In this essay, published for The Nation magazine, I argue that the primarily purpose of the application of generative methods to the production of images (what some insist upon calling AI art) is the creation of an assembly line.
Art without artists, produced via automation.
Several years ago, there was a mini-trend of soft documentaries depicting what would happen to the built environment if humans somehow disappeared from the Earth. How long, for example, would untended skyscrapers punch against the sky before they collapsed in spectacular, downward cascading showers of steel and glass onto abandoned streets? These are the sorts of questions posed in these films.
As I watched these soothing depictions of a quieter world, I sometimes imagined a massive orbital tombstone, perhaps launched by the final rocket engineers, onto which was etched: Wasted Potential.
While I type these words, billions of dollars have been spent on and barely tabulated amounts of electrical power, water and human labor (barely tabulated, because deliberately obscured) have been devoted to large language model (LLM) systems such as ChatGPT. If you follow the AI critical space you’re familiar with the many problems produced by the use and promotion of these systems – including, on the hype end, the most recent gyration, a declaration of “existential risk” by a collection of tech luminaries (a category which, in a Venn diagram, overlaps with carnival barker). This use of mountains of resources to enhance the profit objectives of Microsoft, Amazon and Google, among other firms not occupying their olympian perches, is wasted potential in frenetic action.
But what of alternative visions? They exist, all is not despair. The dangerous nonsense relentlessly spewing from the AI industry is overwhelming and countering it is a full time pursuit. But we can’t stay stuck, as if in amber, in a state of debunking and critique. There must be more. I recommend the DAIR Institute and Logic(s) magazine as starting points for exploring other ways of thinking about applied computation. Ideologically, AI doomerism is fueled in large measure by dystopian pop sci-fi such as Terminator. You know the story, which is a tale as old as the age of digital computers: a malevolent supercomputer – Skynet (a name that sounds like a product) – launches, for some reason, a war on humanity, resulting in near extinction. The tech industry seems to love ripping dystopian yarns. Judging by the now almost completely forgotten metaverse push (a year ago, almost as distant as the pleistocene in hype cycle time), inspired by the less than sunny sci-fi novel Snow Crash, we can even say that dystopian storylines are a part of business plans (what is the idea of sitting for hours wearing VR goggles if not darkly funny?).
There are also less terrible, even hopeful, fictional visions, presented via pop science fiction such as Star Trek´s Library Computer Access/Retrieval System – LCARS.
In the Star Trek: The Next Generation episode, “Booby Trap” the starship Enterprise is caught in a trap, composed of energy sapping fields, that prevents it from using its most powerful mode of propulsion, warp drive. The ship’s chief engineer, Geordi LeForge, is given the urgent task of finding a solution. LeForge realizes that escaping this trap requires a re-configuration, perhaps even a new understanding, of the ship’s propulsion system. That’s the plot but most intriguing to me is the way LeForge goes about trying to find a solution.
The engineer uses the ship’s computer – the LCARS system – to do a retrieval and rapid parsing of the text of research and engineering papers going back centuries. He interacts with the computer via a combination of audio and keyboard/monitor. Eventually, LeForge resorts to a synthetic, holo mockup of the designer of the ship’s engines, Dr. Leah Brahms, raising all manner of ethical issues but we needn’t bother with that plot element.
I’ve created a high level visualisation of how this fictional system is portrayed in the episode:
The ability to identify text via search, to summarize and read contents (with just enough contextual capability to be useful) and to output relevant results is rather close, conceptually, to the potential of language models. The difference between what we actually have – competing and discrete systems owned by corporations – and LCARS (besides many orders of magnitude of greater sophistication in the fictional system) is that LCARS is presented as an integrated, holistic and scoped system. LCARS’ design is to be a library that enables access to knowledge and retrieves results based on queried criteria.
There is a potential, latent within language models and hybrid systems – indeed, within almost the entire menagerie of machine learning methods – to create a unified computational model for a universally useful platform. This potential is being wasted, indeed, suppressed as oceans of capital, talent and hardware is poured into privately owned things such as ChatGPT. There are hints of this potential found within corporate spaces; Meta’s LLaMA, which leaked online, shows one avenue. There are surely others.
