Subscribe
Share
Share
Embed
IIIMPACT is a Product UX Design and Development Strategy Consulting Agency.
We emphasize strategic planning, intuitive UX design, and better collaboration between business, design to development. By integrating best practices with our clients, we not only speed up market entry but also enhance the overall quality of software products. We help our clients launch better products, faster.
We explore topics about product, strategy, design and development. Hear stories and learnings on how our experienced team has helped launch 100s of software products in almost every industry vertical.
Here we are again with another episode of AI from the Make an Impact podcast. I've got Brinley Evans, Joe Craft with me today again, and we've got some interesting topics to talk about today.
Speaker 2:Looking forward to it.
Speaker 1:Yeah. Well, we're gonna start off, I think, today is with the kind of the latest news. And, you know, obviously, the biggest news for this past week is the GPT 4 o release as well as the Google IO where they showcase their their latest AI. And, you know, I think just to point out some really interesting points that came up was around some of the real time translation, some of the emotions that were behind that, some of the face detection around what they showcased, understanding the code and showing off actual code, snippets, and giving some kind of voice commands by debugging it and as well as visually understanding images around, you know, showing, like, data visualization charts or graphs and then actually deciphering what that means. I don't know if you guys had any thoughts on that.
Speaker 1:I I know I for the translation, I actually tested it with my wife who's Lithuanian, and I actually have that I actually have them real time translate what I say in English, and then she says something in in Lithuanian because we're actually going to Lithuania this summer. It's been really tough training, you know, training or teaching our kids to speak Lithuanian. So her relatives, like her her parents and grandparents don't speak English at all, so or very little. So it's been it's really cool to see if that can actually work and have conversations. Exactly.
Speaker 2:I mean, did just the general, eye kind of reveals. We always find, you know, we work with the OpenAI models quite a bit. And every release, you're just like, wow. Really? It can do that?
Speaker 2:You know, we're we're sort of so up to speed with what the previous model can do, and yet the jumps are so big. And I think we're seeing that with with so many things in the industry, the speed at at which it's moving. You think surely they they can't have that already, and, yep, it is. So it almost forces you to continue reevaluate what is possible and what, you know, what will be possible in not, oh, in 5 years time, I'm sure we'll be able to do that. It becomes more of a question of, right, it looks like in 6 months' time, we'll be able to do all those things we wanted to do for years.
Speaker 2:It's quite exciting.
Speaker 3:I think one of the the cooler parts of that demo, I mean, apart from the real time nature of it, you know, which is amazing, is just the inflection that's come in that wasn't there before. So all previous kind of interactions with GPT 3.54 has always been text based. Or if you actually like talking into, like, your phone and and getting a response, you can it's basically you can tell in the background this is converting what you're saying into text, still passing it through to the model, and then you're still getting a text reply back, which is just read out to you. But what was, really cool in that demo was the inflection that was happening there. So I think at one point, when the presenter sort of said, okay.
Speaker 3:Reply in a very sarcastic tone, and there, I would be like, wow. I'm so happy to be doing this. This is great. And, like, that's like that really blew my mind. It's a it's a small thing.
Speaker 3:Like, I don't think you'd even notice it if you weren't aware of it. But, you know, understanding human inflection and tone for something, it's like even understand sarcasm and understand, like, not just the words that are coming out of someone's mouth, but the way that they're saying them and the inflection behind that and what that means and is able to respond with that understanding. That, to me, was huge because if you're having a happy conversation, like, the, you know, the model was replying in, like, a happy way too and laughing with you and, like, you know, making it like a fun event, which is really blew my mind because that's, you know, translating from just something very intelligent to almost very human. So it's like another level of with the chat, which was, yeah, incredible set. Really blew blew me away and keen to see where that goes.
Speaker 1:I was I was going to actually potentially cancel my GPT subscription because Claude was doing so so well. I was like, GPT's been it's been slowing down. But what some of the things that I heard was interesting were there's the Turing test, you know, where if you can trick human beings into thinking it's a it's a real human, then that's passed the Turing test. And so, supposedly, it was tested, and it passed it 50% of the time. And so there's there was some debate I saw on Twitter, Rex, where they were saying, well, it's not, then it didn't really pass it, but then I forgot who was arguing it, but they definitely said it was well, if it does it part of the time, then it actually did pass it.
Speaker 1:Just one time is passing it. So that was really interesting. And then with that, yeah, with the same article they were talking about where, I think the developers of GPT actually there's some kind of, website that where you can test it against other AIs, and it completely blew everybody out of the water. They didn't really say it was GPT 4 o. They just said it was, you know, one of the open AI developers testing out a bot, and it was leaps ahead of anything else that was close and surpassed everything.
Speaker 1:So that was really interesting to to hear about.
Speaker 2:That is just the speed as well. I mean, latency, we've always been aware that's been a problem. And, you know, especially in in business applications, you want quick responses. So just working with it for the past week to see how quickly it can come back with complex answers is another massive jump. That was really what I found with the conversation in, just GPT 4 is you'd ask it something, but there would be that delay, which is a little bit like you can feel the processing and you feel that disconnect.
