Dr. Rachel Horst: AI Anti-Slop AI Writing Machine | Vancouver AI May 2026

0:00Incoming transmission from the future. Vancouver AI community meetup. >> Up first we have Dr. Rachel Horst who I met um in a different lifetime when I was a rock and roll photographer and she was a rock star. And um your label called me to come make some film photos with you up on the on the Sunshine Coast and uh and um I thought from the very first moment that I met you that you were so cool.

0:28Like uh living on the coast, living on the reservation, not a lot of white people choose to live on the reservation. Um teaching music, your family. I've I've long looked up to you. And so when you came back into our world a couple years ago with your PhD um and uh interest in the creative AI, I was all ears. And so I'm very very happy to invite Dr. Rachel Horst here tonight.

0:53Um she's going to tell you more about her work, but she entered a contest, she built an agentic anti-slop creative magnum opus creation tool of some sort. Dr. Rachel Horst. >> Here. >> [applause] >> Thank you so much, Chris. Thank you. All right. Yes, so I am Rachel Horst. Um I am a lecturer at UBC in the Master of Educational Technology program. Um and my research is I I do a lot of different kinds of things, but I'm I I'm predominantly an arts-based researcher. Um and I'm really interested in how digital technologies and AI specifically are are changing our literacy practices, our meaning-making practices, our um our learning practices. So um I haven't uh memorized my whole talk for you all, so

1:45I'm going to be um I have I I was practicing at the beach today, but I didn't practice with a mic in my hand. So um it's going to be interesting, but we'll we'll see how it goes. Okay. Um yeah, so I'm going to the talk today is called building an anti-slot machine.

2:02Um so I'm going to share this research I did with you that it all started a couple months ago. We can go to the next slide. Um with this provocation in the form of a contest. And the provocation of the contest was AI fiction sucks and what is your theory of how to fix it?

2:20Um but I've come to understand that the real question here um and the question and and the the question I think that is makes this talk I hope relative to relevant to to all of you in this room is that AI slop sucks. Um and so I'm going to share a theory about what about what what it what it is that sucks and what we can do to fix it. And I think um what makes this relevant to everyone in the room is that we all use AI in our work I I imagine. Um and we all know like whether it's in your creative practices, whether it's in your studies, uh your organization, even um you know,

2:59organizing your playlist for your workout. I think we all know that we all kind of are pushing against this slopification. Um and it's it's related to Cory Doctorow's enshittification if you if you've heard that. But the enshittification is a little different, right? Um and I think we're all pushing we're we're we're resisting that in our in our work, I think. Um so in this presentation, I'm going to share a theory about that, about how to solve this problem. And I'm going to do it in the form of a fiction machine that I built for this contest and I'm going to tell you about it. Um but I think this talk is not only about fiction and I do and I

3:38don't even think it's really about AI. I think it's about creativity and um how to make things that mean and matter in a world that is increasingly inundated with this endless iteration of meaningless stuff that we are all, you know, um, bombarded by, right? How do we keep making things that matter? Okay, so, um, the contest is run by this company called hyperstition. And I had come across this term, um, in my doctoral studies. My doctoral studies were all about how we use fiction to build the future. We build futures with fiction, and we build technologies through fiction, through fictioning. And so, hyperstition, um, names this this idea. It's It's a combination between hype and

4:28superstition. And it names the the phenomena where some kinds of fiction that some kind of fictions like, um, like myth, like like hype, like propaganda, get get re-circulated and repeated so often that they start organizing behavior, and then they insert themselves into the real and become real. They become real. The fictions become real.

4:52So, um, uh, AI, now, amazingly, interestingly, is operationalizing this hyperstitional loop. If we could go to the next one, yeah. So, we feed the AI, we train the AI with, um, all kinds of data, including fiction, right? Um, uh, our our our all the huge body of of stories, novels, um, uh, AI imaginaries, um, metaphors, and all this stuff feeds the AI, which then trains the behavior, which then informs the output, which then gets inserted into the real, right? So, there's this hyperstitional loop.

5:30Okay, so, I just want to say to, um, cuz I can hear some people, um, arguing. Of course, this is not a one-to-one relationship. And values alignment in AI is a live field, as we all know. Um, but there is research that supports this um this movement of of AI fic uh fiction um kind of guiding the way AIs LLMs end up behaving.

5:58Okay. Can we go to the next slide? Okay, so back Sorry. Back to hyperstition. So, this company had this idea not crazy and kind of um adorable that all our AI fiction is like um can we go to the next slide? Is apocalyptic. It's like disastrous. AI is going to take over humanity. It's going to like burn down the world, right? Like there's very little AI fiction about like well-behaved, happy, you know, futures with AI. And so their thought, which is not crazy, is like this fiction is going to you know, train the systems towards this apocalyptic imaginary and we need to counteract that with helpful, aligned, safe, well-behaved fiction.

6:53So, they did it. They They made this huge corpus, huge corpus of like well-behaved We'll just stay there for a second. Well-behaved AI fiction and now it's on hugging face. You can download it. And I know you all you're all going to run home and download it and read it tonight, right? No, you're not cuz it's the they I'm sure they encountered the problem >> [laughter] >> that it's terrible, right? Like it's going to be boring as hell. Horrible stuff. Not good fiction.

