This episode of Make Me Data Literate features Associate Professor Kazjon (Kaz) Grace discussing the use of AI in creative pursuits, design, and education, and how powerful it could be if we used it to help us think more instead of less.
“I think the tech industry is very good at human focused things. I mean, in the same way that a chicken farmer understands chickens.“
“I do really think that it’s a cool world that we live in. I just hope that experiencing that transparent access to stuff, that cool world, doesn’t become a luxury good.“
“So when we talk about creativity, I think AI can be really useful for individual creativity, in that it can help someone do something that they wouldn’t necessarily have the skills to do, but that if you increase your skill level, then you’re gonna run up against the problem that the large scale generative AI systems are just going to be a regression to the mean, right? They they produce mediocrity on tap. They are infinite extruders of mid. And so once you reach that point, there is somewhat of a a limit on what you can make with them. “
Linda McIver (00:00)
Welcome back to another episode of Make Me Data Literate. I am back on my bad habit of recruiting people I met at parties to be on my podcast. I’m excited about this one. As soon as I met this guest and found out about what he does, I was like, oh man, we have to talk about this on Make Me Data Literate. So this is gonna be a fun conversation. welcome, Associate Professor Kaz Grace.
Kaz Grace (00:24)
Hi Linda, it’s great to be here. Thanks for having me on.
Linda McIver (00:28)
It’s gonna be great. so can you tell us who are you and what do you do?
Kaz Grace (00:33)
Hi, I’m I’m Kaz Grace. I’m an Associate Professor at the University of Sydney’s Design School. My title is Associate Professor of Computational Design and I I’ll definitely get into what that means. I’ve spent about twenty years now teaching people designers, people who are interested in in how to make things, how to understand what people need, how to understand products and services, teaching designers how to think about data, as well as doing research in data and AI and machine learning. I actually started about twenty years ago, I I did in my PhD, I I started getting interested in how AI can be used in design and creativity, which means I’ve spent about fifteen of the last twenty years saying, hey everybody, wouldn’t it be really cool if AI could be used in creativity? And then the last five years saying, no, no, not like that. That’s not what I meant. That’s absolutely not what I was doing.
Linda McIver (01:31)
Must have been a super crunching gear change.
Kaz Grace (01:34)
Yeah, it it I mean, it’s pretty cool to have the kind of stuff that you were working on as sort of the strange PhD student off in the corner doing creativity stuff. I was often thinking, like, maybe I should work on me- medical image detection or something like that. it’s kind of great to have it be part of the mainstream conversation, but there’s really some challenges too.
Linda McIver (01:53)
So before we get into the challenges, which is where I’m spending a lot of my time these days, let’s talk about the creativity side. How did you foresee the possibilities for AI and creativity?
Kaz Grace (02:06)
So I actually my PhD supervisor, John Gero, is a an expert in the field of the cognitive science of design, like the psychology of creativity, how people think when they’re designing, what makes design, whether it’s design done by a software engineer or design done by someone who’s working on a a a mobile phone app or a chair or a table or the workflow in an organisation, like just design in abstract, how is that different from other kinds of complex thinking and open ended tasks?
And so I really started out on the human side, and then drifted into how do we model that computationally, like how can we understand those processes better by building computational models. and everything that we were talking about back then was about how people unless they are very expert usually don’t consider enough alternatives. They don’t iterate, they don’t think broadly, they don’t think about different perspectives on a problem. whenever you’re teaching design, I I’ve had a a a really cool bunch of opportunities teaching design in a bunch of different contexts. And whenever you’re teaching design, students always want to cling to their first good idea like it’s a life raft on the Titanic. They’re like, I’ve had-
Linda McIver (03:23)
Yeah.
Kaz Grace (03:24)
I’ve had a good idea and you can’t make me let it go. And and honestly, I think part of that is because they they’re just not confident in their ability to have a another good idea, right? Like why would you let go of the life raft and and strike off swimming into the distance? but what we teach them in class and what I was trying to use AI for was to help people have more different ideas to iterate on what they’re doing, to consider different perspectives, to reframe, to explore. So imagine my surprise when a whole family of algorithms came along which do the exact opposite and push people to explore less and say, yes, absolutely I can do that whole thing for you right now. You don’t have to think about it even a little. so that that’s that’s where the majority of my challenges start in the creativity end. but there are others.
Linda McIver (04:14)
Yeah.
Yes. I I can imagine we are definitely going to get into that.
I I’m gonna start with the the the starting question before we we get into the the the nitty gritty of the challenges. But what did you have to learn to do what you do? What was missing from your formal education? I imagine starting to do AI twenty years ago, there was not a lot in your formal education that prepared you for that.
Kaz Grace (04:44)
Yeah. Right. So I mean I had done a design degree and I had done a bunch of software stuff along the way. So I knew how to program and I knew how to design. mostly digital design stuff. and so when I started researching, I had to learn a lot of psychology and cognitive science theories really quickly. So that was like a lot of reading and scholarshipy kind of stuff. It wasn’t really until about, I don’t know, two thousand and ten ish when I started needing to learn machine learning. and so I was largely self taught trying to figure out with my relatively meager Python programming skills. Trying to figure out how to throw together different kinds of neural network in the early days of deep learning, experimenting with things that I didn’t really understand, like Boltzmann machines and recurrent neural networks and various different off the wall things like Kohonen networks for self-organization and things like that. And I found myself working in domains where there wasn’t like a lot of good labelled data.
So I spent a lot of time looking at unsupervised learning and reinforcement learning and stuff like that. Things that are super common now, because the scale of compute and the capability of our inference engines is is just lots lots better. but they were weird little ornaments and toys back in two thousand and nine ten. so it’s a it was a ton of fun having to having to learn that stuff. Finding groups of people who were doing similar things. I found myself publishing in venues where there were like as many artists as there were computer scientists, which was super fun. and then drifting towards trying to build things that were a little bit higher scale, like models of curiosity and and stuff like that for for designing meant again, trying to learn the the human psychology on one side and some machine learning stuff on the other. I’d make a really terrible machine learning engineer. Like if you wanted me to make anything that was like maintainable and scalable and secure, then that would be really bad for whatever system you’re plugging me into. But I feel like I’ve learnt how to tinker in a lot of different spaces. And it was really valuable to learn that stuff back before it was popular. Because everyone will try and sell you a six week AI course now. but it was actually it was there was a lot of scrabbing around try to do it fifteen years ago.
Linda McIver (07:19)
I I think you overestimate the maintainability, scalability and and security of current systems, but that’s Yeah. Yeah.
Kaz Grace (07:27)
Yeah. Let me let me clarify. The ones that I build definitely won’t be. I can’t say anything about other people’s.
Linda McIver (07:37)
That’s a problem. How did you what drew you to AI back back when it was not
Kaz Grace (07:40)
Honestly, my my supervisor was really interested in it. So all like he had been working in what can the role of a computer in design be since his PhD, which was like finished in nineteen sixty eight or something like that.
Linda McIver (07:55)
Wow.