Speaker 2:Whereas I think what connects us as humans is we process things so quickly and you respond immediately. So you expect that. And that's something that lends itself to some believing something is more relatable. So if you have a quick response, it feels like you've been heard and someone's answering you. There's waiting and thinking, okay.
Speaker 2:Now waiting for it to go up to the cloud, be processed, come back down, makes a big difference.
Speaker 1:And I don't know if you saw the the Google IO where conference where they're talking about, you know, obviously with Soarer, where that's a competitor where the you can do text to video, Google had the same thing. So that should get pretty interesting when making movies and doing those type of things where I know that's such a big controversy with the acting industry of taking people's likenesses and voices and making movies from that. So where that is being taken should be pretty pretty fascinating.
Speaker 2:Will be. I always joke that and we probably even mentioned it on a previous one, but in a way, you literally will be able to, you know, ask for the type of movie you wanna watch. Like, I wanna see this actor or this actress and, you know, include my son in there. You know, it'd be fun to see him in an action role. Mhmm.
Speaker 2:And, right, you know, and you literally would be able to do something like that and watch a feature length that's generated on the fly. And is this 10 years? Is it 15? Or is it 5 years' time? It's interesting to to see where it's gonna go.
Speaker 1:Yeah. I think the other 2 that was, interesting was the ability to show, like, maybe something not working, like a product not working and recognize what it is, some ways to solve it. So I'm curious to see how much vaporware that it actually is. You know, they showed it, I think, on a, like, a pot or a pan demo saying something wasn't working. It was able to detect a brand and give some suggestions.
Speaker 1:And then, you know, Gemini being integrated within Google products, suite of products, and then cross collaborate collaborating between the products. So, you know, add some receipts to, you know, some Google doc or Google folder, and then you can calculate it into Google sheets. That should be interesting how that cross pollinates between the products of Google.
Speaker 2:Yeah. And just integrating it into to products as well.
Speaker 3:I hope this isn't where the topics that you guys are gonna talk about. Let me know if it is. We can go back to it. But I was watching an interesting video around generative AI and some studies that were done in terms of, you know, the complexity and how smart it can be, I guess, in quotes, compared to where they can see it going. And there's some studies of which are showing that generative AI is kind of peaking to a point where, you know, it's not really gonna get that much smarter.
Speaker 3:The amount of processing required to get it, you know, a little bit better is, like, you can only throw so much data at it, and the more data you throw isn't necessarily going to make it smarter and smarter. It's sort of incrementally expanding its knowledge base, but it's not necessarily going to get better answers or or do better. So it's a bit of a different perspective because at the moment, we're sort of, like, open eyed because we've seen how fast it's come within, like, 2 years. And it seems like it's on a huge trajectory of just, like, boosting upwards in terms of, like, you know, we're saying, like, who knows what will be, like, in a year or 2 years. But the at least the study is saying, like, we're kind of at a point where it's as smart as it's gonna be.
Speaker 3:And anything new that comes out of it is just us improving the tooling around that model. Model itself or is any capable of still doing so much. So it may be able to create videos, but those videos will still be, you know, you know, kind of you can see they're very easy AI generated, and there'll be artifacts. And the the writing behind it won't be too good and look a little silly and, like, we're really improved beyond that point.
Speaker 2:I I haven't seen anything directly with that, but I think it ties in something I'm I'm gonna chat about later with, small language models and just some interesting approaches with that. But also then you look at how much more do we want from things when, you know, they've been trained on almost anything. Even at the moment, they can arm kind of almost any question. And if there's comprehension and there's you know, there there are a number of other factors, but then can't those be, augmented by sort of tying them to different automation software or, you know, at what point do we need any more except in terms of the sort of responses and that kind of fine tuning that we're talking about. Interesting to see.
Speaker 2:And also a different model, not a kind of large language model, but a neural network or, you know, that that's then tied in or I don't know. It is an interesting point though.
Speaker 1:Yeah. I think I think it's pretty restrictive to think that way. It's almost like you're teaching a little child and they're growing up at the moment and, you know, it's just learning and to make it this far this fast can only increase it. In my opinion. You know, me studying robotics, you know, a long time ago, I was studying I took a neural networks class, and it was just fascinating to see the mathematics behind what it takes for a brain to learn something and the feedback loops of training that.
Speaker 1:Just touching on that, you know, I'm no genius at that at all. But just to see that and getting a taste of what the mathematics are behind learning back then 20 plus years ago and where it is now, and the hardware behind that is also increasing and the speed in which it's increasing is just I think it's just going to keep moving upwards. There's no way it's going to plateau. I mean, maybe it might slow down in certain respects, like right now, oh, wow. We we waited what what was between GPT 3 to 4.