7:23So, I wasn't surprised very soon after this um I learned about this project that they announced We can go to the next slide. The um onslaught fiction contest. Okay, so this is the contest. And I was very excited to to read about it. So, the the way it works is and we could just fill this whole slide up. So, you build a prompt harness.

7:51You design it, you build it, and all you can do as a human being is enter the tweet length prompt and press go. And then you submit the your output. You aren't [clears throat] So, this is there's no co-design, there's no co-imagining, there's no all the words that we use to talk about how we write with AI.

8:12None of that, okay? So, you just build the prompt harness, press go, and and submit the output. This is not as easy as it seems and the the story that I ended up submitting, there were lots of places where I wanted to like just fix it a little, but no, you're not allowed to do it and it's yes. So, so this is the the setup.

8:37Okay, let me just So, the first [laughter] part of the contest was to submit your theory about what was was wrong with AI fiction and then how you were going to fix it and then they were going to select from those essays and submissions as some select some contestants. I was among them and they gave you $280 Canadian of money to pay for compute, train, and test your system.

9:07So, in order to do this, in order to fix something, you have to understand it, right? So, what is slop? Um and why is it so offensive to our sensibilities, right? It's like gross. It sucks. Why? Um so, I want to So, I'm going to share a theory about about why and then and then my fix.

9:31So, if we could go to the next slide. So, my answer to the question of why AI fiction sucks is actually that it's pretty good. Like, it's fine. It's fine. It works. It has structure, it has style, it has imagery, emotional beats, you know? That's it. It's death by It's death by fine.

9:52Um and I think here, if we can just fill this slide up now, I think it's helpful to make a distinction between something that is bad and something that is slop. And this is an interesting distin- distinction. Not all slop is bad. I I have I have I have some very favorite sloppy artworks that I love. Shrimp Jesus is my favorite. Um right?

10:15Um And certainly not all bad things are slop. So, bad and and as an educators in the room, we want bad in our in our submissions. We don't want slop, we want bad if we're going to have one or the other. Bad is awkward, it's strained, it's embarrassing. Um it can be overdone, underdone. Um but some It's something in a particular direction, right? It has some kind of human pressure in it.

10:42Um it can tell you something about what the person wanted to do, the intention there, um but just couldn't quite make happen, right? Slop is different. Slop is offensive because it lacks intentionality. It's too easy. It works. It gives you the shape of meaning without the necessity of meaning. Um it gives you the feeling of a story without a story.

11:07Um and nothing in it has to be there. So, I think the secret heart of of AI slop is this concept of fungibility, which is the swap-ability. Anything can be swapped out for anything else.

11:27One character for another, one metaphor, one trauma in this endless supply of iterations of fine in slightly different ways, right? Just endless. Um AI is the ever swapper, right? None of these shapes matter. Well, I can just keep giving you shapes till till Tuesday, right? Um just just keep just keep going. Um And and this is the thing. LLMs are made this way for a very good reason.

11:57They are meant to be useful, responsive, opt- optimizing, frictionless. Um they're not meant to stop and be like, "Ooh, which shape should I choose? I don't know. It's like for this context." They're not meant to do that. You don't want your LLM to do that. That's going to mess your flow, right? Um They optimize. They complete. They fill in the blanks. Left and left to their own devices, >> [gasps] >> they go they go to the middle of things, right? Where And that's the that's where slop lives, right? In the middle of whatever you direct your LLM at, they go to the middle. And that's that's where where they land.

12:34Another thing for the next one. Okay, so this effect. So, when I sat down to move from theory to actually building this thing, I realized that I had some pretty obscure language for the kind of story that I wanted to make and how I thought we would get there. I would say things like, "Write me a sentence that resists closure.

13:00That has windows open for meaning to move through." And I soon found out that the LLM had no idea what I was talking about. Um and when it didn't understand, it said, "Uh-huh. Uh-huh." [laughter] And it simply filled things in for me. And that filler read like filler. So, I have this good friend who was going to come tonight, but couldn't.

13:22And she's been my editor over the years, and she's moody and irritable and opinionated. I can say that cuz she's not here. And >> [laughter] >> And she has zero passion for how I sometimes slip into dense and complicated language. Um that can sometimes, it's true, only sometimes, obscure the fact that I don't really know what I'm saying. Um and she always says, "Stop, Rachel.

13:52What are you What are you trying to say? Five words or less." Um the LLM won't do that to you, right? It's not going to say, "What the hell are you talking about?" Sometimes it does, but rarely. Um you know, if we're honest. And it will just quietly nod and say, "Hmm." And uh and fill in whatever imprecision you've left open for it, right? And it does it quietly. If you could just go to the next one.

14:21So, since building my magnificent machine, I have gone back in to develop this talk and look at the wondrous thing I made. And and I've discovered lots of places in in the in the instructions and in the agents that I didn't write. Like, place, you know, um instructions I didn't write. Um all the smoothing that happened in the build. Um where it just smeared across the distance between one point and another and filling fills filled things in with this stuff. And I think this is an important um insight for everyone who works with AI is that you need to be really vigilant about that smoothing and averaging and blurring unintentional aspects to the build that you're working

15:07on. Because I do think it has effects downstream in the output. So, yes. So, I'd like to move on to the next slide. Okay. So, I want to make um the system itself be a kind of artwork. This is the essential irony of the contest. They wanted a work written entirely by AI.