Kaz Grace (07:55)
And so he was programming mainframes and and doing stuff on punch cards, and getting involved in CAD before CAD was CAD, and and was always really interested in AI as kind of this other frontier for what the role of a computer could be in the design process. And so he inspired me to take the things that I was interested in, which is like, how do people make associations? When you interpret a problem, what are you actually doing? Why can people with similar backgrounds interpret problems in completely different ways? how, you know, all those sorts of kind of problem-centric reasoning issues. And he encouraged me to consider computational ways of exploring those topics. And then I got hooked.
Linda McIver (08:47)
That’s awesome. I I feel like the so when I did my PhD I used to tell people that I was studying the human side of computing and people were like, computing has a human side. Particularly computer science-
Kaz Grace (09:00)
Yeah.
Linda McIver (09:00)
-was not known for its humane and and you know, human centered perspectives at the time. not sure that it is now either.
But but the idea that computers have a creative side is something of course that we’re seeing pushed very hard by organizations that are absolutely not making computers programs that are able to be creative. So what’s what’s your perspective on on the the creative side of computing?
Kaz Grace (09:31)
Well, to to momentarily throw a little bit of shade at at Silicon Valley, I think the tech industry is very good at human focused things. I mean, in the same way that a chicken farmer understands chickens.
I think that there is actually a lot of great stuff that software can do to help people be creative. let’s say software for a second before we talk about AI.
I I think part of it is about agency, as in like I’m a relatively poor drawer, like I’m bad at sketching. I did a design degree, so I should be at least like workable at sketching, but like just between you and me, I didn’t do so great in those units. Once we got our keyboard in front of us, my my my performance improved. I’m I’m just not a I’m just not a good at sketching. I just can’t I I just can’t think about it. The visual thinking is fine. Like I’ll I’ll draw boxes and arrows and things like that, but I just can’t I can’t draw a leaf for the life of me. So having the ability to work with a drawing program, do 3D modeling or or even just make sketches on a on a digital canvas provides a bunch of agency that I don’t have.
And so for a lot of people, software programs, even stuff that we don’t think of as like professional software, like let’s take Canva. That’s a great example. I know a bunch of people at Canva, they’re they’re very cool. and Canva was derided by a lot of people in the design industry for for a decade or so as you know, it’s like designed for people who aren’t designers. But another way to think about that is that it enables or gives access to a kind of artifact a a kind of thing that you would otherwise need ten thousand hours of practice to make.
I I can think of similar things in the software world like low code tools, right? Like you can put together something, you can string together an app or or whatever. Now are there limitations at using a tool like Canva? Are there potential downsides to using a low-code tool to put together a mobile app or whatever? Yeah, absolutely. Right? Not least of which is your app’s gonna show up 1.2 gigabytes because it’s got an entire like browser running inside it or whatever. But for people who genuinely couldn’t have done that without that then that’s cool, right? Like we should we should celebrate that as cool. And that even extends to some of the more derided parts of generative AI, which is like text-to-image systems. a very good friend of mine is a special ed teacher, and she works with people with profound physical and mental disabilities, and they love generative AI. These kids really, really love getting to make songs or getting to make a an image with their favorite thing on it, you know. Like this is actually awesome.
That said, creativity has like at least two definitions. I think it’s got like a whole bunch, actually. It’s one of the more overloaded words. But when we can say a person is being creative, we’re saying they’re they’re producing something that is original and valuable and interesting to them. Right? So when my seven-year-old draws, no one would say she isn’t being creative.
But would an art critic look at the productions of a kid and say, yes, this really does need to be picked up by a gallery. Well have our you know, I I don’t want to throw any shade at at art critics, right? But there’s there’s a point where something can be societally creative, you know, so the society recognizes it as being interesting, novel, useful, whatever the the right labels are in that domain.
And something can be individually creative. This is Maggie Boden actually. it’s always cool to call out female computer scientists who are foundational in their field. Margaret Boden, who unfortunately passed away within the last year or so, was an absolutely foundational person in the field of AI and creativity. And she was the first one that I know who articulated this difference between individual and societal. She called it historical.
But then you get into arguments about which society invented something first. Like was inventing gunpowder in Europe less of a creative invention because China had had it for five hundred years already or something like that. So let’s not worry about historical. I prefer societal in the in the context of a of a society, right? So when we talk about creativity, I think AI can be really useful for individual creativity, in that it can help someone do something that they wouldn’t necessarily have the skills to do, but that if you increase your skill level, then you’re gonna run up against the problem that the large scale generative AI systems are just going to be a regression to the mean, right? They they produce mediocrity on tap. They are infinite extruders of mid. And so once you reach that point, there is somewhat of a a limit on what you can make with them. Really frontier LMs are a little bit different in some cases, just because I think I think that talking to a twisted mirror of yourself can sometimes move your own thinking, even if the mirror doesn’t say much useful. like I use LLMs like that. I I I find them useful in an inspirational or brainstorming context, and it’s probably not what it’s saying that’s helping me think. It’s like more like what I’m saying that’s helping me, but I I still consider the tool pretty useful. But for like a text to-
Linda McIver (14:48)
Yep. Like
Kaz Grace (14:50)
-image thing, it becomes it becomes challenging for it to be useful the more skilled you are.
Linda McIver (14:58)
It’s it’s like rubber duck programming, you know. I I one of my year elevens, they that year they they gave me I used to do a virtual lab coat, which was like a you know, because year twelves sign the lab coat kind of thing. I used to get them to-
Kaz Grace (15:15)
Nice.
Linda McIver (15:15)
-to write a comment on my website and and this time they made me an actual lab coat and he wrote something on it like, thank you for standing over my shoulder watching while I bug fixed all my dodgy programs and he would almost invariably find that I would come and I’d be like, So what’s the problem? And he would start explaining the problem to me and as he was explaining the problem to me, he like, “ohhh” I very rarely had to say anything at all, but I had to be there.
Kaz Grace (15:43)
You really do. And and I I gotta say, like classrooms are changing. like people don’t talk as much. maybe maybe they don’t want to be disruptive, but I’ve I’ve really noticed it that there’s a room full of twenty students and they’re all sort of looking at their laptops and there might be a couple of people chatting or something or but less than there used to be. And so you come over and you say, Hey, what’s going on? and that might be the first time someone has verbalized the issue that they’re having and then they’ll immediately rubber duck themselves and you as the tutor could go, I’m happy to help. Good job! Yay. Teaching.
Linda McIver (16:14)
Yay me!
Learning what when not to intervene and when not to say something is the I think one of the hardest parts of teaching. Certainly for me, who, you know, I’d rather chat the whole time than learning to shut up oof!
Kaz Grace (16:30)
Me too, dude. I talk too much. Yep.
Linda McIver (16:38)
Hmm. So my next formal question is what do you wish everyone knew about data? But I’m going to for you add in and AI.
Kaz Grace (16:52)
I would really love if everybody understood that even the really fancy frontier LLMs, experts tend to think they’re good at everything except what they are expert in. And so someone goes, you know what, this is really helpful at recipes for helping me cook, or this thing’s really helpful at talking about this, that, and the other. But it probably doesn’t know that much about my thing, but it knows about everybody else’s thing.
And I just think if everybody stopped and thought about the systemic implications of that-
Linda McIver (17:26)
Yeah.
