Re: Spamming the Data Space – CLIP, GPT and synthetic data

[Luke Munn]

> I question whether the notion of pastiche makes any sense at all without
> interpretation (preumably Luke has something to say about that). What's
> certain is that the autonomy question is becoming urgent.
>
>
If an AI model produces a tasteless, derivative image in a forest and
there's no human there, is it pastiche? ;-)

Sure, you're right, pastiche is everywhere, and is very much in the eye of
the beholder. Even without direct AI production of images, a lot of the
shows on Netflix, for example, feel strangely familiar, precisely because
they cut-and-paste elements of 'cult' or 'successful' shows together.
'Stranger Things' was essentially engineered from viewer data, and Netflix
knew it was going to be successful even before it launched.

And this kinda brings me back to some of the genuine questions I have
around AI models and cultural critique - the lines here seem really
arbitrary. Why is my DALL-E generated image 'derivative' or a 'tainted'
image, according to the tech commentators I mentioned earlier, and my
'manual' pixel art not? I honestly don't see what the difference is between
a MidJourney collage and a Photoshop collage. The same question goes for
text. Why is Hardy Boys #44 (or whatever formulaic fiction you like) a
'real' book and the same story cranked out by GPT-3 'contaminated' or
'synthetic'?

Molly, in her excellent post, raises the issue of 'false production'. But I
genuinely don't know what that would look like. Is it false because it's
based on existing images (but collage artists already do this), or because
there's not enough 'intentionality' (prompting feels too easy?), or because
it's generated too quickly? In some ways, the exact same critique of AI
production could be levelled at meme creation - choose your base image, add
the classic meme typeface, and pump out a million permutations - but these
images are somehow considered an 'organic' part of the internet, while what
comes next is going to be artificial, mass-produced, spam, synthetic, and
so on.

At the core of all this, I think, is the instinct that there's something
unique about 'human' cultural production. (Even though AI models are
absolutely based on human labor, designs, illustrations, books, etc - their
packaging and obfuscation of this data makes them 'machinic' in the popular
consciousness.) Terms like 'meaning', or 'intention', or 'autonomy' gesture
to this desire, this hunch that something will be lost, that some ground
will be ceded with the move to AI image models, large language models, and
so on. I'm sympathetic to this - and don't want to come across as an
apologist for Big Tech, Open AI, etc. But I guess my struggle is to put my
finger on what that 'special something' is. Many of these posts have
suggested future autonomous zones where 'synthetic' culture is banned. What
would be the hallmark or signature of these spaces? No digital tools or
algorithmic media may come to mind, but these overlook the most crucial
element to 'new' cultural production: reading or listening or viewing other
people's works.

- 'Perplexed in Mianjin/Brisbane'



>
>
>
> On Fri, Dec 23, 2022 at 8:54 AM Francis Hunger <
> francis.hun...@irmielin.org> wrote:
>
>> Dear Luke, dear All
>>
>> Interesting essay Francis, and always appreciate Brian's thoughtful
>> comments. I think the historical angle Brian is pointing towards is
>> important as a way to push against the claims of AI models as somehow
>> entirely new or revolutionary.
>>
>> In particular, I want to push back against this idea that this is the
>> last 'pure' cultural snapshot available to AI models, that future
>> harvesting will be 'tainted' by automated content.
>>
>> At no point did I allude to the 'pureness' of a cultural snapshot, as you
>> suggest. Why should I? I was discussing this from a material perspective,
>> where data for training diffusion models becomes the statistical material
>> to inform these models. This data has never been 'pure'. I used the
>> distinction of uncontaminated/contaminated to show the difference between a
>> training process for machine learning which builds on an snapshot, that is
>> still uncontaminated by the outputs of CLIP or GPT and one which includes
>> generated text and images using this techique on a large scale.
>>
>> It is obvious, but maybe I should have made it more clear, that the
>> training data in itself is already far from pure. Honestly I'm a bit
>> shocked, you would suggest I'd come up with a nostalgic argument about
>> purity.
>>
>> Francis' examples of hip hop and dnb culture, with sampling at their
>> heart, already starts to point to the problems with this statement. Culture
>> has always been a project of cutting and splicing, appropriating,
>> transforming, and remaking existing material. It's funny that AI
>> commentators like Gary Marcus talk about GPT-3 as the 'king of pastiche'.
>> Pastiche is what culture does. Indeed, we have whole genres (the romance
>> novel, the murder mystery, etc) that are ab