Speaker 1:You know, the timing, it wasn't that bad it wasn't that bad as far as, like, the length of time, but what we see between it is pretty fascinating. I think what I'm worried about is the amount of noise that is is being generated because it is being trained. So there's noise, you know, garbage in, garbage out kind of thing. So somebody has to be there to check it and make sure it's really right. But also the evil, you know, with great power comes great responsibility.
Speaker 1:So everything is gonna start picking up as well as cyber hacking as well, I think. And anything else, you know, trying to hack into using people's voices, using people's look alikes to say, hey. I'm this person. Let me can you send that money to this bank account? So those are the kind of things that I think will also increase as well.
Speaker 2:Yeah. Absolutely.
Speaker 1:Yeah. I think it's a good way to say segue into what you wanna talk about, Brinley, is the small language models.
Speaker 2:Yeah. But this is a kinda news item, but I started unpacking it, and it really became a lot more. And I thought, I'm not gonna I'm not gonna mention too much about it and get both of your reactions because it took me by surprise. I mean, we used to have large language models. They're trained on everything.
Speaker 2:You know, they consume a huge amount of information, everything they can you know, sort of be trained on that exists almost. And, you know, that forms the basis of this model that you can interact with and and ask questions. So, you know, they large language models, they excel at sort of complex reasoning over these vast datasets. So if you look at, you know, something like legal or even medical or pharmaceutical where there are peer reviewed journals, there are white papers, clinical trials, reports, you name it. It can go through all of that and does an amazing job of being able to compare data, understand your intent and actually fetch and, you know, the pieces of data you need.
Speaker 2:So way more powerful than sort of the human brain is, you know, is able to do and can comprehend really so much more. But you still send your data over to them. Their responses stream back and there are latency issues. They need to be hosted and they need to have specific hardware. So there are limitations in that even with 4 GPT 4 o.
Speaker 2:We've seen much better latency speeds, but it still needs to go out. You need an Internet connection, that sort of thing. So Microsoft recently announced the fee 3 or was it 5 3 or 4, fum? I don't know. But family of open language or small open small language models, SLMs.
Speaker 2:So these are really the sort of most capable and cost effective models for this size. They have incredibly low latency. And what sort of struck me is almost this evolution that we're seeing now where, you know, like we did in web development, you know, you had a large sort of application, you know, that was compiled from a big data source. And, you know, then we started going into client side frameworks with components. And, you know, everything was sort of these bite sized of of applications, which had all those performance improvements and number of other advantages.
Speaker 2:And it's kind of doing the same thing. So what we would look at is we've got these smaller models that are effective at doing specific things. And you would actually have a whole portfolio of different models. This is kind of coming back to what you were talking about is, you know, are we going to plateau? And maybe the answer is yes, we'll plateau with large models, but it's about capitalizing on these sort of smaller models and having a portfolio of different sort of disciplines.
Speaker 2:So one thing to sort of unpack on these is well, before I go into that, the the other point that was interesting is these small language models can often out of the large language models on specific areas like language or coding and math. So you've got these, you know, these sort of giants that are incredibly good at a lot of different things, but they haven't been down. So that was an exciting point that I thought, you know, that you actually get better performance. But one thing that's also and and this sort of goes into the discussion is the parameters on each model. I didn't know a lot about parameters.
Speaker 2:And from the research, it seems fairly complex. I don't know, Makoto or Joe, if you know much about the parameters in a model. But I have a good analogy if you don't and for our listeners. So I'll go into that. So to understand the sort of model sizes, we look at the the parameters.
Speaker 2:So you can imagine analogy we could use is kind of baking chocolate chip cookies. So imagine you have a recipe book for making cookies. So each recipe has a list of ingredients. So you got things like your flour, your sugar, your chocolate, and you've got specific amounts of each ingredient. So in this analogy, we would say, alright.
Speaker 2:Our parameters are like the amounts of each ingredient in the recipes. So is it a cup of flour? Is it 2 cups of flour, half a cup of sugar? That sort of thing. The model is like the entire recipe book.
Speaker 2:So it's got all different recipes in it, or you could think of rules for making different kinds of cookies. And those are kind of your responses or outputs. So when you're baking or you could say we're training the model, that's an analogy of baking, you can adjust those the amounts of those ingredients or really the parameters to get the best tasting cookies. So really the most accurate responses. So the better you adjust these amounts, the better the cookies or the output will turn out.
Speaker 2:So it's an interesting way of understanding. So in kind of more complex recipes, larger models like GPT 4, you might have many more ingredients and need to be more precise with the amounts. Just like larger models have more parameters to fine tune. So that's sort of an overview of what the parameters do. So now you see where where PHE3 minutei comes in.
Speaker 2:It comes in with 3,800,000,000 parameters. So it sounds like a lot, but you're looking at a model that's about 14 gigs in size. So pretty manageable with most storage on a lot of devices. That's something you could fit in. Now you look at GPT 3, which is supposedly a 175,000,000,000 parameters.