Among a dizzying collection of falsehoods, the tech industry’s greatest lie is that it is building the future. Or perhaps, I should sharpen my description: the industry may indeed be building the future but contrary to its claims, it is not a future with human needs centered. It is possible however, to imagine and build a different computation and we needn’t turn to Silicon Valley’s well thumbed library of dystopian novels to find it. Science fiction such as Star Trek (I’m sure there are others) provide more productive visions
What should we think about AI? To corporate boosters and their camp followers (an army of relentless shouters) , so-called artificial intelligence is a world altering technology, sweeping across the globe like a wave made from the plots of forgotten science fiction novels. Among critics, thoughts are more varied. Some focus on debunking hyped claims, others, on the industry’s racist conceptions (such as the presentation of a cohort of men, mostly White, who work with ‘code’ as being the pinnacle of human achievement) and still others, on the seldom examined ideology of ‘intelligence’ itself.
For Dan McQuillan, author of the taut (seven chapters) yet expansive book, ‘Resisting AI: An Anti-Facist Approach to Artificial Intelligence’ AI, is, under current conditions but not inherently, the computational manifestation of ever present fascist ideologies of control, categorization and exclusion. McQuillan has written a vital manifesto, the sort of work which, many years from now, may be recalled, if we’re fortunate, as being among the defining calls to arms of its age. In several interviews (including this one for Machine Learning Street Talk) McQuillan has described the book’s origin as a planned, scholarly review of the industry that, as its true state became clearer to him, evolved into a warning.
We can be glad he had the courage to follow the evidence where it led.
Both In and Of the World
“The greatest trick the Devil ever pulled” the saying goes, “was convincing the world he doesn’t exist.” The tech industry, our very own Mephistopheles (though lacking the expected fashion sense) has pulled a similar trick with ‘AI’ convincing us that, alone among technical methods, it exists as a force disconnected from the world’s socio-political concerns. In short order, McQuillan dispenses with this in the introduction:
It would be troubling enough if AI was a technology being tested in the lab or applied in a few pioneering startups, but it already has huge institutional and cultural momentum. […] AI derives a lot of its authority from its association with methods of scientific analysis, especially abstraction and reduction, an association which also fuels the hubris of some of its practitioners. The roll out of AI across swathes of industry doesn’t so much lead to a loss of jobs as to an amplification of casualized and precarious work. [emphasis mine] Rather than being an apocalyptic technology, AI is more aptly characterized as a form of supercharged bureaucracy that ramps up everyday cruelties, such as those in our systems of welfare. In general, […] AI doesn’t lead to a new dystopia ruled over by machines but an intensification of existing misery through speculative tendencies that echo those of finance capital. These tendencies are given a particular cutting edge by the way Al operates with and through race. AI is a form of computation that inherits concepts developed under colonialism and reproduces them as a form of race science. This is the payload of real AI under the status quo. [Introduction, pg 4]
Rather than acting as the bridge to an unprecedented new world, AI systems (really, statistical inference engines) are the perfect tool for the continuance of existing modes of control, intensified and excused by the cover of supposed silicon impartiality.
Later, in chapter two, titled, ‘AI Violence’ McQuillan sharpens his argument that the systems imposed on us are engines of automated abuse.
AI operationalizes [a] reductive view through its representations. […] , Aľ’s representations of the world consist of the set of weights in the [processing] layers plus the model architecture of the layers themselves. Like science, Al’s representations are presented as distinct from that which they claim to represent. In other words, there is assumed to be an underlying base reality that is independent of the practices by which such representations are constructed. But […] the entities represented by AI systems- the ‘careful Amazon driver’ or the ‘trustworthy citizen’- are partly constructed by the systems that represent them. AI needs to be understood not as an instrument of scientific measurement but as an apparatus that establishes ‘relations of becoming between subjects and representations. The subject co-emerges along with the representation. The society represented by AI is the one that it actively produces.
We are familiar with the categories McQuillan highlights such as ‘careful drivers’ from insurance and other industries and government agencies which use the tagging and statistical sorting of discrete attributes to manage people and their movements within narrow parameters. AI, as McQuillan repeatedly stresses, supercharges already existing methods and ways of thinking, embedded within system logic. We don’t get a future, we are trapped in a frozen present, in which new thinking and new arrangements are inhibited via the computational enforcement of past structures.