A.I. Lenin: What is to be Done Today, by ChatGPT and Dmytri Kleiner

[Dmytri Kleiner]

A.I. Lenin

What is to be Done Today, by GPT and Dmytri Kleiner

In the early 20th century, Lenin recognized the importance of an 
all-Russian newspaper as a means of building the capacities and 
capabilities necessary for a revolutionary movement to overthrow 
capitalism. Similarly, a digital agency can help build the capacities 
and capabilities necessary for a revolutionary movement to overthrow 
capitalism in the modern world. As Lenin wrote in "What Is To Be Done?": 
"Without a newspaper, it is impossible to unite, to direct, to arouse, 
and to organize the masses" (Lenin, 1902, p. 28). For Lenin, it was not 
the newspaper as such that was important, but the organization capable 
of publishing it. In this sense, a digital agency is not an end in 
itself, but a means of building the abilities necessary for a 
revolutionary movement to coordinate and communicate with supporters, as 
well as carry out propaganda and agitation efforts.


One key capacity that a digital agency can help build is the ability to 
disseminate information and propaganda effectively. Social media 
platforms and other modern technologies have become major channels for 
the distribution of information and ideas, and a digital agency can help 
a revolutionary movement leverage these channels to reach a wide 
audience and spread its message. In the past, Leninist organizations 
used underground newspapers and smuggled literature to disseminate their 
message, often at great risk to their own safety. Today, a digital 
agency can help a revolutionary movement use modern technologies to 
reach a wider audience and spread its message more effectively. As Lenin 
wrote in "The Tasks of the Russian Social Democrats": "The spread of 
revolutionary ideas among the masses depends above all on the degree of 
their own organization" (Lenin, 1898, p. 42). And as he wrote in 
"Left-Wing Communism: An Infantile Disorder": "The party must have its 
own press, its own organization, and its own set of tactics" (Lenin, 
1920, p. 22).


Another key capacity that a digital agency can help build is the ability 
to mobilize and organize supporters. In the past, Leninist organizations 
relied on secret meetings and underground networks, known as 
konspiratsiya, to coordinate their activities. Today, social media and 
other modern technologies can provide similar capabilities, allowing a 
revolutionary movement to quickly and effectively organize and mobilize 
supporters. For example, a Leninist organization might use a closed 
social media group or a secure messaging app to organize meetings, 
distribute propaganda, and coordinate actions. As Lenin wrote in "One 
Step Forward, Two Steps Back": "Only konspiratsiya can ensure the 
freedom and independence of the Party" (Lenin, 1904, p. 80).


Konspiratsiya refers to the practice of maintaining secrecy and keeping 
activities hidden from the authorities in order to avoid detection and 
repression. In the past, Leninist organizations used a variety of 
tactics to maintain secrecy and avoid detection, such as using code 
words and symbols, holding meetings in secret locations, and using fake 
names. In the context of a digital agency, this might involve using 
aliases and pseudonyms, as well as secure communication methods and 
infrastructure, to protect against surveillance and to maintain secrecy. 
As Lenin wrote in "Left-Wing Communism: An Infantile Disorder": "The 
party must have its own press, its own organization, and its own set of 
tactics" (Lenin, 1920, p. 22).


A digital agency can also help build the capacity for a revolutionary 
movement to engage in digital activism and direct action. This might 
involve creating and distributing memes and other digital content that 
can go viral and spread the movement's message, or using hacking and 
other digital tactics to disrupt the operations of capitalist 
institutions. In the past, Leninist organizations used a variety of 
tactics to engage in direct action, such as strikes, boycotts, and 
sabotage, as well as organizing demonstrations and protests. In the 
context of a digital agency, these tactics might be adapted for the 
digital realm, such as organizing online boycotts or launching cyber 
attacks against capitalist institutions.