Speaker 2:So you've made the jump there from 3,800,000,000 with the V3 Mini to GPT 3, an old model, 175,000,000,000. So you're looking at a 650 gig model. So then we go to the the GPT 4, and that's rumored. I say rumored because I think it was Sam Altman who was like, no. It's BS that it's 1,000,000,000,000 parameters.
Speaker 2:But a lot of things online are saying it's it's round about 1,000,000,000,000 parameters, which is 3.6 terabytes. That's massive. So that's not now we're purely looking at storage of the model. We haven't looked at any of the the processing or anything like that. So GPT 4 o is still to be confirmed, but you've got massive size differences there.
Speaker 2:And this is where I find it exciting. So the advantage of that small size is that these models can be deployed on devices without network connectivity, which is massive. So imagine you have something specialized in teaching or mathematics or anything like that. You can have it on your phone. You don't need an internet connection and it can still do the tasks it's trained for.
Speaker 2:So maybe that's translation like you're talking about earlier, Mikado, anything like that. Whereas, they would be able to do that with a large language model. So then we look at things. Yeah. This it is.
Speaker 2:It just kind of opens up the possibility of what you can actually do and what you can build on. And what they've really achieved is is high level of accuracy by finding new ways of training and scoping the material to kind of ensure the quality and condensed content as well, which is really smart. And that's kind of the refinements that I think we'll see. Like the plateau you're talking about. Yes.
Speaker 2:We can probably only get models that'll be, you know, a certain size. So let's say a few terabytes big, but, you know, can we then break down? So you have a, you know, a medical model and a pharmaceutical model and a engineering model and know, mathematics and all those things in there tweaked for sort of efficiency in those. But you also looked at some of the other things. You said, well, even if you could fit a large language model, your phone's gonna run out of battery very quickly, and it just doesn't have the the CPU or GPU to even handle the the processing.
Speaker 2:So that's where it's smart as well because the energy efficiency of those small language models, they they need a fraction of the processing power and energy requirements to actually run them. You know, we look then at kind of, well, what do we see if we project out? Like, what would would these small language models be used for? And you could think things like like we mentioned, language translation, image recognition, tutoring, all things like that. And then you also have the possibility of hybrid.
Speaker 2:So imagine you have this local this local small language model on your phone, but you have something that's out of its scope. And it can just farm out over a Internet connection and go get you a much more detailed response if that's what you need. So pretty interesting how, you know, how this will work out. And I guess we could then ask, well, what does this really what does this really mean for us? The first thing I thought is I don't have to grumble it anymore for not understanding a question.
Speaker 2:I'm really looking forward to it because I'm really tired of that. Sorry. Didn't understand what you're asking. So I guess we're gonna get really capable virtual assistants that are gonna really sort of, you know, update. I mean, I'm sure the updated Siri will use something like this soon.
Speaker 2:Any other kind of virtual assistants. So and then other things would be like yeah, car entertainment systems where your dictation is going to be kind of accurately captured, just so important. And then wearables, teaching aids, those sort of things. So what really started out as sort of, small language models, what would you need those for? Really sort of opened my eyes in terms of, wow, that is a really valuable alternative to have.
Speaker 2:And I'm thinking even what we're working on, it's you know, you start to think, well, maybe we could have an application that has one of these models built into them because you get 90% of what you need potentially, and that would be enough.
Speaker 1:Is this where we're talking about before? Or may I mention where, you know, the large companies like OpenEye, you know, Google, whatever, those are going to start falling not on the wayside, but they're not gonna that's not where the business is at. It's where you have these specialized AIs built for different companies, and you have these little niches of AI is built in it. And if that's the case, it's funny too because and this is for sci fi nerds out there that like to read some of the sci fi stories. One story is called expeditionary force.
Speaker 1:It's by, Craig Allison. You've gotta read it. There's this AI called Skippy. I think he's actually appeared in certain video games with really smart ass AI that was developed by some, you know, futuristic alien species or whatever. But the human find a human finds it, and it helps it, like, kinda, you know, beat up the big bad aliens.
Speaker 1:But long story short, it's pretty funny where this super intelligent, you know, massive AI builds these little tiny AIs to control missiles, rockets, lasers, whatever. And he has arguments with the little AIs that he creates where they band together and wanna have a committee about how they're gonna blow up an alien spacecraft. And it's kinda funny because that's, like, the major AI having these little AI, you know, things that he's created, and now it's causing problems because they're becoming smarter. So just a funny thing.
Speaker 2:It is. It it just it surprises me as well just looking at a lot of sort of sci fi references now where you watch a a good sci fi movie that was made even a couple of years ago. You're like, wow. This is supposed to be 20 years in the future. The AI is pretty basic.
Speaker 2:You're like, we've got way better AI now. So a lot of those fall flat now, and you think, wow. I don't think anyone predicted this sort of growth that we're actually experiencing.
Speaker 3:Kind of interesting too just thinking of specialized AI models. I've always, you know, had the back of my mind. We want to get to a point where you have assistance that can do things for you. And so say to it, like, you know, book me an airline ticket to Canada for the 20th of next month and find me the best rates. And, you know, there's a few steps involved in that.