Necropolitics
For me, the most powerful diagnostic section of the book is chapter 4, ‘Necropolitics.’ Although McQuillan is careful to not declare AI systems fascist by nature (beginning the work of imagining other uses for computational infrastructure in Chapter 5, ‘Post Machinic Learning’) he does make the critical point that these systems, embedded within a fraying political economy, are being promoted and made inescapable at a moment of mounting danger:
Al is entangled with our systems of ordering society. […] It helps accelerate a shift towards far-right politics. AI is emerging from within a convolution of ongoing crises, each of which has the potential to be fascism-inducing, including austerity, COVID-19 and climate change. Alongside these there is an internal crisis in the ‘relations of oppression’, especially the general destabilization of White male supremacy by decolonial, feminist, LGBTQI and other social movements (Palheta, 2021). The enrollment of AI in the management of these various crises produces ‘states of exception’ – forms of exclusion that render people vulnerable in an absolute sense. The multiplication of algorithmic states of exception across carceral, social and healthcare systems makes visible the necropolitics of Al; that is, its role in deciding who should live and who should be allowed to die.
As 20th century Marxists were fond of saying, it is no accident that as the capitalist social order faces ever more significant challenges, ranging from demands from the multitudes subjected to its tyranny to the growing instability of nature itself as climate change’s impacts accelerate, there is a turn, by elites, to a technology of command and control to reassert some sense of order. McQuillan’s urgency is born of a recognition of global emergency and the ways the collection of computational methods called ‘AI’ is being marshalled to meet that emergency using what can clearly be identified as fascist approaches.
There’s much more to say but I will leave it here so you can explore on your own. Resisting AI: An Anti-Facist Approach to Artificial Intelligence, is an important and necessary book.
As the hype, indeed, propaganda about AI and its supposed benefits and even dangers (such as the delusions about ‘superintelligence’ a red herring) are broadcast ever more loudly, we need a collectivity of counterbalancing ideas and voices. McQuillan has provided us with a powerful contribution.
[In this post, I imagine that I’m writing to a researcher who, disappointed, and perhaps confused by the seemingly unstoppable corporate direction their field is taking, needs a bit of, well, not cheering up precisely but, something to help them understand what it all means and how to resist]
My friend,
Listen, I know you’ve been thrown by the way things have been going for the past few years – really, the past decade; a step by step privatization of the field you love and education pursued at significant financial cost (you’re not a trust funder) because of your desire to understand cognition and just maybe, build systems that, through their cognitive dexterity, aid humanity (vainglorious but, why not aim high?) You thought of people such as McCarthy, Weizenbaum, Minsky and Shannon and hoped to blaze trails, as they did.
When OpenAI hit the scene in 2015, with the promise – in its very name – to be an open home for advanced research, you celebrated. Over wine, we argued (that’s too strong, more like warmly debated with increasing heat as the wine flowed) about the participation of sinister figures such as Musk and Thiel. At the time, Musk was something of a hero to you and Thiel? Well, he was just a quirky VC with deep pockets and an overlooked penchant for ideas that are a bit Goebbels-esque. “Form follows function,” I said, “and the function of these people is to find ways to generate profit and pretend they’re gods.” But we let that drop over glasses of chardonnay.
Here we are, in 2023… which for you, or more pointedly your dreams, has become an annus horribilis, a horrible year. OpenAI is now married to Microsoft and the much anticipated release of GPT-4 is, in its operational and environmental impact details, shrouded in deliberate mystery. AI ethics teams are discarded like used tissues – there is an air of defeat as the idea of the field you thought you had joined dies the death of a thousand cuts.
Now is the time to look around and remember what I told you all those years ago: science and engineering (and your field contains both these things) do not exist outside of the world but are very much in it and are subject to a reality described by the phrase you’ve heard me say a million times: political economy. Our political economy – or, I should say, the political economy (the interrelations of law, production, custom and more) we’re subject to, is capitalist. What does this mean for your field?
It means that the marriage between OpenAI and Microsoft, the integration of large language models with the Azure cloud and the M365 SaaS platforms, the elimination of ethics teams whose work might challenge or impede marketing efforts, the reckless proliferation of algorithmically enacted harms is all because the real goal is profit, which is at the heart of capitalist political economy.