Kaz Grace (17:26)
-it’s not expert at very much. now, we have to put that in the context of particularly things we’ve seen in the news in the last couple of weeks, like a solution to the Millennium Challenge around the Navier-Stokes fluid dynamics problems, which I have to admit I don’t have enough math knowledge to fully understand. So the the the AI can know more about me there. But looking at these these various different things, a lot of these AI systems are attacking classes of problems successfully where two things are true. It’s the classic computer science thing, the definition of a problem that computers are good at is one where there is a low cost to trying again and where it is easy to check if your solution is good.
If you think about it, this this characterises a lot of classes of problems in theoretical computer science. If you can try again over and over and over and over and over again with a low expenditure of resources, and if you can easily check if the solution is accurate, then you you have something that a computer is very good at. And that’s held true for traditional optimisation, it’s held true for just sort of like generating test approaches, and now it’s holding true for Machine Learning. A lot of these sort of the style of proof, and again, I’m not a mathematician. I can’t actually verify the proofs. I have taken some graduate math classes and folks are well beyond me in that regard.
Linda McIver (18:48)
Yeah.
Kaz Grace (18:51)
But the style of progress that we’re seeing, is a lot of this these sorts of things where you can just sort of try out a bunch of different initial conditions, get some feedback on how you’re on how you’re going, and then try a lot more things. And we’re seeing these very incremental progresses. I’m actually less interested in whether AI can do stuff by itself.
Like it’s a really good tech demo for the AI companies, but the real value will come in whether AI plus one or more humans can do something that just the humans couldn’t have done by themselves. Right? So the the the unit of progress will be AI plus human rather than AI alone.
That’s just harder to verify and write blog posts about ’cause if you get some very high end fancy mathematician and they discover a thing, well, that’s not gonna make news. That’s water is wet, right? Mathematician does maths.
Linda McIver (19:47)
Yeah.
I like that. That’s really cool. so bearing in mind that for all of these questions we can add AI where we have data. the next one of my questions is what are the worst data mistakes that you’ve seen? you can talk about data or you can talk about AI. Or both.
Kaz Grace (20:13)
So I’m actually going to talk about data to start with. So I I have the the fun privilege of teaching information visualisation to final year design students each each year. I’ve done that for about five years now. it’s a really fun course to teach because you have people who’ve spent three years getting good at understanding people. and then we have to say, All right, everybody, completely throw out everything that you’ve learned just for a second. We’re gonna learn about data, we’re gonna learn about systems, we’re gonna b learn about, you know, cybernetics, we’re gonna learn about data structures, we’re gonna learn about, you know, the difference between correlation and causation, which doesn’t come up in a design degree, right? Like why would it need to?
Linda McIver (20:55)
Yeah. Yeah.
Kaz Grace (20:56)
And it’s a lot of fun because well, I I I enjoy teaching things. but it’s a lot of fun because it is a very different way of thinking and I encourage students to sort of reflect metacognitively on how the different approaches are different while while teaching this. But that means that I have seen a lot of mistakes about data from people who are very knowledgeable and very skilled and very creative, just not at this.
Linda McIver (21:24)
Mm-hmm.
Kaz Grace (21:24)
Ao I I I think the biggest mistake is not considering the data source. It’s so easy these days to go out and find relatively large scale data on whatever it is, the thing that you’re interested in. And I I set briefs like, make me a thousand word data journalism style article something on something related to inequality.
And I’ll have people go out and find really, really cool briefs like shade inequality in Western Sydney. So a bunch of less socioeconomically advantaged suburbs with significantly lower canopy cover from trees and what are the impacts on that and like that’s that’s a really cool series of of data points about canopy cover and climate and roof temperatures and et cetera, et cetera. and I’ll have people looking at like global fashion, like fast fashion, where do all of the tons of of clothes end up? And then I’ll have people who took an interesting topic which might be a good example from last semester was where you are born affecting where you’ll be able to get a job right so just looking at inequality of inequality of opportunity based on birth nationalism right like so so birth citizenship but then it’s a cool idea.
The website looked good, but the sources were you’re like combining things from different countries and you’re not necessarily thinking-
Linda McIver (23:00)
Mm-hmm.
Kaz Grace (23:00)
-about how to do the data interoperability and and it’s-
Linda McIver (23:03)
Hm.
Kaz Grace (23:03)
-really, really hard to take people who are not skeptical about data from the beginning.
I’m I’m gonna use that as my bridge to get to the AI topic because people who are not trained to think in that, how is this data lying to me? How might this data be wrong? What is the potential-
Linda McIver (23:25)
Yeah.
Kaz Grace (23:26)
-perspective of the person have made it? what are the pitfalls in putting together this kind of data? What kind of missing entries or interpolations or et cetera, et cetera, et cetera?
People who haven’t had the opportunity to do that critical digital literacy stuff. are the ones who tend to get themselves in trouble making data charts and the ones who tend to get themselves in trouble naively using AI, right? Because the AI says something. It’s not is the AI right or wrong about this, it’s where did it get the data from?
Linda McIver (23:56)
Yeah. Yeah.
Yeah. And that’s so that’s a the foundational question that I always start my data projects with, that I that I teach teachers and students to do is what’s wrong with the data? Because if it’s real data, there’s no such thing as perfect data. So where are the issues? And that skepticism doesn’t come naturally. We’re weirdly trusting species, but it can be trained. And the more you think about it, the more likely you are to do it next time. You know, every time you you ask that question, it sort of reinforces the neural pathways and you get better at thinking, well, hang on, what was the sample size? And where did they collect it? And you know, if it’s if it’s survey data, you’re gonna get a different result outside of my church on Sunday compared to outside of my supermarket.
In the middle of the day versus in the evening or outside of a school or outside you know, outside of a meeting of Skeptics Australia or you know in the middle of the CBD-
Kaz Grace (24:59)
Right, yeah. Yeah.
Linda McIver (25:02)
-like what what who is missing from your from your data?
Kaz Grace (25:05)
Yeah.
Linda McIver (25:07)
Who is overrepresented? All those kinds of questions that we don’t and and even what are the questions you’re asking. You know, I teach evaluation surveys and and people are always like, How great was my workshop? And I’m like, mm-
No, you have to start with how did you find the workshop, you know, and then a scale from terrible to great, you know, like you have to you have to you have to frame it neutrally otherwise you are absolutely, you know and very obviously seeking a result which you’re more likely to get and
Kaz Grace (25:38)
Which they’re absolutely gonna give you because humans are not only trusting, we are naturally fairly polite to strangers, right? We’re actually, once-
Linda McIver (25:45)
Yeah.
Kaz Grace (25:45)
-you know someone, you might be more willing to to be honest with them. But you’re you’re likely going to say the thing that a stranger wants to hear unless you have some prejudice, honestly. yeah,
Linda McIver (25:54)
Yeah, exactly.
Kaz Grace (25:56)
I I do a lot of work in human-computer interaction, and HCI research has a real problem with this, which is that I build a thing, and so we do a little controlled experiment where people it’s usually within subjects because recruitment is hard. People use prototype without the thing and a prototype with the thing. And don’t worry, we we r randomize the order so half of the people use this one first and half the people use that one first, and then we had them to do we had them do a survey after each one and look, my thing makes them better at whatever scale we care about.