15:30But I think the way to combat AI slop is to reassert the idiosyncrasies and situatedness of human judgment. Yeah, all all throughout Thank you. All throughout the automation um and leave as little up to the LLM as possible. So [laughter] so It's work. Oh, it's work. Uh you're not going to get a story about like optimization here. I'm sorry.

15:59Uh like that's Or this is automation, but it's like hard automation. >> [laughter] >> Okay. So I needed to get really articulate in my instructions to build a container, a pasta machine, if you will, made up of constraints and pressures. Each element of the machine is a macro and micro judgment, um a human judgment, um So the idea is to take the stuff of the LLM, the medium of the LLM, like dough, and design conditions to resist the smoothing tendencies and that drift towards the middle, and maximize the other tendencies because I do love AI and LLMs. They have They are beautiful and they have wondrous capacities. Um their incredible reach, their precision, their their

16:56their their um yeah, their computational mastery, all these wonderful wondrous elements that we can maximize and minimize, you know, the the sloppy stuff. So and here's my my friend put this sounds hard. If we go to the next So yeah, authoring the process and not the product. Moving the authorship, so there's authorship here. It's just in the process and not the product.

17:26So, can we go to the next slide? So, technology can technology reveals process differently. So, if you know this iconic photograph of the the horse's gait, right? Photography allowed us to see the elements of the horse's gait that were you can't see with the human eye and we understood something deeper about the movement of a horse with with the aid of that technology.

17:53Similarly, I think um we can build automation in such a way that it brings us into an understanding of the process that we're that we're automating. So, if we if we go ahead to the next one as I was building this yeah, so we automate to understand a process. The next one? Yeah.

18:15So, drawing, I don't know if there's anyone who draws in the in the house. Um I'm sure many of you do, probably. Yes, good. So, we draw um we don't draw to replace the thing or to solve it um or to make a final statement about the thing or so now we've drawn it, it doesn't need to be here anymore.

18:36Um but we draw to notice, right? We draw to reveal, we try draw to interpret, um to discover. So, this is how I began to think about automation of in this particular instance of the machine was not a a shortcut to the product, but as a sketch of the process of of writing a story.

19:00Okay, so the next well, so we're going to So, in order to do this, in order to articulate the conditions for something like a story to happen computationally, it's kind of cool. Um I needed it to be really clear about what the LLM was going to draw from and what it was going to do with it. And I wanted things to happen along the way, things that happened to me along the way of my process when I'm doing something creative. Um, I wanted false starts. I wanted disagreements. Uh, I wanted choices to be made and then unmade. Um, I wanted play with language and if it's not, you know, play because I don't something like play with language. Um,

19:42I wanted the story to be non-linear and to move towards those non-fungible, load-bearing elements that if you took it out of the story, the story would break. I wanted details if I wanted details that grounded the story in a specific world, not any world. So, now I'm going to show you the system itself and then I'm going to show you some of the inner workings um, with one of the stories that that it generated. So, we'll go to the next slide.

20:20So, the first So, the world. This is what what the story cares about. And so, before that before anything happens, before anything is written, there's all all the stories in the Antelope system are deeply interested in AI and the impacts these technologies have on human beings. So, Antelope stories take it as a given that AI is already making changing human experience um, and changing ordinary human experience. So, then we have the corpus.

20:51This is what the story remembers. So, in the system that I designed, there are two persistent um, bodies of text. One of them is the theory. So, I have a theory agent that goes out into the internet and finds open-source um, wild, uh, theory about AI, recent research, it pulls that theory in and that it generates ideas and it adds us to adds this to the corpus that keeps growing as the stories, um, perpetuate.

21:21It's [snorts] cool. Okay, thanks. >> [panting and laughter] >> And then I'm going to show you a little bit from it. And then, um, the other body is my own writing, um, and it's indexed and chunked into little into little bits that can be semantically, um, used as context for voice. So, the drafting agent draws on that for for a voice.

21:41Then we have the seed, which is the tweet-length, um, story idea and this the story the system that I built has gone through six generations and each seed has some kind of it has a human and a technology in some kind of, um, conflictual relationship. Um, and then we have the agents. So, the agents are designed ways of paying attention.

22:04And there's eight of them in this in this system. And they have processes. So, that's what the agents do. So, they search, they read, they deposit, they interview, they draft, they annotate, revise. I have one that just cuts. Um, and then each of these processes creates an output. And so, that's the traces of the processes. And so, all of these outputs are, um, um, made into, um, markdown files in a file structure so each story, and I'll show you in a sec. And then I can go in and read those outputs and if I don't like them, I go back to the instructions and tweak the to So, the whole the whole, um, journey to the drafts, the final

22:53element is the drafts. Each story has three drafts and a final PDF and you can see the story changing. So, you can see the story sort of emerging all through through this process. So now I'll just show you the file structure. So here is um yes. So that's the that's one story and you can see all the elements, the drafts, the characters, um the notes, the letters, the theory deposits, and the faces.

23:21Okay. So I'm going to um I'm going to share the the story seed for a story. And then I'm going to give you a rich synopsis of the story and then show you some pieces in between the seed and the final thing, okay? Um so this is the the seed. The custodian of a regional archive is told the AI will preserve everything. She begins quietly to misfile documents testing what counts as preservation when nothing can get lost.