And we needn’t stop with Microsoft; there is no island to run to, no place that is outside of this political economy. No, not even if your team and leadership are quite lovely. This is a totalitarian (or, if you’re uncomfortable with that word, hegemonic) system which covers the globe in its harsh logic.
Oh but now you’re inclined to debate again and it’s too early for wine. I can hear you saying, ‘We can create an ethical AI; it’s possible. We can return to the research effort of years past’ I won’t say it’s impossible, stranger things have presumably happened in the winding history of humanity, but taking the whole fetid situation into account – yes, the relationship between access to computation and socio-technical power, the political economy, it’s not probable. So long as you continue believing in something that the structure of the society we live in does not support, you will continue to be disappointed.
Unless, that is, that structure is changed.
What is to be done?
I don’t expect you to become a Marxist (though it would be nice, we could compare obscure notes about historical materialism) but what I’m encouraging you to consider is that the world we grew up in and, quite naturally take for granted as immutable – the world of capitalist social relations, the world which, among other less than fragrant things, has all but completely absorbed your field into its profit engine is not the only way to organize human society.
Once you accept that, we can begin to talk about what might come next.
A few years ago, I started this very blog, devoted, as the subtitle reads, to “AI industry analysis without hype and techbro-ism.” Writing, when seriously pursued (whether money is exchanged or not) is a demanding activity, requiring time and often, a reduced number of social interactions, things that are becoming ever scarcer in our decaying world of enforced busy-ness and endlessly distracting ‘discourse.’
Considering the difficulty, why bother writing? And why bother writing about the tech industry generally, and its so-called ‘AI’ incarnation specifically? Until very (very) recently, the unchallenged cultural consensus was that Silicon Valley is populated by a wondrous horde of luminous creatures, the brilliant young who, armed only with wafer thin laptops, dreams, and that sorcerer’s wand, code, were building a vibrant future of robot taxis, chatbot friends and virtual worlds filled with business meetings attended by cartoon dinosaur avatars.
Who could resist this vision, this nirvana of convenience? Well, as it happens, yours truly.
It was while watching a left-leaning (and at the time, supposedly Marxist) YouTube show that I realized there was an acute need for a pitiless, materialist critique of the tech industry. One of the show’s co-hosts opined that it would not be long before robot trucks replaced actual truckers, changing the political economy of logistics in the US. This is not remotely close to happening (as one of that program’s guests, a trucker, pointed out) and so, I wondered why this idea was asserted with the same confidence of a Tesla press release about full self-driving…happening, any day now.
The reason is a lack of understanding of how actually existing computational systems work. This isn’t a sin; the world is complex and we all can’t be experts in everything (though there’s a large army of men who assume they can, for example, perform surgery, fly fighter jets and wrestle bears – the scientific term for such men is idiot). As it happens, my decades of experience with computation, combined with an unequivocally Marxist (therefore, materialist) understanding of capitalism seemed to make me qualified to fill this niche from a unique perspective – not from the distance of academics but feeling the cold chill of data centers.
And so, I started this blog, a sisyphean effort, of unknown utility but necessary, if only to help me achieve some measure of clarity.
But, how to write about the tech industry? What ‘voice’, to lean on a cliche, should be used? In the beginning, I wrote like a war correspondent (or at least, what I supposed to be the attitude of a war correspondent) : urgent, sparse, accessibly technical. The enemy was clearly identified, the stark facts countering mythology plainly stated. There was no time for leisurely applied words. In an earlier age, when fedoras and smoking on planes were common, this might have been called a ‘muscular’ style (which evokes the image of a body builder, busily typing on a keyboard after leg day at the gym). I imagined myself in a smart, yet disheveled suit, sitting on-set with Dick Cavett in a forever 1969 Manhattan, a Norman Mailer of tech critique, though without the nasty obsession with performative manliness.
Something pulls at me, another ‘voice’ which has moved me, by degrees, away from reports from the front to an even sharper-edged approach, one informed by a combination of disdain for the target – an intrusive and destructive industry – and deep concern for its victims: all of us, nearly everywhere. This writing personna is closer to my day to day self – not a perfect mirror, but more recognizable.
This person at the keyboard, this version of Dwayne who tries to convey to you, esteemed reader, the true danger posed by the tech industry and the various illusions it promotes, is a man who refuses to be fooled or, at least, to walk into delusion willingly, without struggle.