Linda McIver (26:30)
Mm-hmm.
Kaz Grace (26:31)
Except, you were the same person administering the survey as doing the tests, it’s really obvious which of the prototypes has the special sauce in it, right? It’s-
Linda McIver (26:44)
Yep
Kaz Grace (26:44)
-the equivalent of feeding someone a plain burger and then a burger with literally special sauce in it and then looking at them with stars in your eyes saying, How was it? How was it? How was it?
Linda McIver (26:53)
Yeah, yeah.
Kaz Grace (26:54)
It’s not like a carefully blind randomized controlled trial of a new medicine where you’re comparing a placebo to a you have to be really skeptical about where that kind of data comes from. No one’s gonna sit there, eat both your burgers, use both your mobile phone apps or whatever, and then say-
Linda McIver (27:10)
No.
Kaz Grace (27:14)
-sauce didn’t help at all actually.
Linda McIver (27:16)
Yeah. Yeah, I did a I did a a a psychology experiment one time where they asked me to watch a video and like I don’t know, rate my attitudes to something or stress levels. I can’t remember what they were looking at, but one video was, you know, a well being a set of well being exercises and the other video was a video about waste recovery.
Like I feel like you might have signalled the
Kaz Grace (27:50)
Yeah, I wonder which one of these is the control.
Linda McIver (27:53)
I just I don’t think anyone is seriously looking at the impact of a video on waste recovery on well being. Like I mm
Kaz Grace (27:57)
Yeah. Mm. Yeah. But this just shows how well developed I mean, okay, mainstream medical science does, does cop some deserved flack for a number of the different ways that things get advanced and things get worked with. There’s a number of perverse incentives when it comes to innovation and large corporations and et cetera, et cetera, et cetera. Challenges in funding and sustaining public health campaigns. I’m not going to pretend that it’s perfect. However, you really have to compare how well developed and rigorous a field like medicine is compared to something like human-computer interaction or psychology, because they have pre-registration of the kinds of data that you’re going to do and the kinds of methods and the kinds of and conditions and a not just a an culture but in some cases a regulatory requirement that your control be the existing best available therapy.
Whereas if you compare that to human-computer interaction, where let’s say we’re studying something to do with augmented reality glasses or something rather than say, hey, does this kind of thing work better than that kind of thing? The scientists will probably just go with whatever the PhD student cooks up as the control, right? Like,
Linda McIver (29:06)
Yeah.
Kaz Grace (29:09)
I’m not suggesting it’s always bad. There’s tons of really cool research in these fields, but you just have to sort of on that skeptical lens, you have to step back and say, what is the culture of this field when it comes to constructing experiments and how much effort is put into making sure that the control is genuinely the best existing thing that’s out there. So that when you’re comparing your hot new thing, that you’re are you comparing it to something a PhD student vibe coded last weekend or are you comparing it to the best extant therapy? and there’s a lot of there’s-
Linda McIver (29:35)
Yeah.
Kaz Grace (29:39)
-a lot of little things like that. Now, I want to make it really clear that that doesn’t mean that I think we should throw the baby out with the bathwater, right? Like all of these scientific fields, all these particularly innovation-centric fields that make new things have have tons of value. But you you’ve really just got to be healthily cautious, right? You can’t over-interpret single results and say, well, these seventeen participants found this thing. Now let’s r redesign everything around around that. which is the kind of thing that someone who used a chatbot, which then pulled a paper from a respectable conference last year, that did a study like that. and then the AI goes, based on the best available research, you should do things this way. And now it’s been filtered through a filter through a filter, and someone goes off and uncritically thinks that that’s the best way to approach your problem.
They don’t stop and consider alternatives. They don’t think critically about where that data came from or where that result came from. They don’t think about where the AI got that information from. ‘Cause of course if OpenAI or Claude or whatever says this is the best way of doing it, and they’re the best AI, well that must be the best answer.
Linda McIver (30:51)
Yeah. Yeah.
Kaz Grace (30:52)
That’s logic.
Linda McIver (30:57)
excuse me while I head the desk.
Kaz Grace (30:58)
Yeah.
I guess I don’t know, honestly, I’m gonna flip the script here, ’cause you’ve got more experience in this space than me. Linda, how can we fix this in high school? How can we make this sort of critical thinking, critical digital literacy around not just data but AI? How can we make this something that people come out of school already thinking about? I don’t know.
Linda McIver (31:23)
I have I have big feelings on this. and and you know, this episode is not about me, but I will-
Kaz Grace (31:30)
Mm.
Linda McIver (31:30)
-say that I do think we have the answer to that. And the answer to that is more getting kids to solve real problems and critically evaluate their own work as well as the work of others, and less getting kids to produce the right answers. And-
Kaz Grace (31:48)
Hell yeah.
Linda McIver (31:49)
-we we still do too much.
Give me the answer, look it up in the back of the textbook, mark it as right or wrong, and move on. And we don’t do enough of that. You know, you were talking about how AI is really good at solving problems where you can really easily check the right or wrong. We need to be getting kids to solve problems where you can’t easily check the right or wrong, where there isn’t actually a right or wrong, and you have to go, who does it help? Who does it harm? Where does it work? Where does it need improvement? And then kids come out knowing that they don’t have a right answer. They don’t have a perfect solution, but their own work has flaws always because real meaningful work has flaws. And so the the the bit that you assess, the bit that you prioritize is can you tell me how your how your answer is flawed and how you might improve it? Like that that becomes the focus. That becomes something that we build into the the fundamentals of every level of education. You can start doing this in kindergarten and then you you take it from there. You so that you’re you’re spending the whole time not telling kids this is this is the right answer, you’re telling kids critically evaluate your work and and it it just it changes everything.
Kaz Grace (33:10)
What I love about that, that’s that’s so cool. What I love about that is moving from teach kids to get the right answer to teaching kids to defend their answer when there is no right answer. What I love about that, and and also, you know, not just defend as in blindly, but also critique.
Linda McIver (33:25)
Yes.
Kaz Grace (33:25)
What I love about that is that that is hard for exactly the same reason as why AI is bad at it.
Because solving these problems doesn’t scale easily. And it doesn’t scale easily in training data, but it also doesn’t scale easily in a public education system. Right?
Linda McIver (33:41)
Mm-hmm.
Kaz Grace (33:41)
Like it is very hard to spend enough time with each individual student to explore and further those kind of skills. Now I’m I’m gonna be a little biased and I’m gonna say that that’s kind those are design skills. What you’re talking about is the kind of thing that we that we teach in in design studios. studio based education, which is the way that design degrees have worked for a couple hundred years, grew out of the kind of atelier model of like you’d have a fancy architect and then the fancy architect would take on a number of apprentices who would be asked to do mostly menial jobs here and there, finishing things off, touching things up, getting stuff out the door, handling clients, you know, the drudgery.
Linda McIver (34:24)
You’re disturbingly good at that voice.
Kaz Grace (34:30)
I chose that one because even though Atelier would have been a French practice that emerged largely in Paris, you really don’t want to hear me attempting to do fancy French. I’m gonna stick with I’m I’m-
Linda McIver (34:37)
Ha ha
Kaz Grace (34:38)
-going to stick with British.