23:53So that's the the seed. Now I'm going to give you a rich um synopsis and I'm just going to read it to you. Read that one. Now it's story time. Okay. So So Margie So the story is called The Hand That Will Not Lift and the stories get that their titles as part of the their output. It doesn't start out with the title, obviously.

24:18Okay. So Margie is the custodian of a regional archive. And it's almost this magic realist archive. I didn't give the the system any information about that, but it just came out that way. It was really cool. Um so like containing all the paper uh of a community, the parking tickets, the personal letters, all the stuff in the and even um um you know, pins and different things.

24:42And so Margie um is this custodian worked there for years and she helps bring in an AI system meant to preserve and retrieve everything in the collection and to automate her job, essentially. So at first the system seems efficient and accurate, and but it begins surfacing documents that Marjeet has deliberately allowed to remain hidden.

25:05Letters, affidavits, and records whose placement [clears throat] in the archive has been shaped not by formal procedure, but by her ethical judgment, privacy, grief, and even care. So, her old practice, she had this practice of misfiling um elements in the archive, and it wasn't negligence, it was her way of honoring informal requests from uh protecting descendants, even recognizing that some records, though preserved, should not be easily found.

25:37So, as the AI makes everything searchable, Marjeet starts to lose her sort of intuitive bodily sense of archival responsibility. >> I didn't write this. >> Yeah. I know. Um so, her hand that once knew where things belonged to no longer knows what to do. And the crisis becomes personal when Marjeet realizes that the same system that exposes other people's buried histories can also expose the lie she has maintained for 33 years about her brother Peter's death.

26:12She has been protecting her elderly father by letting him believe that Peter is alive elsewhere rather than letting him know that Peter died by suicide in '91. When Marjeet goes home one evening and searches on her personal access to the system from home, searches her brother's name, she the the archive returns the police report, the coroner's report, um the newspaper clippings, and she understands in that moment that her private act of protection has now become part of the system's retrieval record.

26:46And the story ends with Marjeet suspended between action and inaction. Her father has called unexpectedly, and she can't decide whether to answer with truth. Um her her hands still won't lift. Um and so yeah, >> [laughter] >> the story is about um about preservation versus protection. Um it's about the ethical the archive as an ethical practice which is, you know, very human.

27:16Um and then, you know, and AI as this the end of informal forgetting. It's pretty cool pretty cool, right? Okay. So, now we're going to go back and I'm going to introduce to you the beginning of the story from the seed and what we did. So, the next one, please. So, [clears throat] from the seed, the first there's two agents that go out.

27:42Yes, just stop here for a sec. Okay. So, the the character agent whose name is is Francis, um that's an important detail. Um is multimodal and it draws three images at random from a bank of 67 portraits that I'd pulled from Unsplash database and the photos are all openly licensed and the system retains the the metadata and the photographer credits and everything.

28:10And the character agent builds a character profile with a with a physical description of the person, a photo, a history and um and builds their voice through through an interview. So, for this system, the the this portrait was was drawn. This is Margit. If you could go to the next one. So, this is Margit Halloran, the archivist and protagonist of Selective Forgetting.

28:38Uh well, that's the first name of the story. Um so, I'll just give you a brief history. She's from She's from Duluth, Minnesota. She's got an Irish-American father and a Hungarian mother and this is where the detail about her um um her brother having been born in '68 and having died of an overdose in a in a motel.

29:01And the central contradiction of her life that the character that the character agent composes is she is the most rigorously honest person in the in any room she enters. She will correct a citation, a date, a misattributed quote. Um and yet she's been lying to her father about her brother for 33 years.

29:20In her father's mind, he's in Oregon doing well, just not good at calling. She invented a job for him. She invent invented a girlfriend. So it goes on. There's all this rich stuff in there and not all of it gets storied into the final output, but these are some of the details that the later agents picked up on. So at the same time, the theory agent goes out and this is I think my favorite part of the [clears throat] the system. The theory agent goes out and draws from like if anyone's into AI research, like there are so many crazy, weird, cool stuff being published every day about these weird beasts. Um so the theory agent goes out

30:02and draws those papers, three of them, and then creates these theory deposits. So this one paper, so for this story, there's a paper called how do language models learn facts, um dynamics, curricula, and hallucinations. So I'm just going to give you a little brief synopsis of the paper. Um so large language models learn in sharp phases, not smooth phases. The same fact can be learned through different mechanisms that interfere with each other. Hallucinations according to them aren't random noise.

30:36They're structured failures where facts learned early on get masked by later training in predictable patterns. So, um the theory agent then takes that theory and thinks, "What does this mean to the to the humans in the story?" So, the the theory writes writes this. So, the custodian doesn't know it, but she's stumbled into the same insight. This AI system has been told to preserve everything, but everything isn't a flat list. It's a sequence. It's a curriculum.

31:08What gets ingested first becomes the substrate. What comes later can mask it without removing it. By misfiling elements in an archive, she isn't destroying facts. She's rewriting the order in which the system encounters the world. If the paper is right that hallucinations are structured failures, her misfiling produces structured truths of a different kind. A document in the wrong drawer doesn't disappear. It grows a different neighborhood. It grows a different neighborhood.

