompt, the closer your image is going to be to it. So you can be really specific. So you can say things like, "I want to have a knight in shining armour with a broad sword. There is a dragon that he is fighting; and in the background, there's a Gothic castle that may have come from Romania; and the night sky is showing the moon with the stars out." And if you were to put all of that in as a prompt, you would get an image pretty much that looks like you've asked for. The less information you give it, the more the AI has to try and interpret it itself. So you end up with images that are very different to what you're looking for and maybe just very loosely based on what your original few words were in the prompt. You can also put negative things in the prompt. So you can say, "I don't want to have a damsel in distress in this image," for example, because it will know that if you've got a knight in shining armour and a dragon and a castle, there could be a princess in there as well. But if you're saying, "I don't want a princess," it won't put that in. Text is a big problem for all AI at the moment. Creating text that is readable, it knows about text, it doesn't understand about words, it doesn't understand about stylized words. So you end up, if it's something that has words in it, like you can ask it to create a logo for something, and it might put some text around, but it's just going to be gobbledygook, basically; text that's placeholder rather than actually means anything. Yeah, you can't read it. It's not readable at all. But you can spend hours and hours crafting the perfect prompt so that you end up with an image that looks amazing. And you can give it styles like you said, you can give it, you can say you'd like a watercolour; you want a line drawing; you want a charcoal image. It understands all of this because it's been fed thousands and millions and thousands of millions of images so that it's learning what everything is. And it's just phenomenal what you can do. I have Midjourney, I have a paid account for it. It's not a lot of money. I think I pay something like $10 a month and I get to get to play with it. And it's getting better all the time. I mean, we are on version five currently for Midjourney. Version four came out like a month ago, and version three was like four months before that. It's like the amount of time that it's taking to get to the next generation of AI is very, very short. And it's getting shorter. Because obviously the AI is learning quicker. Let's move to the next part of AI, which we've discovered, which is the "Chatbot," if you like, the "Smart Chatbot."
[00:16:41.020] - Taylor
Yeah. So we've got a number of them. We've got Chat GPT, which is the one that came out first by Open AI. We also have Bing.
[00:16:50.040] - Gordon
Which is using Chat GPT.
[00:16:52.240] - Taylor
They created their own search engine bot.
[00:16:55.850] - Gordon
Version of Chat GPT, yeah.
[00:16:57.480] - Taylor
And I think other... Microsoft, I think, have released one. Google. They've all released their own versions of it using the same underlying AI / information they fed in, but normally they will have fed in information on top of that. So they use Chat GPT underneath the hood of all of them, but they've all given them and trained their models to work better with whatever the actual product that they're using. So if it's for Google or for something like Bing, they're obviously going to be trained to answer questions in a way that they're going to give the most relevant responses and most likely will give you websites that are most relevant.
[00:17:45.820] - Gordon
Because they're a search engine, that's their business.
[00:17:48.120] - Taylor
Whereas obviously things like Jasper are trained to give you blocks of text that can help you and you can use.
[00:17:56.390] - Gordon
And they have their own Chatbot as well. And what's interesting is when they started, when it first came out, it could answer simple questions. You could ask it to write lyrics for a song. You could ask it to write a poem. You could ask it to write code.
[00:18:12.320] - Taylor
Maths equations.
[00:18:13.460] - Gordon
Maths equations. Yeah. So it knew all about this sort of thing, which has an answer, whether that be a logical, a factual, or a creative answer. We've moved on again, though. We're on the next generation of Chat GPT, which is version 4. And that's now allowing it to search the web. So it can now go out onto the internet to grab data and bring it back and answer your question. There are plug-ins now available. So you can add features to Chat GPT and add to its knowledge base. And that's all part of the paid version again. Haven't looked into the price of that, but I'm assuming it's not going to be a lot again.
[00:19:01.330] - Taylor
I think it's about $30, I think, something like that. Because you have Wolfram?
[00:19:06.190] - Gordon
Wolfram, yeah.
[00:19:07.460] - Taylor
That's a plug-in. Up until recently, it's only had up to 2021, I believe, in terms of historical data. So it didn't know anything about anything that's happened in 2022. So if you asked it who was the Formula One Champion of 2022, it wouldn't know because it doesn't have that data. Whereas with these plug-ins that are coming out, it's being fed newer information and it can stay up to date and things like that, and obviously read on the internet and find out stuff.
[00:19:35.440] - Gordon
So that's an overview of AI. And AI has been used in other places, like in photography. And Photoshop, for example, now has AI plug-ins that will take content - they've had it for a while, actually, as a content aware fill thing, which is a sort of AI, where you have an image and you have lots of people in the background, for example, and you can cut around them and say "AI fill that." And it will look at...
[00:20:01.180] - Taylor
I'd say that was AI.
[00:20:01.640] - Gordon
Yeah, well, it is. I'm not sure it was... It's not as we know AI now.
[00:20:07.110] - Taylor
No.