Linda McIver (34:40)
Good.
Kaz Grace (34:41)
But the idea is that by being part of those different parts of the design process, the apprentices would learn through osmosis, and then there would be this critical practice, literally called critiques, where the apprentices would have some time set aside to come to the master and say, you know, I I’ve been kind of working on this. And this would often be conducted in groups and there would be discussions and critiques on each other’s work. And so in return for doing honestly unpaid drudge work for some fancy architect, you would get to be part of critical creative community that is discussing and sharing and moving forwards. And you have like a little nudge from the master or a little tap on the shoulder saying, you know, that one’s actually not that bad every now and then.
Now, that didn’t have to scale because the master and the apprentice were all upper class. Right? Like let’s be honest, you weren’t getting into one of those things if you’re a chimney sweep. You already had to have a family supporting you and you probably had to have that family pay a generous donation to the creative works of such a really really prestigious artist in our city in order to to buy yourself spot.
We still face the same challenges in design degrees a couple of hundred years later trying to scale up studio practice. And so Australia’s education system, for better or worse, and we could definitely talk about that, but it’s maybe another episode.
Linda McIver (36:03)
Mm.
Kaz Grace (36:04)
We don’t support our tertiary education, our universities sufficiently from the public purse. and as a result, in order for us to have enough spare resources to conduct research, which is a very costly endeavor for the first little while. and in order for us to to deliver on our sort of taxpayer mandates, we enroll a lot of international students who can be called upon to pay higher tuition. And opinions are split on whether those international students are being provided an English language education in a respectable institution, which they can then take back to their home countries or anywhere else in the world and use as sort of a currency for employment, or whether we’re sort of a pipeline for getting permanent residency with a couple of extra steps and a piece of paper along the way. That’s sort of the positive or the negative spin on that one. But as a result of that, my design classroom has a couple hundred students in it and would not be recognizable as a studio to those fancy French architects from 1850, right? and so scaling up that kind of critical and reflective thinking, what role can technology play in that?
How do we make sure that we can even recruit so if I need to recruit like a dozen tutors who will then each run a classroom, maybe I’ll run one of them and I need eleven tutors or something along those lines. Class size is roughly about twenty. And I need to recruit people who don’t mind working for sixty bucks an hour on a very on a casual job. 60 bucks an hour sounds pretty good if you’re doing 40 hours a week, but if you’re doing 10 hours a week, it can be really disruptive. How to structure that at that scale is already incredibly challenging. and I would really, really love someone to put in the hours to figure out how we could do a similar thing with all that kind of creative critical, reflective, metacognitive stuff in even, you know, stage five and six of high school would be a challenge.
Linda McIver (38:14)
Bear in mind those classes are already about that size, right? They’re they’re twenty to twenty five sometimes. I don’t I think in Victoria the the hard limit on class sizes is twenty five at at school. so like you you don’t have this quite the same scaling issues you do at universities, where often the the the default template is one lecturer and four hundred students in the room and you know, some some others online and it d you don’t have to scale, you know, because you don’t have to split them into groups of twenty. Schools are already in groups of twenty. So it’s a shift of teaching style rather than a shift of resourcing. there that you do need more resources, but not at the scale that you need more resources to do that at university. So it’s a it’s a different question. But also, you know, you said that universities are woefully underfunded, see also public schools.
Kaz Grace (39:06)
Yeah. I think the the scale issue is how do you how do you I’m just gonna pick a number, I think it’s probably in the right order of magnitude, how do you teach twenty thousand public school teachers who need a week off and a raise already? How do we give them two weeks of training and even then two weeks-
Linda McIver (39:20)
Mm, mm, mm.
Kaz Grace (39:23)
-of two weeks of training is probably not enough, right? that’s
Linda McIver (39:26)
Mm. Well, you know, it it’s it’s actually there are a ton of really skilled teachers out there who are doing this kind of work already. and to some extent we need to give them more time and that’s where the resourcing comes in. You know, I did a when I was teaching in secondary school, I did a little calculation on how much time I had to spend per student and how much time I had to spend on preparation and it was horrific, if you had me working my scheduled hours, which of course nobody does, you know,
Kaz Grace (39:59)
Yeah, yeah, that sounds about right.
Linda McIver (40:02)
Which is problematic in its you know like this there’s so many issues there. but also you you need to give teachers more trust and less constraints. so take out a ton of the admin would be a really nice start. And then you also remove all not all the facts, but you remove a lot of the fact based curriculum. You you remove the the requirement that we’re going to the end of the year and we have to regurgitate this set of facts, otherwise we haven’t succeeded in whatever way, you know, like that. Yeah. Anyway, this is Yeah. Yeah.
Kaz Grace (40:37)
Yeah. It’s a classic yeah. This this is this is a separate tangent that we’re getting at. But it’s it’s a classic scale problem, right? The fact based stuff is is in the curriculum because it was easy to scale. But we’re getting to-
Linda McIver (40:48)
Mm.
Kaz Grace (40:48)
-the point where it no longer serves the needs of society to assess no they’re not gonna pretend that teachers are just teaching facts, right? They’re as you say, they’re teachers doing sorts of amazing amazing different things in the way that they’re getting kids with experiential learning and problem based learning and and small group learning and all this kind of awesome stuff. But the assessment component comes down to regurgitating facts. And why would you learn to regurgitate basic facts in twenty twenty-six? It’s hard to get motivated to do that.
Linda McIver (41:14)
Yeah, it it absolutely is. And and it teaches kids the wrong thing. So my kids are both really good at maths, but they believe that they suck at maths because they weren’t great at the times table challenge in primary school. And so obviously they’re bad at maths. Now, the times table challenge is a memory exercise, not a maths exercise.
Kaz Grace (41:35)
Yeah.
Linda McIver (41:36)
And they couldn’t be bothered learning the like memorizing the things and why would you? But you know, you say it’s not not fit for purpose anymore. I would argue it never really was. I would argue if we had been teaching critical thinking from the get go, we would not have a society where we were still using fossil fuels. You know-
Kaz Grace (41:59)
That is such a good point. Yeah.
Linda McIver (42:01)
-like where anti vax was such a thing. Like we where-
Kaz Grace (42:04)
Right. Yeah.
Linda McIver (42:05)
-inequality was such you know, we we’d have solved all this shit if we’d actually been doing critical thinking properly.
Kaz Grace (42:09)
Where where we had these weird ideas that you know, what we put in the atmosphere doesn’t matter, but what is inside other people’s underwear does matter. It’s like hold up a second, right? Like which one of those sorry, sorry, how much do you interact with other people’s underwear and how much do you interact with the air that you breathe?
Linda McIver (42:27)
Yeah.
Kaz Grace (42:28)
Like let’s take a step back and think about that one.
Linda McIver (42:37)
yeah, that’s perfect.
Kaz Grace (42:38)
It you know, the w- I really love that perspective, right, where you said that it probably never did, which means that AI generative AI, if we’re to try and put a positive spin on this, it might be the Emperor’s New New Clothes story, right? Generative AI is the little kid who goes ‘The education system’s not wearing any pants’, right? Sorry, I’ve suddenly gone on a whole-
Linda McIver (42:58)
Mm. Yeah.
