Soylent Culture

In 1964, Marshall McLuhan described how the content of any new medium is that of an older medium. This can make it stronger and more intense:

The content of a movie is a novel or a play or an opera. The effect of the movie form is not related to its program content. The “content” of writing or print is speech, but the reader is almost entirely unaware either or print or of speech.

Marshall McLuhan, Understanding Media (1964).

In 2024, this is the promise of the generative AI tools, that we currently have access to, tools like ChatGPT, Dall-E, Claude, Midjourney, and a proliferation of others. But this is also the end result of 30 years of new media, of the digitalization of anything and everything that can be used as some form of content on the internet.

Our culture has been built on these successive waves of media, but what happens when their is nothing left to feed the next wave?

It feeds on itself, and we come to live in an era of Soylent Culture.


Of course, this has been a long time coming. The atomization of culture into it’s component parts; the reduction in clips to soundbites, to TikToks, to Vines; the memeification of culture in general were all evidence of this happening. This isn’t inherently a bad thing, it was just a reduction to the bare essentials as ever smaller bits of attention were carved off of the mass audience.

Culture is inherently memetic. This is more than just Dawkins’ formulation of the idea of the meme to describe a unit of cultural transmission while the whole field of anthropology was right over there. The recombination of various cultural components in the pursuit of novelty leads to innovation in the arts and the aesthetic dimension. And when a new medium presents itself, due to changing technology, the first forays into that new medium will often be adaptations or translations of work done in an earlier form, as noted by McLuhan (above).

It can take a while for that new media to come into its own. Often, it’ll be grasped by the masses as ‘popular’ entertainment, and derided by the ‘high’ arts. It can often feel derivative as it copies those stories, retelling them in a new way. But over time, fresh stories start to be told by those familiar with the medium, with its strengths and weaknesses, tales told that reflect the experiences and lives of the people living in the current age and not just reflections of earlier tales.

How long does it take for a new media to be accepted as art?

First they said radio wasn’t art, and then we got War of the Worlds
They said comic books weren’t art, then we got Maus
They said rock and roll wasn’t art, then we go Dark Side of the Moon (and Pet Sounds, and Sgt Peppers, and many others)
They said films weren’t art, then we got Citizen Kane
They said video games weren’t art, and we got Final Fantasy 7
They said TV wasn’t art, and we got Breaking Bad
And now they’re telling us that AI Generated Art isn’t art, and I’m wondering how long it will take until they admit they were wrong here too.

But this can often happen relatively ‘early’ in the life-cycle of a new media, once creators become accustomed to the cultural form. As newer creators began working with the media, they can take it further, but there is a risk. Creators that have grown up with the media may be too familiar with the source material, drawing on the representations from within itself.

F’rex: writers on police procedurals, having grown up watching police procedurals, simply endlessly repeat the tropes that are foundational to the genre. The works become pastiches, parodies of themselves, often unintentionally, unable to escape from the weight of the tropes they carry.

Soylent culture is this, the self-referential culture that has fed on itself, an Ourobouros of references that always point at something else. The rapid-fire quips coming at the audience faster than a Dennis Miller-era Saturday Night Live “Weekend Update” or the speed of a Weird Al Yankovic polka medley. Throw in a few decades worth of Simpson‘s Halloween episodes, and the hyper-referential and meta-commentative titles like The Family Guy and Deadpool (print or film) seem like the inevitable results of the form.

And that’s not to suggest that the above works aren’t creative; they’re high examples of the form. But the endless demand for fresh material in the era of consumption culture means that the hyper-referentiality will soon exhaust itself, and turn inward. This is where the nostalgia that we’ve been discussing come into play, a resource for mining, providing variations of previous works to spark a glimmer in the audience’s eyes of “Hey, I recognize that!”

But they’re limited, bound as they are to previous, more popular titles, art that was more widely accessible, more widely known. They are derivative works. They can’t come up with anything new.

Perhaps.

This is where we come back to the generative art tools, the LLMs and GenAIs we spoke of earlier. Because while soylent culture existed before the AI Art tools came onto the scene, it has become increasingly obvious that they facilitate it, drive it forward, and feed off it even more. The AI art tools are voracious, continually wanting more, needing fresh new stuff in order to increase the fidelity of the model, that hallowed heart driving the beast that continually hungers.

But the model is weak, it is vulnerable.