Now, as I write, my thoughts turn to an essay about detective fiction Raymond Chandler penned for The Atlantic magazine in 1950 titled, ‘The Simple Art of Murder.’ About writing, as a craft, Chandler wrote:
“The poor writer is dishonest without knowing it, and the fairly good one can be dishonest because he doesn’t know what to be honest about.”
Honesty. This is the goal; an honest accounting of the situation we’re in and what we’re up against – capitalist political economy, supply chains, resources extraction and data centers as a form of sociotechnical power – a rejection of the Californian Ideology; no, not just a rejection, but a hard boiled reaction to it, a Noir response.
To close this, which is a work in progress, let’s return to Chandler’s essay about detective fiction, “The Simple Art of Murder” –
“It is not a very fragrant world, but it is the world you live in, and certain writers with tough minds and a cool spirit of detachment can make very interesting and even amusing patterns out of it. It is not funny that a man should be killed, but it is sometimes funny that he should be killed for so little, and that his death should be the coin of what we call civilization. All this still is not quite enough.”
The world itself may be lovely but the world the tech industry has built and which it seeks to entrench is ‘not very fragrant’ indeed; in fact, it is a nightmare. Resistance requires passion but also, as Chandler wrote of his fictional hero, Phillip Marlowe, a tough mind and cool spirit of detachment. No, I will not celebrate AI and each gyration of an industry whose goal is to act as the means through which labor’s power is suppressed.
Enough wide eyed belief; time for productive cynicism.
I have avoided writing about ChatGPT as one might hurriedly walk past a group of co-workers, gathered around a box of donuts who’re talking about a popular movie or show; to avoid being drawn into the inevitable.
In some circles, certainly the circles I travel in, ChatGPT is the relentless talk of the town. Everyone from LinkedIn hucksters who claimed to be making millions from the platform, only moments after it was released, to the usual ‘AI’ enthusiasts who take any opportunity to sweatily declare a new era of machine intelligence upon us – and of course, a scattering of people carefully analyzing the actually existing nuts and bolts – everyone seems to be promoting, debating and shouting about ChatGPT.
You can imagine me, dear reader, in the midst of this drama, quietly sitting in a timeworn leather chair, slowly sipping a glass of wine while a stream of text, video and audio, all about ChatGPT, that silicon, would-be Golem, washes over me
What roused me from my torpor was the news Microsoft was investing 10 billion dollars in OpenAI, the organization behind ChatGPT and other ballyhooed large language model systems (see: “Microsoft’s $10bn bet on ChatGPT developer marks new era of AI”). Even for Microsoft, that’s a lot of money. Behind all this, is Microsoft’s significant investment in what it calls purpose built, AI supercomputers such as VOYAGER-EUS2 to train and host platforms such as ChatGPT. Although tender minded naifs believe corporations are using large scale computation to advance humanity, more sober minds are inclined to ask fundamental questions such as, why?
The answer came from the Microsoft article, “General availability of Azure OpenAI Service expands access to large, advanced AI models with added enterprise benefits.” Note that phrase, ‘enterprise benefits.’ The audience for this article is surely techie and techie adjacent (and here, I must raise my hand) but even if neither of these categories describes you I suggest giving it a read. There’s also an introductory video, providing a walkthrough of using the OpenAI tooling that’s mediated via the Microsoft Azure cloud platform.
As I watched this video, the purpose of all those billions and the hardware it bought became clear to me; Microsoft and its chief competitors, Amazon and an apparently panicked Google (plus, less well known organizations) are seeking to extend the rentier model of cloud computing, which turns computation, storage and database services into a rented utility and recurring revenue source for the cloud firm that maintains the hardware – even for the largest corporate customers – into the ‘AI’ space, creating super rentier platforms which will spawn subordinate, sub-rentier platforms:
Imagine the following…
A San Francisco based startup, let’s give it a terrible name, Talkist, announces it has developed a remarkable, groundbreaking chat application (and by the way, ‘groundbreaking’ is required alongside ‘next generation’) which will enable companies around the world to replace customer service personnel with Talkist’s ‘intelligent’, ‘ethical’ system. Talkist, which only consists of a few people (mostly men) and a stereotypical, ‘visionary’ leader, probably wearing a thousand dollar t-shirt, doesn’t have the capital, or the desire to build the computational infrastructure required to host such a system.