31:40The marriage license next to the demolition order, the land deed next to the deportation list, new facts emerge, not invented but composed by adjacency. Um Right? It's really So, it's really cool stuff. Um and and the the story the the stories are are full of these juicy bits of like and this is the point this this AI theory tells us about our lives and like what it is to to know things and remember things.

32:16Uh yes, anyways, the point I wanted to demonstrate by showing you these two agents is that I tried to make the system answerable to specific things in the world, specific faces, specific research, specific material and social realities. And when you point the system at something specific like that, like the stuff that it came up with was nuts.

32:40Okay, so we can just like zoom through this one. Yeah. >> Where are you at? >> Uh sorry, am I Am I almost done? >> I'm just asking where you're at. >> Um I'm almost done. Two more builders. >> Because I got 10 questions and I bet all of them do too. >> [laughter] >> Okay. Well, let me let me just let me get there. Okay, this is tofu by the way. Okay, so one thing I really did not want the system to do was tell it to write like something.

33:07Um I think part of what turns our stomach about AI slop is that it surfaces as disguised um it's pretending, it's trying to trick you. Somebody's tricking you, they telling you they wrote it and it wrote it, right? Like it's this it's it's this disguise of human labor. Um which hides an absence somewhere of care, uh thought, or intention. So, when the LLM and I don't know if you've experienced this, when it thought it was writing literature Oh my god, the the most disgusting sentences of all time, right? Just vapid, shiny stuff. Um it could mimic literature and poetry. Um but that mimicry is the problem. It's all convention, no intention. Um it sounded like literature without being any about

33:58anything. But the system became much more interesting when I forced it to stop pretending and instead do the things it was good at, right? It's precision, its specificity, um its unexpected unexpected connections. It could pull details I'd never heard of. It could It could make a mother speak Romanian to her child. Um you know, it knows all the side streets in Sudbury, Ontario.

34:25So, I thought a lot about cooking the pasta and also the tofu, this one. So, um in cooking you transform ingredients towards an output. You bring skills, sequence, you work with that ingredient, and you you create the conditions for the flavors to happen, right? So, tofu is not failed meat. Um, but when you dress it up like a turkey, um, you know, or when you say this is bacon, um, it fails because you sit down and taste for the thing that it is not, right? And you're always going to find a lack.

35:02Um, but yeah, tofu is not failed meat, and um, LLMs are not failed humans. Um, we need to when if you're reading the LLM output for a human voice, I think that's where we find that icky like, ooh, it is, but it isn't. Um, but it's computational. It's not It's not human, and that's its That's its strength, I think. The point is not to hide the LLM, hide the computational nature.

35:30Um, the question is how to cook with the medium that is the LLM, and create the conditions for its own strange, specific, non-human capacities to come to come through [clears throat] with our human care and intention. Okay. So, I'm not going to read the this passage, but if you're interested in the story, the full story's here if you if you want, and I can I can share it. Um, and and other ways.

36:02Um, but this was the second So, I um, this was the second my second favorite story. My first favorite story is is still um, being um, adjudicated. >> know yet. >> We don't know yet. Uh, sorry. No, don't know yet. Okay, so I want to I want to turn return So, yeah, it's not the point. Winning is not the point.

36:26It is. 10,000 bucks, too. >> Woo! It is. We'll see. We'll see. I'll let you know. Um but I want to return again to the the the the meat, the tofu of the issue. Um So I want to return to hyperstition. Um to the idea that facts don't simply represent the world. They enter the world fiction rather, sorry, not facts.

36:47Oh my god, I don't care about facts. Um >> [laughter] >> They enter the world and begin to produce effects there. And I This is where slop comes in and becomes more than an aesthetic problem. Slop is a hyperstitional force. It circulates, it repeats, it normalizes, it trains us towards the middle, towards the normative, um towards the exchangeable. It teaches systems and it teaches humans, too, I think, that fine is enough and that specificity is optional and that everything is replaceable without consequence.

37:23Woo. So the danger of slop is not that it sucks, it's that it's world-building. And um the question then becomes an existential one. How do we continue to make things that matter in a world inundated by slop? And this is not to turn away from AI at all, but understanding and employing it with deep human intentionality.

37:46So I'd like to leave you with three takeaways. Um First, I think that slop is diagnostic and we need to look at it more and understand it more. It shows us where process has been too generic, uh too smooth. It reveals absence of pressure and it tells us about the conditions of its origin. And I think we need to get uh serious about uh diagnosing it and and understanding it.

38:10Second, moving authorship upstream can return us to our own human processes. So building this machine brought me back to my my creative process in a in in profound way, um which is central to my humanity. And I I do believe that we can use these systems to deepen our understanding of ourselves as as human beings. I don't see the tension between us there.

38:33And finally, LLMs are not failed humans. And that is their strength. Um, we when we ask them to pass as human, they become artificial. And this is my problem with the term artificial intelligence. We are always looking at the LLM as if we're trying to see human intelligence and it's artificial. But But that's not what it is, right? Um, we sense the lack in the human lack in them because we're looking for the human in them. But if we just sat down and and with them as computational systems, um, then the And that's what they are, right? Um, then the real possibilities, I think, start to emerge.

39:14And the point is not to make the machine human. The point is to work with its other-than-human nature. And I don't think there's We can talk about, you know, binaries and all that at another time. I won't go on. Okay. >> [laughter] >> Um, but the point is that we can learn through these systems. We can learn more about ourselves.