Another key capacity that a digital agency can help build is the ability 
to defend against digital threats and attacks. This might involve 
developing and implementing security measures to protect against cyber 
threats, such as hackers and malware, as well as implementing measures 
to protect against surveillance and monitoring by the authorities. In 
the past, Leninist organizations often faced repression and persecution 
from the state, and they had to develop tactics and strategies to defend 
against these threats. In the context of a digital agency, this might 
involve using encryption and other secure communication methods, as well 
as developing contingency plans and backup systems to ensure the 
continued operati

Re: Spamming the Data Space – CLIP, GPT and synthetic data

[Molly Hankwitz]

Thank you, Francis for this very interesting work, and for the responses from all. Some of you may not know me at all, so by way of introduction, I am an independent researcher/writer/feminist artist working between film, media and cultural studies on history and theory of new media in the context of post-war arts culture.  I have been teaching about AI, AL, and varied artists works in these fields for a few years with great internet as a new medium (AI) comes in. Some salient features of this analysis and the varied responses stand out for further critique. If I am correctly reading your text, Francis, and I thank you Francis for the separation of the two modi, on the one hand there is the potential for new forms of data sets (as assets and commodities) which are “walled” and controlled - as collections or arrays - which will undoubtedly become, indeed already are being made:  Some examples here: Lev Manovich’s PhotoTime projects, Trevor Paglen’s experimental arrays in which he inputs the dataset and then works with it; or experiments with MS-made Genesis Maps software used on/with the METs digital collections. Some of you may have seen or been at this event written about here: https://www.metmuseum.org/perspectives/articles/2019/2/artificial-intelligence-machine-learning-art-authorshipFrancis writes of something similar - controlled data spaces (like online communities maybe, where contributing or donating is controlled by membership?) “We are going to create separate data ecologies, which prohibit spamming the data space. These would be spaces, comparable to the no-photo-policy in clubs like Berghain or IFZ with a no-synthetics policy. While vast areas of the information space may be indeed flooded, these would be valuable zones of cultural exchange.” On the other hand, there is the ability for AIs to remake data based on anything it is allowed to pull in, and to then feed that invention back into data pools and streams. So the question arises about who and how data will be controlled. This question is familiar. Capitalist interests will have their reasons and methods - maybe what Brian alluded to about understanding that we have seen these mechanisms at work before…and there is the decided problem of having capitalist interests further empowered through the accelerated process of reproduction and distribution of it’s wares, especially when this uncontrolled reproduction/replication is prone to prejudices, biases, falsehood, and omissions. Let’s see what we can foresee. Can we foresee cultural progress continuing as a growing alienation among larger sectors of global populations, no longer able to find any useful or rational meaning in a plethora of false production? (- a great resignation from the virtual spectacle and an end to capitalism of this kind - literally it’s collapse from over extension) Will there be regulation of AI in the net, further regulating the circulation of information? Will  those with the least access, those most vulnerable to data mining of their lives, become even more susceptible to exploitation, as they fail to understand how to navigate this new synthetic reality to their own advantage or where self-interest is a manufactured byproduct of the spectacle? Whatever we might foresee, the bandaid of tweaking SEO for the sake of a false surface of BIPOC to add in is lame (Francis’ example about a few more pictures of BIPOC being worked in) Molly On Dec 23, 2022, at 12:48 PM, Luke Munn  wrote:Hey Francis,Thanks for your response. Just briefly I think there was some misframing of my response. It was not meant as a takedown or critique of your text, but rather a more general response to this idea tabled in the opening line, For the last time in human history the cultural-data space has not been contaminated. This is an idea I had seen a few times from others the same day. 'The internet is now forever contaminated with images made by AI,” Mike Cook, an AI researcher at King’s College London' . 'How AI-Generated Text Is Poisoning the Internet' etc. I was kicking around this idea and seeing what it implied, hence delving into this idea of post-truth authenticity etc which your text doesn't talk about. I know you get this stuff, and can see issues with some of these assumptions, which is precisely why I mentioned your dnb example as culture of remixing. Nevertheless I did think your original piece (and most people's work including mine) can be strengthened by thinking longer term / historically, as the presentist framing of tech is often very powerful. That was my only real feedback - but i can see how this more general 'essay as jumping off point' discussion could be misread, so apologies for that. -LOn Sat, 24 Dec 2022, 01:03 Francis Hunger,  wrote:
  

  
  
Dear Luke, dear All

  
  Interesting essay Francis, and always appreciate
Brian's thoughtful comments. I think the historical angle Brian
is pointing towards 

Re: Spamming the Data Space – CLIP, GPT and synthetic data

[Luke Munn]

Hey Francis,

Thanks for your response. Just briefly I think there was some misframing of
my response. It was not meant as a takedown or critique of your text, but
rather a more general response to this idea tabled in the opening line,

For the last time in human history the cultural-data space has not been
contaminated.