Speaker 3:I had to, like, you know, find the airline service. Maybe it books you a taxi there. Figure it knows already that maybe you're a vegetarian, so, you know, preorders your food in the plane for you. Maybe a plan to trip a bit better for you, you know, rental car for this day, books you in Airbnb, all these sort of separate tasks that it needs to do. But in order for it to do it, I've always sort of had the idea in my mind that, you know, it would sort of farm out those certain specific tasks, each one of those steps in that process, in this case, 2 different models.
Speaker 3:Right? So, you know, okay, this is the booking model for, you know maybe even more technical. This is the the model that understands APIs really well. I want to interface with the Airbnb API, and this model is specialized in that. So the guy's gonna grab that model, make use of it for that specific task, and then it sort of is done with that model and discards it.
Speaker 3:Okay. Next step, I need to, you you know, do something else technical, like maybe work out, you know, how much this is gonna cost. Okay. Let me get your finance model, load up all your finance data into it, and, you know, and try and figure out what the billing is going to be for you and how much this is gonna cost. And, okay, now we've got a total for that and sort of forms out all these different kind of specialized tasks, not just to one big GPT model.
Speaker 3:Like you're saying, it's, like, rather specialized individual ones. It's not really the same as what you're saying. You the benefits to what you're focusing on is really still accuracy in a small contained scope that you can have in your phone. So but similar design. Right?
Speaker 3:There's microservices for AI where you have some specialized sort of, you know, points and everyone knows, oh, well, this is the best API model and, you know, definitely use that one if you wanna communicate to those points and things like that. It's just an interesting sort of perspective to
Speaker 1:it too.
Speaker 2:I think it is really interesting. And and also, you know, how accessible are these SLMs going to make AI as well? Because not everyone will necessarily be able to host a large language model, even if they're open source. You still need pretty decent sort of servers to and processing power. So the smaller ones, you'd think more people can host them, potentially more people can train them as well, but easier to train.
Speaker 2:You don't need the same resources to do that initial sort of
Speaker 3:Yeah. Definitely fine tuning, first of all.
Speaker 2:So I think there's a lot you know, imagine everyone is empowered to to create these different models. Some of them stand out a lot more than others. And certain people take different approaches. I think, you know, if this is open source, or at least I'm sure it will move certainly in the sort of open source realm, We'll just see a lot more innovation when people can actually start training their own models and have them, as you say, specifically for things that they're good at. And I mean, how different is it really if you went to buy a car to you speak to the salesperson, and then you go to the finance person and chat to them.
Speaker 2:And then maybe you go to the service center and and pick it up and chat to someone there or whatever the case is. But, you know, there's all sort of these We arrange ourselves in kind of, you know, society has, you know, our kind of independent sort of expertises. Maybe it'll be pretty similar to that. So
Speaker 1:Well, you know, just to I'm sure you've guys seen her movie. It's Joaquin Phoenix where he's talking to the AI and it's his girlfriend. So I was driving home and I was playing with, GPT 4 and and having conversations with GPT 4. I'm gonna say her because I have the female voice. So it was, for 25 minutes was a pretty fun conversation.
Speaker 1:And, of course, my wife is like, who's that? So there's gonna be more more than just business, you know, applications, but, actually, you know, the Girlfriend AI niche or the wife niche or the, you know, boyfriend niche, whatever. You know? It's, yeah, it's scary.
Speaker 3:Yeah. I think it's gonna become really big. Yeah. The sort of companionship AI, I think, is gonna have a huge market. Kind of kind of, again, I still want to show if that's a good or a bad thing.
Speaker 3:It's good because there's sort of, you know, there's people out there that could really help them just to be able to talk and to be able to, like, discuss things with something better than with no one still wrapped up in their own head. But, again, it's like, well, you know, is this actually a good thing, though? You know, is this, you know, AI just going to agree with them and just sort of, you know, say everything is fine and it's all okay, whereas, like, a human might be like, no. Like, you need to stop, you know, your fascination with, you know, Sam Altman. Like, get over it.
Speaker 3:It's not good to be so obsessed with him. And, like, you know, it'll be interesting just to sort of see how that pans out if it's like a very pandering companion or or, you know, it's still something that can challenge you or push you forward or say, you know, you should leave your data and job and you should take that plunge in certain new career path that's good for you. Like, you know, kind of what humans do. Like, we can sort of recommend and sort of guide people in ways which, you know, they may not wanna do, but we can tell it's the best for them. Whereas, like, can an AI companion do that?
Speaker 3:Or will it just be, like, someone nice to talk to just sort of, you know, makes you feel better? But maybe that in in itself is enough.