39:32So, I just like to, uh, leave with a small gift. Um, all the diagrams in this, um, presentation, uh, were created with a vibe-coded, um, diagramming platform that I've been building. Um, and uh, Nessa encouraged me. I'm going to put a tutorial on it. Um, but it's quite intuitive and easy to use and it's super fun.

39:54Um, and if you do make something cool, I would love you to reach out and share it with me. THANK YOU VERY MUCH. >> [applause and cheering]

40:07>> PLEASE [cheering] STICK AROUND. >> [applause] >> STICK AROUND, DR. HURST. I don't think we're done with you yet. Um, Those are beautiful. >> Thank you. >> I really I realize you used AI to generate them, but I feel like all of us should develop like a personal style for our own >> No, I didn't use AI to generate them. I used AI to vibe code the platform that has no AI involved. But it's it's very different.

40:35>> Fair enough. Fair enough. Beautiful. Beautiful. Um >> I definitely used AI, yes. >> Guilty. >> Vibe coded. But the platform itself doesn't use AI. >> That was my favorite Vancouver AI talk in 29 months, and we've had some [ __ ] badasses here. >> Thank you. >> [applause] >> Thank you. Thank you very much.

40:58>> Experts agree coding has been solved. >> [laughter] >> Hear me now, believe me later. Experts agree coding has been solved. What hasn't been solved is the qualitative stuff that we're talking about here. But we're applying the same agentic processes. Your work tree, your your file folder structure looks exactly like when I'm building software.

41:23But there's a lot of us, Kevin included. I kind of want to throw to you, too. If you Is there an extra mic? >> Yeah. >> Yeah. Um >> I killed AI. >> We stole those from SIGGRAPH, so it's okay. They're going to die eventually. It's the It's okay. It had no soul anyways. Yeah.

41:42>> [laughter] >> So, you are agentically trying to solve creativity in some ways. >> Well, I would just Can I just say something? Um I think that I'm not trying to solve anything. Um my design of that system would look entirely different than every single person's in this room. And if I sat down to do it again, it would be entirely different.

42:06>> Good. Don't give away the secret yet because Kevin spent like 6 months in Hollywood recently working with a company that's trying to develop like a genetic movie writing capabilities. Take that seed, >> Right. >> develop an anti-slop script and treatment, >> Yeah. >> then use generative techniques that he's pioneering already to kind of like make like why why does your story got to be the end? Why not have like a couple clips for a film festival coming on the other side? So like could you the reason I asked you to come up here is like I know you've been messing with this stuff in a different way and almost like a whole epoch ago, 6 months ago.

42:40>> Well, yeah. Um what I actually marveled about the structure that you showed was like, you know, uh it was NDA safe, marginally close to what we were tackling, you know. >> Sorry, I didn't mean to out you. >> No, no, you're good. >> [laughter] >> No, and and I want to remark on that like, you know, what what what you have looked at as like story as data is is in every agentic system that's out there now and certainly what we were messing around with and is about to be unveiled as well.

43:12So it's it's quite fascinating that in all of these sort of like silos of self-belief and self-development that there is a consensus of just how the you know, story and and the agentic sort of like accoutrements and touch points on it is wildly similar where it says that like the humanity of storytelling has a very common consensus in my mind, right?

43:40>> Interesting, very interesting. Um I would hazard that I would get fired from that job really fast. >> [laughter] >> Because I built in like the story started costing a lot of money. >> Mhm. Yes. >> And I >> Not not as much as a writer's room, which is what we're talking about.

44:02>> No, yeah. >> And you know, you You if you got a sharp seasoned writers' room, I'd fair say in the revision process that is now amplified by these sort of cheap agentic processes that the token costs are almost as good as a mid-level crack writers' room these days. >> I didn't mean to cut you off, Dr. Hurst.

44:22>> Um no, so I'm just thinking about it. So so like the the this Unslop system, the stories that it generates are all sort of within the world of this Unslop thing. So it's almost like its own like um you know, genre or something like that. But then you would get tired of those and then you'd have to build a different system pretty fast, right?

44:43>> Absolutely. Absolutely. We found that for sure. Yeah. >> it. And also, yeah, yeah, yeah, yeah. This is a big a big uh conversation. >> It quickly becomes a trope machine. Absolutely. >> yeah. You guys both step up here for a sec. So now we're going to do some Q&A, William Jordan.

45:05>> Thanks, Dr. Hurst. Just quickly, you were in music at one point. I'm wondering what that artistic process helped you like inform about this process. >> Mic in your face. >> Sorry. Yeah, mic in my face. Sorry. >> Also, like could you use this for music as well or have you considered using it for lyrics or concept albums or something like that?

45:25>> Um well, I haven't I haven't considered that. I think that one could. Um one of the Actually, well, one of the um inspirations for this for this idea was um of course now I'm What's her name? What's her name? Um I'm forgetting her name because I don't have it written down and there's all these people in the room and I forget my own name when that happens.

45:49But anyways, Herndon, Sally Herndon. Anyway, is that her? Yeah, she She is a musician who um has trained uh trained in AI on her own voice and then you can sing. So I I think the point is not I the point is looking for creative ways to use AI to to do art in different ways. Um and I think yes, musicians doing this all the time. And and perhaps that I don't know um um composition the same that I that I do with writing. So, I can't I can't really speak to that, but I know there are artists out there doing like really wonderful, exciting stuff with music, for sure.