This is an idea I had seen a few times from others the same day. 'The
internet is now forever contaminated with images made by AI,” Mike Cook, an
AI researcher at King’s College London' . 'How AI-Generated Text Is
Poisoning the Internet' etc.

I was kicking around this idea and seeing what it implied, hence delving
into this idea of post-truth authenticity etc which your text doesn't talk
about.

I know you get this stuff, and can see issues with some of these
assumptions, which is precisely why I mentioned your dnb example as culture
of remixing.

Nevertheless I did think your original piece (and most people's work
including mine) can be strengthened by thinking longer term / historically,
as the presentist framing of tech is often very powerful. That was my only
real feedback - but i can see how this more general 'essay as jumping off
point' discussion could be misread, so apologies for that.

-L

On Sat, 24 Dec 2022, 01:03 Francis Hunger, 
wrote:

> Dear Luke, dear All
>
> Interesting essay Francis, and always appreciate Brian's thoughtful
> comments. I think the historical angle Brian is pointing towards is
> important as a way to push against the claims of AI models as somehow
> entirely new or revolutionary.
>
> In particular, I want to push back against this idea that this is the last
> 'pure' cultural snapshot available to AI models, that future harvesting
> will be 'tainted' by automated content.
>
> At no point did I allude to the 'pureness' of a cultural snapshot, as you
> suggest. Why should I? I was discussing this from a material perspective,
> where data for training diffusion models becomes the statistical material
> to inform these models. This data has never been 'pure'. I used the
> distinction of uncontaminated/contaminated to show the difference between a
> training process for machine learning which builds on an snapshot, that is
> still uncontaminated by the outputs of CLIP or GPT and one which includes
> generated text and images using this techique on a large scale.
>
> It is obvious, but maybe I should have made it more clear, that the
> training data in itself is already far from pure. Honestly I'm a bit
> shocked, you would suggest I'd come up with a nostalgic argument about
> purity.
>
> Francis' examples of hip hop and dnb culture, with sampling at their
> heart, already starts to point to the problems with this statement. Culture
> has always been a project of cutting and splicing, appropriating,
> transforming, and remaking existing material. It's funny that AI
> commentators like Gary Marcus talk about GPT-3 as the 'king of pastiche'.
> Pastiche is what culture does. Indeed, we have whole genres (the romance
> novel, the murder mystery, etc) that are about reproducing certain elements
> in slightly different permutations, over and over again.
>
> Maybe it is no coincidence that I included exactly this example.
>
> Unspoken in this claim of machines 'tainting' or 'corrupting' culture is
> the idea of authenticity.
>
> I didn't claim 'tainting' or 'corrupting' culture, not even unspoken. Who
> am I to argue against the productive forces?
>
> It really reminds me of the moral panic surrounding algorithmic news and
> platform-driven disinformation, where pundits lamented the shift from truth
> to 'post-truth.'  This is not to suggest that misinformation is not an
> issue, nor that veracity doesn't matter (i.e. Rohingya and Facebook). But
> the premise of some halcyon age of truth prior to the digital needs to get
> wrecked.
>
> I agree. Only, I never equaled 'uncontaminated' to a "truth prior to the
> digital", I equaled it to a snapshot that doesn't contain material created
> by transformer models.
>
> Yes, Large language models and other AI technologies do introduce new
> conditions, generating truth claims rapidly and at scale. But rather than
> hand-wringing about 'fake news,' it's more productive to see how they
> splice together several truth theories (coherence, consensus, social
> construction, etc) into new formations.
>
> I was more interested in two points:
>
> 1.) Subversion: What I called in my original text the 'data space'
> (created through cultural snapshots as suggested by Eva Cetinic) is an
> already biased, largely uncurated information space where image data and
> language data are scaped and then mathemtically-statistically merged
> together. The focus point here is the sheer scale on which this happens.
> GPT-3 and CLIP are techniques that both build on massive datascraping
> (compared for instance to GANs) so that it is only possible for well funded
> organizations such as Open-AI or LAION to build these datasets. This
> dataspace c

Re: Spamming the Data Space – CLIP, GPT and synthetic data

[Brian Holmes]

[This was written yesterday, so it responds mostly to Luke and Felix.]