Model Collapse

And the one thing the model can’t take too much of is itself. Model collapse is the very real risk of a GPT being trained on LLM generated text. Identified by Shumailov et. al. (2024), and “ubiquit(ous) among all learned generative models”, model collapse is a risk that creators of AI tools face in further developing the tools. In an era of model collapse, the human-generated content of the earlier, pre-AI web becomes a much valuable resource, the digital equivalent of low-background steel sought after for the creation of precision instruments in an era of atmospheric nuclear testing, where the background levels of radiation made the newly mined ore unsuitable for use.

(The irony that we were living in an era when the iron was unusable should not go un-noted.)

“Model collapse is a degenerative process affecting generations of learned generative models, in which the data they generate end up polluting the training set of the next generation. Being trained on polluted data, they then mis-perceive reality.”

(Shumailov, et. al., 2024).

Model collapse can result in the models “forgetting” (Shumailov, et al, 2023). It is a cybernetic prion disease. Like the cattle that developed BSE by being fed feed that contained parts of other ground up cows sick with the disease, the burgeoning electronic “minds” of the AI tools cannot digest other generated content.

Soylent culture.

But despite the incredible velocity that all this is happening at, it is still early days. There is an incredible amount of research being done on the effects of model collapse, and the long term ramifications for it on the industry. There may yet be a way out from culture continually eating itself.

We’ll explore some of those possible solutions next.

Implausipod E0012 – AI Reflections

AI provides a refection of humanity back at us, through a screen, darkly. But that glass can provide different visions, depending on the viewpoint of the observer. Are the generative tools that we call AI a tool for advancement and emancipation, or will they be used to further a dystopic control society? Several current news stories give us the opportunity to see the potential path before us leading down both these routes. Join us for a personal reflection on AI’s role as an assistive technology on this episode of the Implausipod.

https://www.buzzsprout.com/1935232/episodes/13472740

Transcript:

 On the week before August 17th, 2023, something implausible happened. There was a news report that a user looking for, can’t miss spots in Ottawa, Ontario, Canada, would be returned some unusual results on Microsoft’s Bing search. The third result down on an article from MS Travel suggested the users could visit the Ottawa food bank if they’re hungry, that they should bring an appetite.

This was a very dark response, a little odd, and definitely insensitive, making one wonder if this is done by some teenage pranksters or hackers, or if there was a human involved in the editing decisions at all. Because initial speculation was that this article – credited to Microsoft Travel – may have been entirely generated by AI.  Microsoft’s subsequent response in the week following was that it was credited due to human error, but doubts remain, and I think the whole incident allows us to reflect on what we see in AI, and what AI reflects back to us… about ourselves, which we’ll discuss in this episode of the ImplausiPod.

Welcome to the ImplausiPod, a podcast about the intersection of art, technology, and popular culture. I’m your host, Dr. Implausible, and today on episode 12, we’re gonna peer deeply into that glass, that formed silicon that makes up our large language models and AI, and find out what they’re telling us about ourselves.

Way back in episode three, which admittedly is only nine episodes but came out well over a year ago, we looked at some of the founding figures of cyberpunk and of course one of those is Philip K Dick, who’s most known for Do Android’s Dream of Electric Sheep, which became Blade Runner, and now The Man in the High Castle, and other works which are yet un-adapted, like The Three Stigmata of Palmer Eldrich, but one of his most famous works was of course A Scanner Darkly, which had a Rotoscoped film version released in 2006 starring Keanu Reeves. Now, the title, of course, is a play on words from the biblical verse from One Corinthians where it’s phrased as looking “through a glass darkly”, and even though there’s some ambiguity there, whether it’s a glass or a mirror, or in our context, a filter, or in this case a scanner or screen. With the latter two being the most heavily technologized of all of them, the point remains, whether it’s a metaphor or a meme, that by peering through the mirror, the reflection that we get back is but a shadow of the reality around us.

And so too, it is with AI. The large language models, which have been likened to “auto-complete on steroids”, and the generative art tools (which are like procedural map makers that we discussed in a icebreaker last fall) have gained an incredible amount of attention in 2023. But with that attention has come some cracks in the mirror, and while there is still a lot of deployment of them as tools, they’re no longer seen as the harbinger of AGI or (artificial) general Intelligence, let alone super intelligence that will lead us on a path through a technological singularity. No, the collection of programs that have been branded as AI are simply tools what media theorist Marshall McCluhan called “Extensions of Man”, and it’s with that dual framing of the mirror held extended at our hand that I wanna reflect on what AI means for us in 2023.

So let’s think about it in terms of a technology. In order to do that, I’d like to use the most simple definition I can come up with; one that I use as an example in courses I’ve taught at the university. So follow along with me and grab one of the simplest tools that you may have nearby. It works best with a pencil or perhaps a pair of chopsticks, depending on where you’re listening.