This is where the Azure/OpenAI complex of systems comes to the rescue of our plucky band of well-funded San Franciscans. Instead of diverting precious venture capital into purchasing data center space and the computers to fill it, that money can be poured into creating applications which utilize Microsoft/OpenAI cloud services. Microsoft/OpenAI rent ‘AI’ capabilities to Talkist who in turn, rent ‘AI’ capabilities to other companies who think they can replace people with text generating, pattern matching systems (ironically, OpenAI itself is dependent on exploited labor as the Time Magazine article, “OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic” shows).
What a time to be alive.
Of course, the uses (and from the perspective of profit-driven organizations, cost savings) don’t end with chatty software. We can imagine magazines and other publications, weary of having to employ troublesome human beings with their demands for salaries, health care and decent lives (The gall! Are there no workhouses? Are there no prisons?) rushing to use these systems to ‘write’ – or perhaps we should say, mechanistically assemble, articles and news stories, reducing the need for writers who are an annoying class (I wink at you dear reader for I am the opposite of annoying – being a delightful mixture of cologne, Bordeaux and dialectical analysis). Unsurprisingly, and let’s indulge our desire for a bit of the old schadenfreude, amusingly there are problems such as those detailed in the articles “CNET Is Reviewing the Accuracy of All Its AI-Written Articles After Multiple Major Corrections. and, “CNET’s AI Journalist Appears to Have Committed Extensive Plagiarism.”
…
Of all the empires that have stalked the Earth, the tech imperium is, perhaps, the bullshitiest. The Romans derived their power from myths, yes, but also, roads, aqueducts and organized violence – real things in a real world. The US empire has its own set of myths, such as a belief that sitting in a car, in traffic, is the pinnacle of freedom and in meritocracy (a notion wielded by the most mediocre minds to explain their comforts). Once again however, real things, such as possessing the world’s reserve currency and the capacity for ultra-violence lurk behind the curtain.
The tech empire, by contrast, is built, using the Monorail maneuver detailed in this Simpsons episode, on false claims prettily presented. It has inserted itself between us and the things we need – information, memories, creativity. The tech industry has hijacked a variety of commons and then rents us access to what should be open. In its ‘AI’ incarnation, the tech industry attempts to replace human reason with computer power, a fool’s errand, which computer scientist Joseph Weizenbaum dissected almost 50 years ago, but a goal motivated by a desire to increase the rate of profit in an era of creeping stagnation by reducing the need for labor.
Rather than being a refutation of Marx and Engel’s analysis as some, such as Yanis Varoufakis with his ‘cloudalist’ hypothesis bafflingly claim, we are indeed, still very much dealing with the human grinding workings of capitalist logics, wearing a prop, science fiction film costume, claiming to have come in peace.
ChatGPT isn’t a research platform or the herald of a new age of computation; it is the embodiment of the revenue stream dreams of the tech industry, the super-rentier.
AT THE END of ´The Wizard of OZ´, Metro-Goldwyn-Mayer´s 1939-released, surrealist musical fantasy, our heroine Dorothy and her loyal comrades complete a long, arduous (but song filled) journey, finally reaching the fabled city of OZ. In OZ, according to a tunefully stated legend, there’s a wizard who possesses the power to grant any wish, no matter how outlandish. Dorothy, marooned in OZ, only wishes to return home and for her friends to receive their various hearts´ desire.
As they cautiously approach the Wizard’s chamber, Dorothy and her friends are met with a display of light, flame and sound; ¨who dares!?¨ a deafening voice demands. It’s quite a show of apparent fury but illusion crumbles when it’s revealed (by Dorothy´s dog, Toto) that behind it all is a rather ordinary man, hidden on the other side of a velvet curtain, frantically pulling levers and spinning dials to keep the machinery powering the illusion going while shouting, “pay no attention to that man behind the curtain!“
Behind the appearance of magic, there was a noisy industrial process, belching smoke. Instead of following the Wizard’s advice to pay no attention, let’s pay very close attention indeed to what lies behind appearances.