I agree that pastiche is a fundamental cultural process - but if it's so
fundamental, then to make any distinctions you have to look at its effects
in specific contexts. One such context, in the recent past, is
postmodernism. It's relevant in some ways, but I agree with Francis that
the present context is quite different.

Postmodern pastiche is the original twist that an individual gives to a
mass-distribution image. In the arts of the 1980s and 90s, the pastiche
aesthetic had the effect of disqualifying a whole range of avant-garde
practices, from neo-dadist transgression to modernist abstraction, all of
which consciously tried to mark off a space *outside* corporate-capitalist
aesthetic production. From one angle, the acceptance of a common commercial
culture was a good thing: it reduced the power of elite gatekeepers, since
the raw material of art was now ready-to-hand, without racial, financial
and educational barriers to access. But the quest for autonomy is another
fundamental cultural process, and in contemporary societies, autonomy from
highly manipulative aesthetic production is crucial. Otherwise, there's
nowhere to develop any divergent ethical/political orientation. As the
focus of commercial culture shifted online, these problems took on new
guises. Most of my own work as a cultural critic in the 2000s was devoted
to autonomy in the communication societies - and then came social media,
making the whole situation dramatically worse.

Today, Francis points to the floods of imagery that are already being
produced by AI/statistical computing, and he predicts second and third
generations of degraded images, synthesized from the initial ones. I was
struck by this word "degraded" in the initial text, and I think it
corresponds to something more than simple entropy on the level of data. The
absence of any individual or subcultural viewpoint at the origin of the
statistically generated images, and the coresponding lack of particular
affects, aspirations, insights or blindspots, renders yet another
fundamental cultural process obsolete - namely, interpretation. I question
whether the notion of pastiche makes any sense at all without
interpretation (preumably Luke has something to say about that). What's
certain is that the autonomy question is becoming urgent.

Autonomy is not about purity, nor self-sufficiency, nor withdrawal. It's
about the ability to establish the terms (particularly the overarching
value orientation) that will guide one's engagement with society. I agree
with Francis that being able to filter out statistically produced images
(and music, and discourse) is going to become a major issue under flood
conditions. And I'd go further. Whoever is not able to form or join an
interpretative community, very consciously dedicated to making meaning with
respect to art or other cultural practices, is going to experience a very
profound alienation during the next phase of the communication societies.



On Fri, Dec 23, 2022 at 8:54 AM Francis Hunger 
wrote:

> Dear Luke, dear All
>
> Interesting essay Francis, and always appreciate Brian's thoughtful
> comments. I think the historical angle Brian is pointing towards is
> important as a way to push against the claims of AI models as somehow
> entirely new or revolutionary.
>
> In particular, I want to push back against this idea that this is the last
> 'pure' cultural snapshot available to AI models, that future harvesting
> will be 'tainted' by automated content.
>
> At no point did I allude to the 'pureness' of a cultural snapshot, as you
> suggest. Why should I? I was discussing this from a material perspective,
> where data for training diffusion models becomes the statistical material
> to inform these models. This data has never been 'pure'. I used the
> distinction of uncontaminated/contaminated to show the difference between a
> training process for machine learning which builds on an snapshot, that is
> still uncontaminated by the outputs of CLIP or GPT and one which includes
> generated text and images using this techique on a large scale.
>
> It is obvious, but maybe I should have made it more clear, that the
> training data in itself is already far from pure. Honestly I'm a bit
> shocked, you would suggest I'd come up with a nostalgic argument about
> purity.
>
> Francis' examples of hip hop and dnb culture, with sampling at their
> heart, already starts to point to the problems with this statement. Culture
> has always been a project of cutting and splicing, appropriating,
> transforming, and remaking existing material. It's funny that AI
> commentators like Gary Marcus talk about GPT-3 as the 'king of pastiche'.
> Pastiche is what culture does. Indeed, we have whole genres (the romance
> novel, the murder mystery, etc) that are about reproducing certain elements
> in slightly different permutations, over and over again.
>
> Maybe it is no coincidence that I inclu

Re: Spamming the Data Space – CLIP, GPT and synthetic data

[Francis Hunger]

Dear Luke, dear All
Interesting essay Francis, and always appreciate Brian's thoughtful 
comments. I think the historical angle Brian is pointing towards is 
important as a way to push against the claims of AI models as somehow 
entirely new or revolutionary.