Speaker 2:You know, the problem with that is is even if they kind of have a level playing field and it's well balanced, you'll get someone who'll say, I want one that does exactly what I say, so I'm willing to pay extra for that. And, you know, it's you're never gonna get away from that, unfortunately, but I hear what you're saying. The other thing that comes to mind is go for it, because
Speaker 1:I was gonna say there's little there's little levers. Motivational, less motivational, less aggressive, introverted, extroverted. Yeah. You know, smart ass, not really. You know?
Speaker 1:Something like that.
Speaker 2:It'd be funny. You just have the random like, what am I getting today? Random. Yes. That's good.
Speaker 3:Yeah. Really, really sarcastic, mean spirited.
Speaker 2:What's wrong with you today? It's not that I hit the random button this morning. But, yeah, the other thing I think is really exciting is when you look at these smaller models and, you know, could those then be used, for a lot of these in game experiences? Or if you're working on kind of mixed reality and and virtual reality still, amazing for the gaming industry to have these sort of non playable characters that are you could literally have an in-depth conversation. They have a backstory.
Speaker 2:They may be able to, you know, voice whether, you know, their backstory is a happy one or a traumatic one or a you know, it's sort of it's very dynamic in terms of what that means for people experiencing, you know, just engagement.
Speaker 3:I do because I've seen demos of that before. And, obviously, there's simple demos where they've got ChatTBT behind, you know, an AI that the person's talking to. He's, like, in a tavern, and he's talking to the barkeep, and, you know, it's being prompted that you're a barkeep in, like, you know, old, you know, sort of, you know, King Arthur times and, you know, whatever the case is. And it does seem pretty good because, you know, asking questions and you can go, you know, the rally group that was here earlier, like, sir Lancelot kicked them out. And, you know, it is kind of engaging and it's fun because, obviously, you can ask the AI any question, you know, to, like, script it to, like, 4 different kind of questions that you can ask.
Speaker 3:But I do kind of wonder you know, that was a simple example, but to sort of listening to it there as well, now that I know it's an AI, I kind of always almost disengaged in a kind of a weird way because I was like, well, is this kind of GPC 4? Maybe that's it's more I guess, this is more introspective question because I've been dealing with AI so much when I know I'm talking to AI. I have that immediate sort of suspension of disbelief, like, you know, this isn't really a character I'm talking to. This is just AI responses. And I could do that myself asking Tatchapici now.
Speaker 3:I'm not sort of immersed as I would have been maybe if I thought maybe a human came up with a story. I don't know. Feeding that out, and I've come across a few examples with AI music where someone send me a link to, like, an AI generated song. That sounds really good and impressive, and it's got a good beat. But then back of my head, I'm like, oh, this is AI.
Speaker 3:Like, does it really, like, have any meaning behind it? No. It's a strange kind of thought process I go through nowadays when I come across AI. Like, I'm no longer super impressed. Well, I am, obviously.
Speaker 3:Like, this AI song is amazing. I'm impressed with it. But because I know it's AI, I'm just like, I I get taken a step out of it. And I'm like, I don't know. It's it's interesting to sort of the reason I'm going down with that and if you guys ever feel the same or
Speaker 1:That's a really important point that you're making because it's almost like I have conversations with my kids who grew up in a Internet age with a Google age. We grew up in a library age, microfiche. We had to go find our information. And so when my kids tell me, oh, I'm not sure how to do this. I'm like, you have basically an omniscient being at the in your hands where I had to go to a library, look up, you know, the where this book was and figure it out.
Speaker 1:You can figure almost anything out in the palm of your hands within seconds. They're growing up in in that environment to just know it's there. So if people or kids that are growing up with now AI, it's just going to be a natural thing. So that progression where all of us that didn't have it before and now can see the difference, we see that difference right now, but the ones that are growing up, it's gonna be different.
Speaker 2:I I agree as well. I mean, there's also the bias that we've built up from dealing with the early, you know, check kind of interfaces where, again, there's latency which disrupts. And I'm sure in that example that you would have seen, there would be bad latency. You sort of ask the the barber, no. I do something.
Speaker 2:And it's sticking around for a while. You're like, no. This is not. Yeah. So just coming back to what we were talking about earlier where, you know, it's the small the not micro interactions, but the small little effects that they put in like the sarcasm or irony or whatever is added into the conversation that makes you believe it's real.
Speaker 2:And if the answers are engaged at a you know, you look at any basic conversation. You'd want to know you're not just interested in asking someone questions. You're wanting to interact. So there needs to be a bit of interest from the other person maybe about what you think or and there's a sort of back and forth. And I think as soon as the balance is fine tuned to get that, we're going to lose that bias that we've built up on.
Speaker 2:Okay. I kind of understand what this means. And then you think of the people that aren't involved. You know, I've spoken to a lot of people recently that have sort of said, oh, you know, is AI doing anything at the moment? And it's like, well, no.
Speaker 2:It's really going quickly. So here to show them what exists now, having not seen anything, they'd be like, wow. This is absolutely incredible. Mind blown. And you can see how people make the jump from, oh, surely this is living because it does its job.
Speaker 2:So or at least it's sentient because it's fine tuned.