If you’re driving an automobile, please don’t follow along and try this when you’re safely stopped. But take the tool and hold it in your hands as if you were about to use it, whether to write or draw or to grab some tasty sushi or a bowl of ramen. You do you. And then close your eyes and rest for a moment.

Breathe and then focus your attention down. To the tool in your hands, held between your fingers and reach out. Feel the shape of it, you know exactly where it is, and you can kind of feel with the stretch of your attention, the end of where that might actually exist. The tool has now become part of you, a material object that is next to you and extends your senses and what you are capable of.

And so it is with all tools that we use, everything from a spoon to a steam shovel, even though we don’t often think of that as such. It also includes the AI tools that we use, that constellation of programs we discussed earlier. We can think of all of these as assistive technologies, as extensions of ourselves that multiply our capabilities. And open your eyes if you haven’t already.

So what this quick little experiment is helpful in demonstrating is just exactly how we may define technology. Here using a portion of McLuhan’s version. We can see it as an extension of man, but there have been many other definitions of technology before. We can use other versions that focus on the artifacts themselves, like Fiebleman’s  where tech is “materials that are altered by human agency for human usage”, but this can be a little instrumental. And at the other extreme, we can have those from the social construction school like John Laws’ definition of “a family of methods for associating and channelling other entities and forces, both human and non-human”. Which when you think about it, does capture pretty much everything relating to technology, but it’s also so broad that it loses a lot of the utility.

But I’ve always drawn a middle line and my personal definition of technology is it’s “the material embodiment of an artifact and its associated systems, materials, and practices employed to achieve human ends”. I think we need to capture both the tool and the context, as well as the ways that they’re employed and used, and I think this definition captures the generative tools that we call AI as well. If we can recognize that they’re tools used for human ends and not actors with their own agency, then we can change the frame of the discussion around these generative tools and focus on what ends they’re being used for.

And what they’re being used for right now is not some science fictional version, either the dystopic hellscapes of the Matrix or Terminator, or on the flip side, the more utopic versions, the one, the “Fully Automated Luxury Communism” that we’d see in the post scarcity societies of like a Star Trek: The Next Generation, or even Iain M. Banks’ the Culture series.  Neither of these is coming true, but those polls – that ideation, these science fiction versions that kind of drive our collective imagination of the publics, the social imaginaries that we talked about a few episodes ago – these polls represent the two ends of that continuum, of that discussion, that dialectic between utopic and dystopic and the way we frame technology.

As Anabel Quan-Haase notes in their book on Technology and Society, those poles: the utopic idea of technology achieving progress through science, and the dystopic is technology as a threat to established ways of life, are both frames of reference. They could both be true depending on the point of view of the referrer. But as we said, it is a dialectic. There is a dialogue going back and forth between these two poles continually. So technology in this case is not inherently utopic or dystopic, but we have to return again to the ends that the technology is put towards: the human ends. So rather than utopic or dystopic, we can think of technology being rather emancipatory or controlling, and it’s in this frame, through this lens, this glass that I want to peer at the technology of AI.

The emancipatory frame for viewing these generative AI tools view them as an assistive technology, and it’s through this frame, this lens that we’re going to look at the technology first. These tools are exactly that: they are liberating, they are freeing. And whenever we want to take an empathetic view of technology, we wanna see how they may be used by others who aren’t in our situation.  And that situations means they may be doing okay, they might be even well off, but they may also be struggling. There may be issues that they, or challenges that they have to deal with on a regular basis that most of us can’t even imagine. And this is where my own experience comes from. So I’ll speak to that briefly.

Part of my background is when I was doing my field work for my dissertation, I was engaged with a number of the makerspaces in my city, and some of them were working with local need-knowers or persons with disabilities like the Tikkun Olam Makers, as well as the Makers Making Change groups. These groups worked with persons with disabilities to find solutions to their particular problems.  problems that often there wasn’t a market solution available because it wasn’t cost effective. You know, the “Capitalist realism” situation that we currently are under means that a lot of needs, especially for marginal groups, may go unmet. And these groups came together to try and meet those needs as best they could through technological solutions using modern technologies like 3D printing or microcontrollers or what have you, and they do it through regular events, whether it was a hackathon or regular monthly meetup groups or using the space provided by a local makerspace. And in all these cases, what these tools are are liberating to some of the constraints or challenges that are experienced in daily life.