Kaz Grace (43:01)
-bunch of pants-related metaphors. I could I could pick a more salubrious metaphorical target. But the the point is that essays were never a particularly good way of evaluating a student’s understanding of a-
Linda McIver (43:12)
Yeah, yeah.
Kaz Grace (43:14)
-first-year psychology course, right?
Linda McIver (43:17)
Mm, mm-hmm.
Kaz Grace (43:18)
Writing really big long reports of everything that you tried while you were solving a project actually wasn’t like where you where you you get marks for how many things you tried, that’s that’s actually like a second order approximation of what it means to tackle a complex problem in design or software engineering or whatever, right? Where we say like it’s important that you try multiple ideas, so you need to document three of the ideas that you tried. That just means you come up with one good one and then you invent two shitty ones along the way, right? Like once you get once-
Linda McIver (43:43)
Yeah. It’s like
Kaz Grace (43:47)
-you get to a good one you just think backwards and you come up with two worse ones.
Linda McIver (43:50)
It’s like the lines of code metric. It’s like, no, that’s not.
Kaz Grace (43:52)
Right. Yes.
Yeah. These are these are you know, so then it is a metrics problem, right? So educational assessment has to metricize. Well, the way we’ve structured education suggests that assessment must metricize. That is assumed that is an assumption. And-
Linda McIver (44:04)
Yep. Yep.
Kaz Grace (44:06)
-so we find things that are metricizable.
And most metrics are garbage.
Linda McIver (44:09)
Mm-hmm. chapter chapter four of my book is measurable or meaningful, pick one. So yes, you’re you’re you’re singing-
Kaz Grace (44:15)
Hell yeah. Yeah. Spot on. Yeah.
Linda McIver (44:20)
-my song here. so to to come back from our glorious tangent for a moment. have you ever seen data deliberately misused and how can we spot things like that?
Kaz Grace (44:33)
I have a have a lecture where I talk about people who are mis- and dis-information with data. And a lot of the the the deliberate cases there, right, are are people who are ex like explicitly manipulating or misusing data. And usually this comes down to media and politics, right?
Linda McIver (44:56)
Yeah.
Kaz Grace (44:59)
There was a classic example from a couple of election cycles here ago where a tabloid here in New South Wales had produced a pie chart showing the percentage of the vote that was going to the Liberals versus the sorry, the coalition versus Labor, and the the segment of the pie chart that was labelled thirty-nine percent was taking up more than half of the chart. And it’s like this is this is not even complicated or sophisticated manipulation. This is just lying. This is just a this is just nonsense.
Linda McIver (45:30)
Mm, mm, mm.
Kaz Grace (45:36)
I have some great examples from Fox News in the US where they’ll be showing a time-
Linda McIver (45:39)
Boy.
Kaz Grace (45:39)
-series and the data points are each data point is labeled on the time series like four four point five, four point six, four point eight, four point nine, it’s unemployment data or something like that. and they don’t even line up with the axis.
And they’ll be highly like these are the Obama ones, these are bad. and they’re just they’re just deliberately moving the points around. Like someone in the art department who clearly wasn’t in the data science department, or or maybe someone in the data science department, just just deliberately twisted the graph. So I think that there are really sophisticated and messy ways of misrepresenting data. I’ll give you one example, let me finish my other point, but we’re actually living in a world where really unsophisticated ways of manipulating data and just fully lying with numbers, just saying things that aren’t true, is so-
Linda McIver (46:23)
Mm. Mm-hmm.
Kaz Grace (46:26)
-prevalent that we are almost living in a post-truth media landscape. post-truth not that not in the sense that there is no truth anymore, in the sense that truth is no longer an important metric. Truth is no longer an important part of what makes our media landscape what it is. And everybody needs to sort of recognize that. And that can be just horrifically unsophisticated like just lying on a pie chart and saying thirty-nine percent, which is the correct number, but then having that be, you know, more than half of the pie chart, right? That’s just not how pie charts work. But then I’ll give you a more sophisticated-
Linda McIver (46:53)
Mm-hmm.
Kaz Grace (46:56)
-example. Someone who works for the US Treasury Secretary gave some data this week about how many barrels of oil are getting through the Strait of Hormuz in the US Navy’s convoy operations to try and get tankers through Iranian controlled parts of the strait. not traditional convoys, they’re not actually sending naval-
Linda McIver (47:17)
Mm-hmm.
Kaz Grace (47:19)
-vessels through. I think that’d be too dangerous, but they’re using uncrewed vessels and they’re using helicopter support and things like that. Sh crews are going as ships are going through at night. They recently released something saying that they had gotten through, I don’t know, like something like eighteen, eighteen and a half million barrels of oil, which is close to the twenty million pre war number. And so like basically they were implying that they did eighteen million on Tuesday last week or whatever it was and that’s back to the pre-war number.
They don’t do the convoys every night. They did one really big convoy that night. If you look at the average over multiple days, which is of course what people in the maritime intelligence space do, the average over multiple days is still like six, seven, eight, nine, depending on the week, right? So forty percent of the original number. But by cherry picking, just that one day, where it was true apparently, I mean it’s very hard to verify these things, they coordinate the convoys between satellite overpasses so that commercial satellites don’t actually pick up the the
It’s a fascinating and complicated logistical platform. I’m not that interested in war, but I’m really interested in systems and logistics, and the two do occasionally run into each other.
Linda McIver (48:28)
Mm.
Kaz Grace (48:29)
So the the long and the short of it is they cherry picked a number from a day where they did get a lot of they get almost pre-war levels through, but they can’t do that every day, and they know they can’t-
Linda McIver (48:36)
Mm-hmm. Mm-hmm.
Kaz Grace (48:38)
-do that every day. So that is a really sophisticated way of lying with data by getting up on all the major news networks in the US saying, you know, we got 18 and a half million barrels through the other day. That’s basically the pre-war number, then you’ve got all the stuff going over-
Linda McIver (48:47)
Mm.
Kaz Grace (48:49)
-land, honestly it’s gone up, is is more than just lying. Like ’cause they did actually get eighteen and a half million barrels through or that appears to be valid.
Linda McIver (48:59)
Yeah. Yep.
Kaz Grace (49:02)
It’s very tricky.
Linda McIver (49:04)
Yeah, I used to do a critical thinking unit with first year computer scientists as part of their communication subject. And I used to we used to talk about statements that can be true while absolutely misleading and that’s you know that’s a a classic of the genre, right? It’s like
Kaz Grace (49:21)
Yes. Yep. Yeah.
I I introduced a similar thing in in one of my lectures and I the anchoring example is the playground passive voice. Okay, what happened here, children? Billy was hit. It’s a true statement.
Linda McIver (49:37)
Yeah.
Yep. Yeah. Yep. See also Billy walked into my fist. Yeah.
Kaz Grace (49:46)
Yes. We were playing and Billy fell over. Yeah. And and you can be very, very precise in such a way that doesn’t that doesn’t give away the fact that you’re implying the exact opposite of what you’re doing. One the really wonderful things about having kids aged seven and nine is that they are getting old enough to detect when I am doing that to them.
Where if if they say, Can I can I-
Linda McIver (50:14)
They’re on to you.