Speaker 3:No. It makes sense. I mean, what you're saying yeah, what you're saying is really, like, the impression I'm getting from my interactions with AI because they're negative because it's so simplistic. Right? Like, it's really simple responses.
Speaker 3:The barkeep had really kind of, you know, simple sentences that he was saying back to me. But you're right. If he had inflection or, like, a really interesting personality, it wasn't just like, yes. Sir Launcelot was here earlier. He kicked out those thugs, and he was more, like, you know, rambunctious and, like, you know, like a proper actor, then, yeah, absolutely, I'd be way more engaged.
Speaker 2:I mean, you mentioned that if he was sort of, oh, where are you from? And you're like, oh, I've come from, I don't know, over the mountains. You're like, wow. Did you not you know, what did you see there? And suddenly you're like, wow.
Speaker 2:This is building the whole backstory. And Yeah. And, again, it's just going to I think the the industries is gonna affect not just gaming, but just, you know, whether it's AI companions or it's, you know, anything. Even just, you know, I found myself just opening it, you know, opening chat GPT conversation and going, well, tell me about you know, I wanna know how quantum physics works, basically. You know, give me the basic breakdown.
Speaker 2:And then, you know, there's that sort of conversation that that, you know, goes on from there that does a good job of sort of going back and forth and she being really engaging. So fast forward a year, a 2, it's it's gonna be amazing.
Speaker 1:Yeah. Have you seen anything where you can train it, personalities? I mean, can you just put it in can you have it listen to a 1000 hours of comedians, you know, and just make it really funny? I mean, is that even possible to to do it that way? I mean, that'd be hilarious to have a companion that's just cracking jokes all the time and making them laugh.
Speaker 2:Well, I mean, from my understanding and dealing with the kind of large language models, you can sort of you could prompt them, but training is is something very separate. I think that's where maybe coming back to the small language models again, you could do that. You could say, here's a lot of material and I'm actually gonna train it. So that initial intake of all the content. And then from there, you're able to maybe pass it sort of you know, just fine tune it through passing it information.
Speaker 2:But, yeah, I don't think, at the moment, it's that easy to, you know, certainly no one really has the resources to train these massive, you know, models. But I think you could. And and certainly by prompting, you could get a, you know, a lot of the way there.
Speaker 3:Yeah. Prompting will get you pretty far. Yeah. If it's a very good prompt, and that's why prompt engineering is it's kind of like a jerk. Well, I mean, not I mean, it's a joke.
Speaker 3:I mean, it's seen as a jerk because some people think it's just like tapping out 2 sentences. You're a rambunctious, and you're, you know, you're serving King Arthur, and that's what they think the prompt is. But prompting makes a huge difference depending on how well you design it, architect it, and, you know, it'll be paragraphs and paragraphs long and, you know, pages long of a really good prompt, which will really start fine tuning, but really tune sort of, you know, the responses that you're getting back. So, yeah, prompting is way bigger than I think people realize, like, how far it actually gets you. And that's the that's a hard thing to sort of get to know how to use.
Speaker 3:I mean, it's a real kind of niche. It's like, you know, the early days of SEO when that came out, and everyone was, like, scrambling around for different sort of ways of optimizing your sites and search engines. I feel like prompting is going through that right now too where there's a lot of specialized knowledge and tricks, and everyone's sort of knowing how to sort of use these pumps and and generate them. And you can find some really good ones, and this guy's doing it. This one that you can copy in and, you know, get a whole different sort of way of it working.
Speaker 3:And, yeah, it's a really cool aspect, which I'd, like to dig into soon. I I really understand more myself.
Speaker 1:Is there anything else you wanna discuss as far as the the small language models or any other topics?
Speaker 2:I don't think so. I think, for me, that was the most exciting one. And, just that I kinda came across lots of potential. Lots to think about.
Speaker 1:We can kinda end this on what jobs are gonna be taken based on the the latest release or what jobs are being threatened, human jobs. And, you know, I think that's from what? Translators probably have to be on the lookout. Digital visualization, you know, that's gonna be something that I think is at high risk. I I think customer support, that's probably that's been at the top of mind to
Speaker 2:to cheat and say that almost any position using a kind of chat service effectively is going to suffer because teams are gonna be that much more efficient. So if you needed I mean, you can, in some ways, probably go I don't know. Joe, what would you say? But, you know, using chat gpt even for, you know, for all different sort of different functions, you can go between probably 20% more efficient to 75% more efficient. I mean, it's, you know, it's, so that means, okay.
Speaker 2:Well, the opportunities that may have been needed as sectors grow won't necessarily be there. A lot fewer people would be able to do the job of what many could before.
Speaker 3:Yeah. Exactly. I think that's the main takeaway. Yeah. Just the efficiency of teams just goes way up.