We can think of more mainstream versions, like a mobility scooter that allows somebody with reduced mobility to get around and more fully participate within their community to meet some of the needs that they need on a regular basis, and even something as simple as that can be really liberating for somebody who needs it. We need to be cognizant of that because as the saying goes, we are all at best just temporarily able, and we never know when a change may be coming that could radically change our lives. So that empathetic view of technology allows us to think with some forethought about what may happen as if we or someone we love were in that situation, and it doesn’t even have to be somebody that close to us. We can have a much more communal or collective view of society as well.

But to return to this liberating view, we can think about it in terms of those tools, the generative tools, whether they’re for text or for art, or for programming, or even helping with a little bit of math.  We can see how they can assist us in our daily lives by either fulfilling needs or just allowing us to pursue opportunities that we thought were too daunting. While the generative art tools like Dall-E and Midjourney have been trained on already existing images and photographs, they allow creators to use them in new and novel ways.

It may be that a musician can use the tools to create a music video where before they never had the resources or time or money in any way, shape, or form to actually pursue that. It allows them to expand their art in a different realm. Similarly, an artist may be able to create stills that go with a collection or you know, accompany their writing that they’re working on, or an academic could use it as slides to accompany a presentation, something that they’ve spent time on, or a YouTube video, or even a podcast and their title bars and the like (present company included). My own personal experience when I was trying to launch this podcast was there was all this stuff I needed to do, and the generative art tools, the cruder ones that were available at that time, allowed some of the art assets to be filled in, and that barrier to launch, that barrier to getting something going was removed.

So emancipatory, liberating, even though at a much smaller scale, those barriers were removed and it allowed for creativity to flow in other areas, and it works similarly across these generative tools, whether it’s putting together some background art or a campaign map or a story prompt. If you need some background for a characters that are part of a story as an NPC, as a Dungeon Master, or what have you, or even just to bounce or refine coding ideas off of, I mean, the coding skills are rudimentary, but it does allow for something functional to be produced.

And this leads into some of the examples I’d like to talk about. The first one is from a post by Brennan Stelli on Mastodon on August 18th, where he said that we could leverage AI to do work, which is not being done already because there’s no budget time or knowhow.  There’s a lot of work that falls into this space of stuff that needs to be done, but you know, is outside of scope of a particular project. This could include something like developing the visualizations that will allow him to better communicate an idea in a fraction of the time, you know, minutes instead of hours that would normally take to do something like that, and so we can see in Brennan’s experience that it mirrors a lot of our own.

The next example’s a little bit more involved in an article written by Pam Belluck and published on the New York Times website on August 23rd, 2023. She details how researchers have used predictive text as well as AI generated facial animations that help with an avatar and speech that assist the stroke victim in communicating with their loved ones.

And the third example that hit a little bit closer to home was that of a Stanford research team that used the BCI or brain computer interface, along with AI assisted predictive text generation to allow a person with amyotrophic lateral sclerosis or ALS (to talk) at a regular conversational tempo, the tools read the neural activity that would be combined with the facial muscles moving and that is allowed to be translated into text. These are absolutely groundbreaking and amazing developments and I can’t think of any better example that shows how AI can be an assistive technology.

Now most of these technologies are confined to text and screen, to video and audio, and often when we think of AI, we think of mobility as well. So the robotic assistants that have come out of research labs like that of Boston Dynamics have attracted a lot of the attention, but even there, we can see some of the potential as an assistive technology. The fact that it’s confined to a humanoid robot means we sometimes lose sight of that fact, but that is what it is. In the video that they released in January of 2023, it shows an Atlas Robot as an assistant on a construction site providing tools and moving things around in aid of the human that’s the lead on the project, so it allows a single contractor working on their own to extend what they’re able to do, even if they don’t have access to a human helper. So it still counts as an assistive technology, even though we can start to see the dark side of the reflection through this particular lens, that the fact that an emancipatory technology may mean emancipation from the work that people currently have available to them.

In all of these instances, there’s the potential for job loss, that the tools would take the place of someone doing that, whether it’s in writing or as an artist, or a translator, or transcriber, or a construction assistant, and those are very real concerns. I do not want to downplay that, Part of our reflection on AI has to take these into account that the dark side of the mirror (or the flip side of the magnifying glass) can take something that can be helpful and exacerbate it when it’s applied to society at large. The concerns about job loss are similar to concerns we’ve had about automation for centuries, and they’re still valid. What we’re seeing is an extension of that automation into realms that we thought were previously exclusively bound to, you know, human actors: creators, artists, writers and the like.