46:28>> Dr. Horsham in the back of the room here with Wayne. Hey Dr. Horsham. Um sorry. Everybody in this room is probably much smarter than me, but I'm still trying to wonder about like I feel like you kind of showed us the answer, but I want to see the work, in a way. Do you know what I mean? Like how did you come up with the idea of like, oh, this is the problem I want to solve?

46:51>> Where does he start? >> Where do you Yeah, where does one start with >> Well, yeah. >> How do you What did you decide to do? Like, oh, okay, here. And what AI What LLM you're going to use? That kind of thing. >> Yeah, well, what I said like it started with write me a sentence with windows and the LLM being like, like I don't understand. And I So, I like I was like, oh, yeah, I'm going to make this system. I'm going to join this contest. I'm going to do this thing. And when I sat down to do it, I realized, oh, I have to articulate what I mean by like what do you So, then I thought,

47:25okay, what do I start with? Well, I will often like it's not like I'm a you know, anyway. But when I go write a story, I I look at things in the world. I go look at pictures. I think about people. I look at you know, I take notes. And that's how And I And I think about theory. And that's why I started with the character agent and a theory agent, cuz I wanted I wanted to start with character. And I wanted to start with a juicy, real thing that that technology does to complicate human experience. And so, that's where I started. But you would start somewhere completely different.

47:56>> Even like the step before that, where did you start? So, like you sat down in front of your computer and you opened your web browser and you went to chat GPT? >> Oh, I used So, I used I used Claude code and I used the $280 to to buy tokens. Um and then I and I and then I worked with Claude um the chat interface and we we talked and talked and talked and talked and >> Um talk through your fingers or talk through the voice recordings?

48:23>> to talk through my fingers, yeah. Um >> I hate talking through my fingers. >> Yeah, I I I prefer it, but um I find and the the voices of these things are not good yet. Chat GPT's voice is so terrible. >> listen to its voice back. You don't put in voice mode. You just record your voice because your fingers are a filter between your head and your heart and imbue yourself into these systems.

48:45>> Anyways, yeah, but I think I think there's a lot of different ways to go about it, but that's how and part of part of what got me there was I was also trying to think about my academic and how and so I'd already been building an a knowledge infrastructure of research and um literature notes and different ways to to write academic papers. So, I was already in that space.

49:08>> So, just to literally follow >> Yeah. >> talk to it? >> I I would talk to Yeah, I was a lot of talking through my fingers. >> You didn't get to it with your finished ideas. You developed your ideas in relationship >> Yes, but I but I wrote the the um the instructions. So, when we got down to I was like, okay, well, I need a character agent and it needs to be able to do this and I don't want to give it too much information cuz that messes things up, too. Uh so, I need to think about what what were the essential pieces, what I wanted that agent to pay attention to.

49:42And this is a really fun thing cuz you have to decompose your creative process and think about where to put point your attention. Um it's just really cool. It's I like it's a it's a cool like I encourage you to just take a "How do I brush my teeth?" And like decompose it into an automation and you're like, "Oh my god, I think about God when I'm brushing my teeth." Like whatever, you didn't know.

50:08>> Yes. >> You didn't know how deep the process was until you tried to make it. >> that came up for me during your talk is you cannot automate It was when you were asking questions. It's like you can't automate a process you don't understand and that's how we get slob. It's you're like, "Yo, write me a [ __ ] story."

50:23>> Yeah. >> Yeah, and you don't know how to write a story or what a story is to you, then you definitely cannot automate that into a pipeline or >> Yeah, no, you have to have >> step of automation is to sit there and break down the brushing your teeth into all of its constituent components such that you can relate that to a machine.

50:38>> That's right. So, in order to automate something, you have to have a theory of that thing. And this is what I understood that I wasn't do I didn't have a theory of creative process. That's where you get slob. Creative process. It's like everything. It's nothing. It means nothing. A creative process for this moment, for this story, for this thing.

50:59>> I was hoping you would have that question. This is Kush. You never know what he might say. I I appreciate your talk very much. Um, first of all, let me just say really, really appreciated the AI slob let let let interrogate it part cuz I feel like any any effort to understand why it is that make it makes us feel that way is like self-knowledge or self-learning. I appreciate that so much. I'm going to investigate your your project more.

51:24>> Wonderful. >> Um, but so so with with that's just a slob part. The other side is like I guess I'm wondering to sort of echo what Chris was just saying, I feel like when people are using AI these days and they find it to be competent in anything at all, whether it's engineering, whether it's writing, whether it's image generation, I find more often than not, maybe this sounds cynical, it's coming from a place of not knowing what the [ __ ] is going on. And it's like, "Oh my god, it's made made me the best software ever." Are you a soft Are you a person who cares about software or you don't? You're enjoying

51:52this? Great. Yeah, because you have no taste and you have no interest. And so you're very happy with the output, right? So similarly, I feel like when it comes to writing or when it comes to anything >> You have to have a You have to have a knowledge of that thing. >> If you have a knowledge of [ __ ] [ __ ] knowledge. Like if you don't care to do the thing, what the [ __ ] do you care? Like we should not We should frankly look down and shame Like we should have like, "Yeah, [ __ ] you for writing when you don't give a [ __ ] about writing, right?" Talking about writing

52:13software. >> want to I want to go back a little bit. >> Yes. So so I feel like what are your I wonder what your thoughts are about like people just doing it for the money or for the something and So it's not like don't do it for the love of it, but don't do it for the money of it. Don't do it for the Don't do it for the You know what I'm saying?