Kaz Grace (50:16)
-have dessert and I say you know what? We really should finish that ice cream… tomorrow
They they’re now they’re like that’s not a yes dad.
Linda McIver (50:34)
And I’ve met your two the you’re in big trouble. That’s so great.
Kaz Grace (50:38)
Yeah is so much trouble. Bit of a rear guard action for me already at ages seven and nine. Yeah.
Linda McIver (50:47)
What’s the first question you ask when you look at graphs in the media?
Kaz Grace (50:53)
After the initial comprehension stuff, where I’m sort of looking at and being like, what’s being shown here? What’s being compared here? What’s what are they trying to what do they tell me? Is it going up? Is it going down? Is it lit linear? Is it that sort of really early comprehension stuff? I usually try to figure out where the data is coming from. I will first go and look at sources. I was like, okay, so we’re seeing a comparison. It’s going up. Is this government data? Is this third-party data? When was this released? Like, let’s go, let’s look at provenance. and I might be looking at an article and skip around a little bit looking for the sources or looking for who they’re quoting or where where the numbers come from. and if it’s something super important, I might open up another browser tab and go and think, well, who are the institute for foundation models? That’s a very generic name in the AI space. That’s very interesting.
Linda McIver (51:36)
Yeah. Yeah.
Kaz Grace (51:39)
Who are who are these folks? and it’s not I want to be really clear.
Campism will be the death of all of us, right? Just being saying, they’re in my camp, they’re the good guys, they’re in the other camp, they’re the bad guys, right?
Linda McIver (51:54)
Mm.
Kaz Grace (51:54)
So you have to be with us or you’re against us. that really is going to be a challenge, particularly for progressive politics, although don’t necessarily need to go down that angle. That’s something that we really struggle with on the left.
Linda McIver (52:08)
Yeah.
Kaz Grace (52:11)
But when I say that I want to know the sources, I will not discard something because I don’t think the source is in my camp. I would really encourage people not to do that. On the other hand, it’s really valuable to know where the data was collected, why the data was collected, what was maybe funding the data being collected. It’s not the same as saying if this study about, let’s say, the health of red meat collected a single penny from The Dairy Farmers Association, then we-
Linda McIver (52:40)
Mm.
Kaz Grace (52:41)
-should consider it propaganda and throw it out and it’s useless. And it was like, no-
Linda McIver (52:43)
Mm, yeah.
Kaz Grace (52:44)
-it’s it’s part of the picture. Provenance is important, but it’s not the only thing that’s important. And you really need to be in between those two things. Ignoring provenance is dangerous. Valuing provenance over everything else and discarding anyone who isn’t in your camp, equally dangerous, if not more.
Linda McIver (53:02)
So we’re back to there’s no right or wrong answer and you have to really critically evaluate it.
Kaz Grace (53:05)
There’s, it’s almost like everything’s complicated, Linda.
Linda McIver (53:12)
Nuance. I hate nuance.
Kaz Grace (53:14)
And you know what the real problem with nuance is? It doesn’t do numbers on social media.
Linda McIver (53:21)
Mm. Yep, I am aware.
Kaz Grace (53:23)
Yeah, I mean that’s that’s ultimately the problem, right? Like nuance is it doesn’t it’s not satisfying.
That’s not true.
Linda McIver (53:33)
Mm.
Kaz Grace (53:33)
Once you really kinda get to the that moment of that dopamine hit when you fully understand something and you recognize what’s going on, you understand a model or you understand a problem or you understand where something’s coming from or you’ve put all the pieces together, that is satisfying. What it isn’t is-
Linda McIver (53:44)
Mm.
Kaz Grace (53:46)
-rapidly accessibly satisfying in the way that watching someone fall over or an attractive member of whatever gender you’re attracted to is satisfying, right? Or even seeing-
Linda McIver (53:48)
Yep. Yeah. All right, cat video.
Kaz Grace (53:55)
-someone yeah, exactly. Even seeing someone that you don’t particularly like be owned, is-
Linda McIver (54:02)
Mm.
Kaz Grace (54:03)
-is more immediately satisfying. and that’s that’s that’s cool. Those like that sort of content is fun. the problem is when it overwhelms everything else. I’m not saying we need to ban YouTube or or whatever-
Linda McIver (54:13)
Yeah.
Kaz Grace (54:16)
-but it’s it’s like it’s like a chocolate bar. You probably-
Linda McIver (54:21)
Yeah.
Kaz Grace (54:22)
-shouldn’t only eat chocolate bars.
Linda McIver (54:23)
Yeah. Yeah. Have a carrot now and then.
Kaz Grace (54:27)
Now and then.
Linda McIver (54:30)
Is there a path from here to AI that actually does support creativity rather than trying to almost eradicate it? I Cory Doctorow likes to use the the saying, if you wanted to get there I wouldn’t start from here. But given that we are starting from here, is do do you see like is there a way out of this mess that we’ve we’ve been kind of thrown into?
Kaz Grace (54:59)
I hope so. I have to hope so because I have to keep writing research proposals trying to bring about that future and it’s very hard to throw yourself into writing a large scale research proposal when you don’t believe in what you’re doing. I’ve unfortunately had to do that a couple of times in the past. It’s not it doesn’t work so good.
Linda McIver (55:15)
No.
Kaz Grace (55:16)
I genuinely believe that the only thing we’re doing wrong is conflating creativity and productivity.
And whoever came up with the phrase creative productivity is not it. but AI companies have a strong incentive to produce products that can hit the largest possible addressable market. Right? So they want to make things for as many people as possible. That means making stuff that can answer the question, I have ricotta and sun-dried tomatoes, give me a pasta recipe.
That is a creative task, right? Like if you’re trying to come up with a recipe that fits some constraints, that’s a that’s creativity. It’s like what would be called small c or little c creativity in the in the literature, right? It’s like the it’s an everyday task requiring some open-ended thinking. It has a bunch of constraints and complexity and no right answer. it’s it’s design but with a little d.
And yeah, AI’s probably pretty good at that. Because it doesn’t really matter if you’re not that wrong or if there was some obvious thing that you didn’t think about or whatever. I think we need a I think that we would need a fundamentally different kind of system to help people be actively creative. Something that was much less about giving you the answers and much more about asking you the right questions. something that knew-
Linda McIver (56:35)
Hmm. I like that.
Kaz Grace (56:37)
-that the role of an AI wasn’t to automate, that was actually probably to slow you down. Help you to think more. Right? I think we need creative anti productivity. Yeah.
Linda McIver (56:49)
Heresy. Heresy. How could you?
Kaz Grace (56:54)
But when it comes to experts, when it comes to people who are knowledgeable about a particular domain that they’re working in, whether you’re a software engineer, whether you’re a fashion designer, whether you are a chef, a high school teacher, someone who anyone who is an expert in an open ended, challenging domain, one which has trade offs rather than just right answers. No, it doesn’t have optimums, it has it has challenges.
Would be called a wicked problem right now, a small a a small wicked problem, to use Rittel’s terminology. and I don’t think that a modern LLM is really good for that. It largely gives you the obvious answer. It gives you then r t regurgitates whatever additional stuff you put into the context.