[00:20:07.110] - Gordon
So you can get the image AI to do what's called in painting, where you can tell it what you want. You can select an area by painting it, and then you can say, "I don't want that, but I do want you to put a cute dog in there," and it will do that. So that's the more up to date version of it. But Photoshop were doing this a long while ago with their content aware fill, where you would just select an area which might have a person in the background and you would say content aware fill and it would just basically look around what was around. So it would try and continue the background as if that person wasn't there, and it would just effectively erase them. I think the Google Pixel phone is doing effectively the same thing these days where you can just use your finger to select something and it'll just get rid of it for you. It's like a magic eraser.
[00:21:01.280] - Taylor
The Apple iPhones can, if you press and hold on a photo, can actually select the subject, whether that's you, basically cut you out just by pressing and holding, and it selects what it thinks is the main subject. So if it's a photo of you, it will select and draw around you. And then you have that as a transparent image of yourself or whatever the subject is, which again, is probably using AI to determine what's the foreground and what's the background.
[00:21:28.740] - Gordon
Yeah. And so AI is being used in lots of different ways. But the three sections that we've mentioned are probably the biggest for consumers, for normal everyday people. So it raises a few questions. So first of all, is AI getting too big too quickly? Because, I mean, what's to come? This time next year, what are we going to be looking at? Is AI going to be taking over people's jobs? Is it going to be putting people out of business? Is it going to create new jobs in different ways? Like every other time that we've had a new technological advancement in the past.
[00:22:14.510] - Taylor
Yeah, I think when computers first came around, people were thinking, "Well, I'm going to have lost my job," because Excel can do the same calculations that someone, a mathematician who was paid money to answer questions and things like that. Then we had Siri and things like that where they were answering questions and again taking away, people relying on their phones more, and so they didn't need dial up or whatever. But with every new enhancement, it means there's more jobs created because people have got to run those things. So I think with AI, it's very much going to be... There probably will be people's jobs that are lost and things like that. But at the same time, I still think that at least in the moment, AI hasn't got the self awareness and control to be able to run itself. So AI can't tell other AIs what to do. And so in that regard, you still need a human. So if you look at it from a manufacturing process, you have all of these amazing robotic machines building cars, building products and stuff in a factory, but they still need someone to press the button to start it. They still need someone to load up a file. You still need someone to maintain them and fix them if they break. And I think we have a while to go before we have robots fixing other robots and being able to figure out what's actually happening because that takes it to another level and another depth of which AI currently doesn't have.
[00:23:49.660] - Gordon
So let's look at jobs that are going to be affected, things like illustrators, painters, authors. There is a completely AI written book that is now in Amazon for sale. It was completely written. Obviously, a person was required there to create the prompts and to put everything in order and to make sure they got the content they wanted.
[00:24:17.600] - Taylor
Upload it.
[00:24:18.150] - Gordon
Yeah, and all that lot. But as far as the writer goes, the AI has taken over that job. So are we going to have a problem with writers being out of work? Are we going to have problems with illustrators and artists who are no longer being required to come and do custom pieces because you can do it with AI now?
[00:24:41.480] - Taylor
I think it will get to the point where you kind of start going back in time. So if you think about music, you had records, you then had CDs, you then had basically streaming and online digital versions, and you've now seen a surge of people that collect records because there's so much, everything's digital, that actually it's nice to have something physical. So I think that...
[00:25:09.850] - Gordon
But also audio files, people who really appreciate the sound, say they prefer the analogue sound of a record, which is bizarre because it's turning it into a digital signal by using the needle. Excuse me. Pardon me. See? Live shows.
[00:25:30.060] - Taylor
I think that in terms of that, I think it may go out of fashion, but then come back into fashion because people were like, "Well, everything is being used. Ai's been used for everything to generate images. Actually, we want an actual physical human to design something else." I think that's probably what's going to happen is that because the computer digital generative images are going to become mainstream and basically the only thing that's going to be used to create art, that actually hiring someone is going to come back into fashion and people are going to be like, "actually, I want to hire someone to do this."
[00:26:06.190] - Gordon
Maybe they'll become even rarer that it's a real person who's created this image and therefore it's going to be worth a lot more.
[00:26:13.930] - Taylor
Well, the records are a lot more expensive now than they probably were when they first were released. And again, I think CDs are in between. You've got records which are some of the oldest, and you've got digital, and they're stuck in between where no one really wants CDs because it's still sort of digital but still analogue.
[00:26:35.280] - Gordon
So the other thing is the copyright of certain things. There's been a lot in the news recently about, "I want to create an image in the style of Andy Warhol." Well, how does that work with copyright? Because obviously Andy Warhol has a style and you're basically copying his style to create your image. So is Andy Warhol the owner of the style and you're the owner of the image? Or how does it work? Are we able to take an image that AI has created or text that AI has created and sell it? Is that something that's possible? At the moment, the answer is yes, but that's probably because no one's actually tested it. There are a lot of artists who are complaining because you can get images in the style of and then put that person in. And they are complaining that there's lots of work out there that looks like theirs, which isn't. And they're not earning a penny from it because the AI has just been trained on their images. And that could be a problem. I think there will be a court case at some point that will determine what's going on. And we may end up being more restrictive as to what we can request. Maybe an artist can opt out and their style will be taken out of the model so that you can no longer have an image created in the style of. What do you think?