52:28>> I just think I I I hear you. I agree and I and I get angry, too. But I do think that we have too much shame right now around AI. And I want to not dig into that. We don't know what people are doing with their AIs. You know, it's like we don't know.

52:49And um And it could So we like I I'm a teacher, okay? And there are teachers out there who are like, "These students are just handing in stuff. They don't care about anything. It's all AI slop." And actually they're working and they're figuring stuff out and they you know, they they ha- it has an AI sheen, but they brought their thoughts to it. And so I It's not my job to divine their intention, right? Um and so and I want to be really careful about that because we are in this transition period and there is so much shame and anger about AI use. So I don't like so So I don't want to get But I I hear you and I agree

53:30because um you know, people will say like I I'm just thinking of this picture in my mind. You know, what's that Ethan Mollick, that guy? He's He's wonderful. But I don't know if you watch him if you follow him on LinkedIn, he he posted this thing and it was like all these famous um paintings in like um designer dresses and then it was in so they were wearing the famous painting and then it was in a desert, right? And it they were beautiful. And I was but then I was like yeah, but like would a designer think so? Would a fashion designer think that dress was like and all the women and then I you know, you take one look and

54:06you're like you but it's that slop thing, right? Like if you don't care about fashion, it just looks like fashion. It looks like fashion. Right? Yeah. So, yes. Yes. Right. But, why are they doing it? They're like I can do fashion. >> [laughter] >> Yes. Yes. Yes. Yes. Yes. But, maybe it will bring you into care. So, that's the other thing. Um you know, yeah, you so so as you you build one thing and you're like, oh, I love that.

54:43Now you want to build something else and now you and then now all of a sudden you're designing a dress, right? Um and then we also know that there was a math there's a we we've talked about the the AI delusions and stuff and how could you ever think that you were that smart? You didn't go to university. Um you know, the the AI made you think you're crazy and yet there's this guy recently who just discovered this math thing through like for real cuz he's really smart with his AI. Like people can do Right? Yeah. So, so like play. We can play. We can we can play with these things and we can make slop. Make slop.

55:22But, but learn. Learn about yourself through it. Yeah. Yeah. Yeah.

55:35>> Hello everyone. Awesome Awesome talks that was really great. I I felt you were talking about meta creation and that's not because I'm the director of the meta creation lab. >> [laughter] >> But you know, we're in this world where now we get to design the creative process. And I have two questions for you.

55:50>> Yeah. >> These are the classic ones but they're the important ones. >> [laughter] >> We should work on it. The first question is through meta creation did you find that what you implemented with those machines is a creative process you would have done yourself? And the second question, do you think it would have been different if instead of agent you would have a team of humans to implement that exact process?

56:12>> Oh gosh, yeah, great. Yeah, totally different if I was doing it myself, 100%. I read >> [laughter] >> I read the synopsis to my husband the other night. He's like, you didn't write that. I was like, I know I didn't write that. Like I could never have written that. Um so yeah, 100% totally different um working with the with the LLM and I had to speak in a way that is not my like I like to be, you know, complicated and I had to be very, you know, articulate and straightforward about my instructions. So but that taught me something and I don't think I would go continue writing like, you know, agents everywhere but it taught me

56:55something about my creative process. And then your point about the humans, I developed a game based on this. >> Is your meta creation process the same as your creation process? >> Oh, totally different cuz I cuz I don't I don't break down my creative process in those in those ways. Um you know, the horse runs.

57:16>> Because you've never needed to automate it before. >> Yeah, yeah. The the I'll I'll just run. I'll just I'll continue being a messy human but but return to myself in a different way and I think I think that when I'm if I'm editing or if I'm stuck or if I might bring my LLM nature into it, maybe, you know? But But I Yeah, I'm still human.

57:38>> I got one more quick one from Finn down here. >> Um Yes. Uh thank you for your talk. And I just wanted to mention um because we saw a kind of alternative version of this last week in the education group. And I got a chance to do that story making game. >> That's what I was going to say to Patrick.

57:59>> And I was just I wanted to mention that because I hope you carry on developing that because that game is fantastic because it gives everyone an insight into what the AI is doing as it's as it's adding those layers of slop onto the creation. >> Yeah, that's right. So I I developed a game because of this. I developed a card game where you build an analog LLM with human beings. And the human beings are the agents. And you put you write something together, sorry. Um and it's it's really interesting. It's fun. Yeah.

58:30>> All right. >> Oh, yeah, that's it. Thank you so much. >> [ __ ] eh, thanks, Dr. Rachel Harris. >> [applause] >> Can I give you a hug? Thank you. >> [applause] >> She lives on the Sunshine Coast. She came down and test drove that talk at our AI and education meet up last week.

58:47Thank you so much to like a group of like 20 people. Thank you. Uh lives on the Sunshine Coast. So she's come down twice in the last week to participate in this, develop her ideas, share them. Thank you so much. And um