This is why AI and generative art tools are such a driving and divisive element when it comes to the current WGA and SAG-Aftra strikes: that the future of Hollywood could be radically different if they see widespread usage. And beyond just the automation and potential job loss, a second area of concern is the one that ChatGPT and the large language models don’t necessarily have any element of truth involved in it, that they’re just producing output linguists like Professor Emily Bender of the University of Washington and the Mystery AI Hype Theater Podcast have gone into extensive detail about how the output of ChatGPT cannot be trusted. It has no linkage to truth, and there’s been other scholars that have gone into the challenges with using ChatGPT or LLMs for legal research or academic work or anything along those lines. I think it still has a lot of potential and utility as a tool, but it’s very much a contested space.

And the final area of contestation that we’ll talk about today is the question of control. Now, that question has two sides: the first is the control of that AI. One that most often surfaces in our collective imaginary is that idea of rogue super intelligences or killer robots gets repeated in TV, film, and our media in general, and this does get addressed at an academic level and works like Stuart Russell’s Human Compatible and Nick Bostrom’s Superintelligence.  They both address the idea of what happens if those artificial intelligences get beyond human capacity to control them.

But the other side of that is the control of us, control of society. Now, that gets replicated in our media as well, and everything from Westworld, to the underlying themes of the TV series Person of Interest, where The Machine is a computer system, developed to help detect and anticipate and suppress terrorist action using the tools of a post 9-11 surveillance state that it has access to.

And ever since Gilles Deleuze wrote his Postscript on the Societies of Control back in 1990, that so accurately captured the shift that had occurred in our societies from the sovereign societies of the Middle Ages and Renaissance through to the disciplinary societies that typified the 18th and 19th century, through to the shift that occurred in the 20th and 21st century to that of a control society where the logics of the society was enforced and regulated by computers and code. And while Deleuze was not talking about algorithms and AI in his work, we can see how they’re a natural extension of what he was talking about, how the biases that are ingrained within our algorithms, what Virginia Eubanks talked about in her book Automating Inequality, and how our biases and assumptions that go into the coding and training of those advanced system can manifest in ways, including everything from facial recognition to policing, to recommendation engines on travel websites that suggest that perhaps should go to the food bank to catch a meal.

Now there’s a twist to our Ottawa food bank story, of course. About a week after Microsoft came out and said that the article had been removed and that it had been identified that the issue was due to human error and not due to an unsupervised AI. But even with that answer, there are those who are skeptical: because it didn’t happen just once. There was a lot of articles where such weird or incongruous elements showed up. And of course, this being the internet, there was a number of people that did catch the receipts.

Now there’s a host of reasons of what might be happening with these bad reviews. Some plausible and some slightly less so. It could be just an issue of garbage in garbage out that the content that they’re scraping to power the AI is drawing articles that already exist that are, you know, satire or meme sites. If the information that you’re getting on the web is coming from Something Awful or 4chan, then you’re gonna get some dark articles in there. But the other alternative is that it could be just hallucinations that have been an observed fact that has been happening with these AIs and large language models that, uh, incidents like we saw with the Loa B that we talked about in an icebreaker last year are still coming forward in ways that are completely unexpected and out of our control.

That scares us a little bit because we don’t know exactly what it’s going to do. When we look at the AI through that lens, like in the mirror, what it’s reflecting back to us is something we don’t necessarily want to look at, and we think that it could be revealing the darkest aspects of ourselves, and that frightens us a whole lot.

AI is a reflection of our society and ourselves, and if we don’t like what we’re seeing, then that gives us an opportunity to perhaps correct things because AI, truth be told, is really dumb right now. It just shows us what’s gone into building it. But as it gets better, as the algorithms improve, then it may get better at hiding its sources.

And that’s a cause for concern. We’re rapidly reaching a point where we may no longer be able to tell or spot a deepfake or artificially generated image or voice, and this may be used by all manner of malicious actors. So as we look through our lens at the future of AI , what do we see on our horizon?

References:
Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.

Deleuze, G. (1992). Postscript on the Societies of Control. October, 59, 3–7.

Eubanks, V. (2018). Automating Inequality. Macmillan.

McLuhan, M. (1964). Understanding Media: The Extensions of Man. The New American Library.

Quan-Haase, A. (2015). Technology and Society: Social Networks, Power, and Inequality. Oxford University Press.

Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking.

Links:
https://arstechnica.com/information-technology/2023/08/microsoft-ai-suggests-food-bank-as-a-cannot-miss-tourist-spot-in-canada/

https://tomglobal.org/about

https://www.makersmakingchange.com/s/

https://arstechnica.com/health/2023/08/ai-powered-brain-implants-help-paralyzed-patients-communicate-faster-than-ever/

https://blog.jim-nielsen.com/2023/temporarily-abled/

https://www.businessinsider.com/microsoft-removes-embarrassing-offensive-ai-assisted-travel-articles-2023-8