And so what we would need instead is something that pushes your thinking along. Helps like organize and note-take, maybe, like because there are some components of brainstorming and being creative that can be automated, right? Like you have to write all the things down. It’d be nice to have a tool that writes all the things down. maybe even organize them a little bit or suggest potential organizations even. But we would have to give up the notion that that tool, firstly, is for everybody.
Because I think you would probably need to have a certain level of skill in the domain that you’re operating in before a tool which was posed in the form of questions would be valuable. You’d need a completely different thing to teach.
I you probably don’t also want a mainstream LLM that’ll just solve things for you, but there’d be like a a third thing necessary in an educational context. But if someone was already kind of at least skilled in a domain and they were tackling an open-ended problem, they would and should be able to use a system that is more about asking questions, that is more about helping them explore, pushing for different kind of framings, keeping them in the problem space reasoning. So you think about make the right thing versus make the thing right.
Right, so there’s problem-space reasoning about am I solving the right problem? And then solution-space reasoning about am I doing it correctly? Am I coming up with best possible solution given that framing? And I think what AI does is take your prompt, your problem space definition, and immediately jump to solution-space reasoning, come up with the most likely thing that it can think of in its training data, and then then kind of hold you there as much as possible. I want something that is about clawing humans back from their innate desire to solutioneer and keeping them in the problem space more. As to how that’s possible, what I would advocate for, on the kind of research that we do in my lab, DWAIL the Designing with AI Lab at at the University of Sydney’s Design School. We have almost no web presence I don’t know why I’m I’m name-dropping my stuff.
Linda McIver (59:28)
Ha ha ha
Kaz Grace (59:31)
Is is about trying to build on open source LLMs and come up with new ways of fine-tuning, of encapsulating in harnesses, of tool use to help bring that about. a really amazing student of mine, Jess, we just submitted a paper that is about comparing two mind mapping systems that are otherwise identical, except one of them only adds little notes to the mind map that are question notes, problem notes, asking you things. And the other one is a more traditional AI that as you’re exploring a problem adds potential solutions. And we found that people were much more likely to reconsider their assumptions and that they had a greater sense of ownership over the result when the AI was asking questions.
What I wish we’d-
Linda McIver (1:00:26)
That’s beautiful.
Kaz Grace (1:00:26)
-done and we hadn’t included in that study, it’s coming in the next one, is include a measure of cognitive load, right? Because I think, just and certainly from the interviews, qualitative interviews as well, certainly from the interviews, it was harder, it was slower. They felt like the questions were actually holding them back and it was appreciated. It was appreciated in a 60-minute study where they sat down, they only had to solve one problem and there was nothing really on the line because they’re just participating in a user study. What I want to explore is how do these things work if you use them a couple hours a day. Is it a system that’s holding you back in questions all the time too much? What are the kinds of balances between problem and solution space reasoning? And is it possible to detect? Is it possible to give the human control over when an AI should think about problem or think about solution? What are the good ways to use an LLM for problem space reasoning? None of this stuff has been properly explored because all of the trillions of dollars in AI are going to addressing the largest possible market, which is just making solution systems. I think that if-
Linda McIver (1:01:23)
Yeah.
Kaz Grace (1:01:23)there is a way out, and I agree with Corey Doctorow, if we were trying to get there, I wouldn’t start here. If there is a way out, it’s open source models that are not controlled by large corporations with an incentive to target the the largest possible addressable market, where we can use these things, not because they’ll have all the answers, but because they might sort of like hold that fun house mirror up to our own ideas and let us see them differently.
Linda McIver (1:01:48)
Yep. Instant profit is never gonna get you there. Like the i when that’s the motivation, it’s it’s always gonna warp things in that way.
Kaz Grace (1:01:58)
I still feel like these things could have economic value, right?
Linda McIver (1:02:01)
Mm.
Kaz Grace (1:02:01)
Like a two-hour brainstorming session by an expert can go in a lot of really interesting places, right? Can take an take an idea really interesting places. And if there were a way to incorporate an AI in that that might even help you do 10% more or 20% more during one of those open-ended creative systems, and sure, if you if you want to, you could then take the result of that and go use a different AI to vibe code up a prototype or something like that. I’m not saying that that’s always bad.
But th I think if experts were ten or twenty percent more capable in little creative sessions where they’re exploring different things, that that could have real serious economic value. Probably not take over the entire stock market value though.
Linda McIver (1:02:42)
Yeah. Yeah. Mm don’t get me on that. We’ll be here for the rest of the day. we’re up to my favorite question, the last question. What excites you about data?
Kaz Grace (1:02:55)
I I mean the thing that excites me about AI is the thing that I just said, right? So we can copy paste that answer into here.
Linda McIver (1:03:00)
Yeah.
Kaz Grace (1:03:01)
The the thing that really excites me about data is that we live in a world where even though the incentives are to be less critical and to think less about data and to just do what the algorithm says, whether it’s on social media or using an AI or whether it’s inside a large scale education system where the algorithm being the the the policy environment, the bureaucratic environment, is just saying that you need to learn you know, to do these twenty quadratic equations.
Linda McIver (1:03:29)
Even
Kaz Grace (1:03:29)
Even though we are in a space that is very, very centralized and controlled and the median level of agency is going down, we do actually have a lot of transparency and access to things. Right? My kids at age nine can, with a bit of adult guidance, jump on the internet and learn real facts about anything, and then mess around with a web tool that allows them to actually explore what does the asteroid belt look like, or to to think about the difference between species of birds, like the accessibility and transparency that is possible for the kids in this generation is is light years ahead of even what was possible for us like thirty years ago, thirty, forty years ago. It was and and that was sort of like at the dawn of the digital revolution. and then things of well the digital revolution happened and then maybe there was a counter revolutionary vanguard that took us in a slightly different direction than we were initially intending. And then again we’re back to the other topic. But
Linda McIver (1:04:32)
but
Kaz Grace (1:04:36)
The the possibilities are actually there. But just because we live in a world with algorithmically mediated pressure, away from thinking about things, you the the the what you can do and what you can play with and what you can put together and I’d sort of include generative AI in that.
Like I have a whole bunch of ideas that I don’t have time to program, and like just v vibe coding up some prototype, which I know is badly written and I haven’t really thought about its design structure very much. and just playing around with it is actually really valuable and interesting. I I do look forward to that. I do r really think that it’s a cool world that we live in. I just hope that experiencing that transparent access to stuff, that cool world, doesn’t become a luxury good.
Linda McIver (1:05:28)
Yeah. Yeah. I like that. That’s a beautiful place to end. Thank you so much. It’s been a great conversation. I’m gonna be thinking about your chicken farmer analogy all week. And boy. It’s been wonderful. Thank you so much. It’s been great.
Kaz Grace (1:05:46)
Thank you so much, Linda. It was absolutely fun to chat with you this morning. I’m so glad that you invited me on at a dinner party. You should keep doing that. These these are really interesting, interesting chats. Yeah.
Linda McIver (1:05:55)
I do. I’m unstoppable.
Kaz Grace (1:06:00)
Cheers.
Outro (1:06:03)
Thanks for listening to Make Me Data Literate. You can find more episodes at ADSEI.org/podcast and you can support our work at givenow.com.au/ADSEI. Have a great day.