THERE’S AN INESCAPABLE MATERIALITY behind what’s called ‘AI’ deliberately obscured under a mountain of hype, flashy images and claims of impending ‘artificial general intelligence’ or ‘AGI’ as it’s known in sales brochures disguised as scientific papers.
At the heart of the success of techniques such as large language models, starting in the latter 2010s, is the graphics processing unit or GPU (in this essay about Meta´s OPT-175B, I provide an example of how GPUs are used). These devices use a parallel architecture, which enables greater performance than the general purpose processors used for your laptop; this vastly greater capability is the reason GPUs are commonly used for demanding applications such as games and now, the hyper-scale pattern matching behind so-called ´AI´ systems.
All of the celebrated feats of ‘AI’ – platforms such as Dall-E, GPT-3 and so on, are completely dependent on the use of some form of GPU, most likely provided by NVIDIA, the leading company in this space. OpenAI, a Microsoft partner, uses that company’s Azure cloud but within those ´cloud´ data centers, there are thousands upon thousands of GPUs, consuming power and requiring near constant monitoring to replace failed units.
GPUs are constructed as the result of a long and complex supply chain involving resource extraction, manufacturing, shipping and distribution; even a sales team. ‘AI’ luminaries and their camp followers, the army of bloggers, podcasters and researchers who promote the field, routinely and self-indulgently debate a variety of esoteric topics (if you follow the ´AI´ topic on Twitter, for example, odds are you have observed and perhaps participated in these discussions about vague topics such as, ´the nature of intelligence´) but it’s GPUs and their dependencies all the way down
GPU raw and processed material inputs are aluminum, copper, clad laminates, glass, fibers, thermal silica gel, tantalum and tungsten. Every time an industry partisan tries to ‘AI’-splain the field, declaring it to be a form of magic, ignore their over-determination and confusion of feedback loops with cognition and think of those raw materials, ripped from the ground.
The ‘AI’ industrial complex is beset by two self-serving fantasies:
1.) We are building intelligence
2.) The supply chain feeding the industry is infinite and can ‘scale is all you need’ its way forever to a brave new world.
For now, this industry has been able to keep the levers and dials moving, but the amount of effort required will only grow as the uses to which this technology is put expand (Amazon alone seems determined to find as many ways to consume computational infrastructure as possible with a devil take the hindmost disregard for consequences), the need for processors grows and global supply chains are stressed by factors such as climate change, and geopolitical fragmentation.
The Wizards, out of tricks, curtains pulled, will be revealed as the ordinary (mostly) men they are. What comes next, will be up to us.
Some Key References:
According to self-satisfied legend, medieval European scholars, perhaps short of things to do compared to we ever-occupied moderns, spent countless hours wondering about topics such as, how many angels could simultaneously occupy the head of a pin; the idea being that, if nothing was impossible for God, surely, violating the observable rules of space and temporality should be cosmic child’s play for the deity…but to what extent?
How many angels, oh lord?
Although it’s debatable whether this question actually kept any monks up at night more than, say, wondering where the best beer was, the core idea, that it’s possible to get lost in a maze of interesting, but ultimately pointless inquiries (a category which, in an ancient Buddhist text is labeled, ‘questions that tend not towards edification’) remains eternally relevant.
At this stage in our history, as we stare, dumbfounded, into the barrels of several weapons of capitalism’s making – climate change being the most devastating – the AI endeavor is the computational equivalent of that apocryphal medieval debating topic; we are discussing the ethics of large language models, focusing, understandably, on biased language and power consumption but missing a more pointed ethical question: should these systems exist at all? A more, shall we say, robust ethics would demand that in the face of our complex of global emergencies, tolerance for the use of computational power for games with language cannot be justified.
The company now known as Meta recently announced its creation of a large language model system called OPT-175B. Helpfully, and unlike the not particularly open OpenAI, the announcement was accompanied by the publication of a detailed technical review, which you can read here.
As the paper’s authors promise in the abstract, the document is quite rich in details which, to those unfamiliar with the industry’s terminology and jargon, will likely be off-putting. That’s okay because I read it for you and can distill the results to four main items:
Here’s an excerpt from the paper:
“From this implementation, and from using the latest generation of NVIDIA hardware, we are able to develop OPT-175B using only 1/7th the carbon footprint of GPT-3.