In particular, I want to push back against this idea that this is the 
last 'pure' cultural snapshot available to AI models, that future 
harvesting will be 'tainted' by automated content.


At no point did I allude to the 'pureness' of a cultural snapshot, as 
you suggest. Why should I? I was discussing this from a material 
perspective, where data for training diffusion models becomes the 
statistical material to inform these models. This data has never been 
'pure'. I used the distinction of uncontaminated/contaminated to show 
the difference between a training process for machine learning which 
builds on an snapshot, that is still uncontaminated by the outputs of 
CLIP or GPT and one which includes generated text and images using this 
techique on a large scale.


It is obvious, but maybe I should have made it more clear, that the 
training data in itself is already far from pure. Honestly I'm a bit 
shocked, you would suggest I'd come up with a nostalgic argument about 
purity.


Francis' examples of hip hop and dnb culture, with sampling at their 
heart, already starts to point to the problems with this statement. 
Culture has always been a project of cutting and splicing, 
appropriating, transforming, and remaking existing material. It's 
funny that AI commentators like Gary Marcus talk about GPT-3 as the 
'king of pastiche'. Pastiche is what culture does. Indeed, we have 
whole genres (the romance novel, the murder mystery, etc) that are 
about reproducing certain elements in slightly different permutations, 
over and over again.

Maybe it is no coincidence that I included exactly this example.
Unspoken in this claim of machines 'tainting' or 'corrupting' culture 
is the idea of authenticity.
I didn't claim 'tainting' or 'corrupting' culture, not even unspoken. 
Who am I to argue against the productive forces?
It really reminds me of the moral panic surrounding algorithmic news 
and platform-driven disinformation, where pundits lamented the shift 
from truth to 'post-truth.'  This is not to suggest that 
misinformation is not an issue, nor that veracity doesn't matter (i.e. 
Rohingya and Facebook). But the premise of some halcyon age of truth 
prior to the digital needs to get wrecked.
I agree. Only, I never equaled 'uncontaminated' to a "truth prior to the 
digital", I equaled it to a snapshot that doesn't contain material 
created by transformer models.
Yes, Large language models and other AI technologies do introduce new 
conditions, generating truth claims rapidly and at scale. But rather 
than hand-wringing about 'fake news,' it's more productive to see how 
they splice together several truth theories (coherence, consensus, 
social construction, etc) into new formations.


I was more interested in two points:

1.) Subversion: What I called in my original text the 'data space' 
(created through cultural snapshots as suggested by Eva Cetinic) is an 
already biased, largely uncurated information space where image data and 
language data are scaped and then mathemtically-statistically merged 
together. The focus point here is the sheer scale on which this happens. 
GPT-3 and CLIP are techniques that both build on massive datascraping 
(compared for instance to GANs) so that it is only possible for well 
funded organizations such as Open-AI or LAION to build these datasets. 
This dataspace could be spammed a) if you want to subvert it and b) if 
you'd want to advertise. The spam would need to be on a large scale in 
order to influence the next (contaminated) iteration of a cultural 
snapshot. In that sense only I used the un/contaminated distinction.


2). In response to Brian I evoked a scenario that builds on what we 
already experience when it comes to information spamming. We all know, 
that mis-information is a social and _not_ a machinic function. Maybe I 
should have made this more clear (I simply assumed it). I ignored Brians 
comment on the decline of culture, whatever this would mean, and could 
have been more precise in this regards. I don't assume culture declines. 
Beyond this, there have been discussions about deepfakes for instance 
and we saw that deepfakes are not needed at all to create 
mis-information, when one can just cut any video using standard video 
editing practices towards 'make-believe'. I wasn't 'hand-wringing' about 
fake news, in my comment to Brian, instead I was quoting Langlois with 
the concept of 'real fakes'.
Further I'm suggesting that CLIP and GPT make it more easy to automate 
large scale spamming, making online communities uninhabitable or 
moderation more difficult. Maybe I'm overestimating the effect. We can 
already observe GPT-3 automated comments appearing on twitter or the ban 
of GPTChat pos