Speaker 3:And it used to be 10 people, and I can get away with 5. I think they're from a market perspective, they're trying to, like, decide, okay. This entire market segment of jobs is gonna go away. I think that really comes down to accuracy and how valued accuracy is in that role because AI can still make mistakes and, you know, hallucinate. And so you can't just sort of throw AI on top of a previous job that needed perfect accuracy like a a financial data analyst, for instance.
Speaker 3:He used to, like, you know, bring up big reports and you know, you know, which are very important to have pure accuracy on those. Hey. I can get that make that in this way more efficient, but, you know, you still need right now at least to make sure that it's very correct and very you can't have mistakes in it. So I think roles like that will still be around for quite a while until those models improve at least.
Speaker 2:Yeah. You can't replace the talent. I guess that's the the main thing. You still require talent, but that talent can Yeah. Do a lot more.
Speaker 1:Yeah. I think that's a good very good point because, you know, that that's the double edged sword where the people that are really good and can recognize when something is wrong or right and they leverage AI, they will surpass. They will become the 10 x producer within the organization. But the people who are average and don't use it, they're gonna be obsoleted pretty quickly. And then you've got the managers who are hiring these people who think, oh, I can just get one person to do a job of 10.
Speaker 1:And they hire the wrong person to do it thinking that, you know, oh, I can just save the company a ton of money by just hiring 1 person to do a job. That takes actually more people to do it right. So there's going to be some of that on both sides that's gonna be a problem. But, yeah, the fact that just how many people that I've spoken to that that have not seen it yet, it's it's scary when you show them how accurate and how fast things work at the moment.
Speaker 2:And if you're not using it, start. Like, it's it's it's like you you don't you don't wanna wait because you're gonna be left behind.
Speaker 3:Yeah. It's I wonder if people are apprehensive because of the chat interface and have to, like, type out full sentences, and it feels kind of, like, slow and clunky. I wonder when the you know, as we see with TPT 4 o as, like, you know, the real time kind of voice feedback comes in, if it's gonna have a huge spike in popularity just because it's so much more accessible for sort of the average person on the street to engage with it. Whereas right now, it's kind of, you know, kind
Speaker 1:of like
Speaker 3:a ticky
Speaker 2:thing. It's funny because I kept typing as a barrier to to actually using it. But then you realize some things you don't actually wanna talk about. So if you're looking at code or you're looking at a strategy, you don't really you wanna go into detail. You need a lot of text.
Speaker 2:Whereas I think I kinda find the conversation interesting if you're wanting to learn about a fact or you're wanting to maybe vet something and say, look. I've got an idea. This is the approach I wanna take. You know? Can you see any holes in it?
Speaker 2:Where do you see I could be going wrong? That's fantastic for conversation. I don't think I think there there's 2 completely different those 2 different channels have very different kind of use cases for them, at least what I found.
Speaker 1:So as a user experience designer, the better the experience to whether that interface is voice or is it visual or is it typing, the better. But, you know, with interacting it in both ways, I did notice that through voice, it doesn't pick up pauses. So as I'm speaking and trying to finish a sentence or a prompt, my pause trying to think about it, it automatically reacts, and that gets frustrating very quickly. So I end up typing things that are a little bit more complex. So once they kinda solve that where the the AI detects where you're not quite done with the sentence, lets you pause and finish.
Speaker 1:I mean, it does pick it up. You can just say it again, but it does get frustrating when it does it a few times, especially in a kind of noisier environment when I'm in a car. It picks up all the other noises, and it it interrupts my conversation and train the thought.
Speaker 2:Believe that is configurable, though. And I know that was one of the the 4 o releases that's coming as well where it kind of picks up on the end of yeah. You're sort of wrapping up. You know? Maybe it's the inflection of, you know, asking a question or doing something, which it is.
Speaker 2:But I do think there even with the the GPT 4, there's a way to actually stop it from answering immediately. I was looking at the settings the other day. So that's but it is. I mean, coming back to the user experience, I think it is I mean, that's nice and simple. And one thing I picked up as well, which was nice is if you're talking about asking for a code reference or something like that, it'll explain a solution and then say reference look at your chat history for the code.
Speaker 2:So, you know, you're not interested in talking about, alright, open this, you know, Openseas to clear your variable. Like, it's, yeah, it's very much okay. You can see an example of what I think the code would look like, but, you know, we're talking about it now. So it's more it's almost recognizing that it's more high level when you discuss something.
Speaker 1:So if you're feeling behind on AI, you have to tune into our podcast, of course, where you're gonna learn a lot more. Yes. Exactly. That's all you need. Everything else is noise.
Speaker 1:But I think with that, we'll wrap it up with this episode. Thanks for everybody for tuning in and, you know, definitely suggest any kind of future episodes, any topics you wanna hear us talk about. We've got a lot of experience with the UX product, obviously integration of AI into different enterprise software and just dealing with making things better when you integrate it into your software products. So with that, thanks for tuning in everybody. Like, subscribe, vote, whatever you need to do to to help the channel out, and, appreciate your time today.
Speaker 3:Thank you. Thanks a lot.
Speaker 1:Bye. Bye, everybody.