[00:28:08.620] - Taylor
I think it's very much a grey area. I think no one knows, it's so early on with AI still that no one knows what it's capable of, what the boundaries are. And I think that you very much have, well, I would say that it is an art form to create a good brief. So then saying that they didn't create it, well, if you were to just put in, "create this piece of art," and you were to use a very basic brief, you are not going to get the same results as someone who has spent time building up a complex brief that gets them a very good result.
[00:28:50.300] - Gordon
So do you think maybe people are going to go to people who are good at creating prompts or briefs, as you call them, and saying, "this is the image I want. Can you make it for me using AI?" So you're effectively going to a third party who is then using AI to get you what you want without you having to learn how to create the perfect prompt.
[00:29:11.060] - Taylor
I think so. I think it's similar to people that come for websites. They don't know how to do it, and they can see what the finished result is, but they don't know how to achieve it, or they can achieve it, but not to the same level of quality. I have a friend who's very good at creating prompts, and I've seen their image and gone, "okay, let me see if I can create something similar." I haven't got anywhere close to the quality of it, even though I can literally see the image that's been generated, and I've typed in literally what the image has in it, and I don't get anywhere near the same results. And they're quite protective over their prompts. They don't want to share the prompts that they've been using because not everyone can achieve it.
[00:29:55.270] - Gordon
It's a skill. Yeah, it's a skill they've developed.
[00:29:58.750] - Taylor
So then, yeah, then you have obviously the copyright of, well, if you've come up with the prompt that generates, and the prompts, if you use the same prompt twice, it won't achieve the same result every single time.
[00:30:09.870] - Gordon
Unless you use the same seed. So the seed is the noise pattern that we started with. If you use the same noise pattern and you use the same prompts, you will get the same result. But it changes each time. You have a different seed, you'll get a different image. And you can keep going until you get one you like. Whatever the outcome is at the moment, for people like us who work in the tech industry, I think it's a really exciting time. It's giving people who aren't very good at writing, people who aren't very good at drawing, it's giving them the ability to create content that they want, and they can keep recreating it until they find something they like. I can't draw, but during the week, I asked Midjourney to draw me a colouring in dog, so a cute dog that a child could colour in, as a drawing. And it gave me exactly what I wanted. And it was perfect for my purpose. I could never have done that myself. So having these tools available to me, and I think that's what we need to look at them as, we need to look at them as tools. Of course. They're not the finished article, you can't just take necessarily what they give you and just use it straight away. You have to use it as a tool, the same way you do a pen if you're writing something. That's just a tool to get the thoughts from your head onto paper. But it is a very exciting time and things are changing very very quickly. If you're watching or listening to this podcast in four months' time, it's currently, what is it, April, early April. So if you're listening to this in a few months' time, chances are a lot of what we're saying is no longer valid because it's moved on in leaps and bounds and is completely different again. It is going mainstream because people can see the advantage of AI and what it can do for them. Any final thoughts, Taylor?
[00:32:16.240] - Taylor
I don't think so. I think we've kind of wrapped them up, really.
[00:32:17.950] - Gordon
So yeah, watch this space. I think this is a topic that we'll come back to because it's very much in the news at the moment. It seems every day there's something new coming along that's using AI. And as computers get faster, more powerful, so AI will learn quicker. And one thing we haven't even touched on is Dream Booth, which is where you can create your own model of your own image, or your face, or your person, or your dog, and then create images based around that, which is something that's getting better and better every day as well. We haven't even touched on that. But that's it, I think. We'll have to stop at this point because we've got AI that needs to transcribe this at some point. So that's it. That's our little look into AI, and it will change and we'll come back to it in the future. But for those who had no idea about it, that's what we think. And if you have a knowledge or an opinion on this, please leave it in the comments because we're very interested to hear what you think about it. And does it affect you? Is it something that you are concerned about? Is it going to make you redundant? You're going to have to find a new job, a new sphere to work in if you like. Let us know in the comments below.
[00:33:47.060] - Taylor
Cool. That's everything from me.
[00:33:48.120] - Gordon
That's everything from me. So thank you for joining us. As always, we can be found on YouTube and on Facebook. Please like, share and subscribe.
[00:33:55.920] - Taylor
You can also go to our website thisweekwith.co.uk to find all of our past episodes and anything we may have spoken about.
[00:34:02.860] - Gordon
We'll see you again next week, 1pm on Friday. Until then, my name is Gordon and I'm from Drayson Design.
[00:34:10.200] - Taylor
And I'm Taylor from The Creative Tinker.
[00:34:11.860] - Gordon
Goodbye.
[00:34:12.790] - Taylor
Bye.