While this is a significant achievement, the energy cost of creating such a model is still nontrivial, and repeated efforts to replicate a model of this size will only amplify the growing compute footprint of these LLMs.” [highlighting emphasis mine]
https://arxiv.org/pdf/2205.01068.pdf
I cooked up a visual to place this in a fuller context:
Here’s a bit more from the paper about hardware:
“We faced a significant number of hardware failures in our compute cluster while training OPT-175B.
In total, hardware failures contributed to at least 35 manual restarts and the cycling of over 100 hosts over the course of 2 months.
During manual restarts, the training run was paused, and a series of diagnostics tests were conducted to detect problematic nodes.
Flagged nodes were then cordoned off and training was resumed from the last saved checkpoint.
https://arxiv.org/pdf/2205.01068.pdf
Given the difference between the number of hosts cycled out and the number of manual restarts, we estimate 70+ automatic restarts due to hardware failures.”
All of which means that, while processing data, there were times, quite a few times, when parts of the system failed, requiring a pause till fixed or routed around (resumed, once the failing elements were replaced).
Let’s pause here to reflect on where we are in the story; a system, whose purpose is to produce plausible strings of text (and, stripped of the obscurants of mathematics, large-scale systems engineering and marketing hype, this is what large language models do) was assembled using a small mountain of computer processors, prone, to a non-trivial extent, to failure.
As pin carrying capacity counting goes, this is rather expensive.
Like other LLMs, OPT-175B has a tendency to return hate speech as output. Another excerpt:
“Overall, we see that OPT-175B has a higher toxicity rate than either PaLM or Davinci. We also observe that all 3 models have increased likelihood of generating toxic continuations as the toxicity of the prompt increases, which is consistent with the observations of Chowdhery et al. (2022). As with our experiments in hate speech detection, we suspect the inclusion of unmoderated social media texts in the pre-training corpus raises model familiarity with, and therefore propensity to generate and detect, toxic text.” [bold emphasis mine]
https://arxiv.org/pdf/2205.01068.pdf
Unsurprisingly, there’s been a lot of commentary on Twitter (and no doubt, elsewhere) about this toxicity. Indeed, almost the entire focus of ‘ethical’ efforts has been on somehow engineering this tendency away – or perhaps avoiding it altogether via the use of less volatile datasets (and good luck with that as long as Internet data is in the mix!)
This defines ethics as being the task of improving a system’s outputs – a technical activity – and not a consideration of a system as a whole from an ethical standpoint within political economy. Or to put it another way, the ethical task is narrowed to making sure that if I use a service which, on its backend, depends on a language model for its apparent text capability, it won’t in the midst of telling me about good nearby restaurants, hurl insults like a klan member.
Within the paper itself, there is the foundation of an argument against this entire field, as currently pursued:
“...there exists significant compute and carbon cost to reproduce models of this size. While OPT-175B was developed with an estimated carbon emissions footprint (CO2eq) of 75 tons,10 GPT-3 was estimated to use 500 tons, while Gopher required 380 tons. These estimates are not universally reported, and the accounting methodologies for these calculations are also not standardized. In addition, model training is only one component of the over- all carbon footprint of AI systems; we must also consider experimentation and eventual downstream inference cost, all of which contribute to the growing energy footprint of creating large-scale models.”
In the fictional history of the far-future world depicted in the novel ‘Dune’ there was an event, the Butlerian Jihad, which decisively swept thinking machines from galactic civilization. This purge was inspired by the interpretation of devices that mimicked thought or possessed the capacity to think as an abomination against nature.
Today, we do not face the challenge of thinking machines and probably never will. What we do face however, is an urgent need to, at long last, take climate change seriously. How should this reorientation towards soberness alter our understanding of the role of computation?
I think that, in face of an ever-shortening amount of time to address climate change in an organized fashion, the continuation, to say nothing of expansion of this industrial level consumption of resources, computing power, talent and the corresponding carbon footprint is ethically and morally unacceptable.
At this late hour, the ethical position isn’t to call for, or work towards better use of these massive systems; it’s to demand they be halted and the computational capacity re-purposed for more pressing issues. We can no longer afford to wonder how many angels we can get to dance on pins.
When I was young, the word, ‘supercomputer’ evoked images of powerful, intelligent systems, filling the cavities of mountains with their humming electronic menace.
Science fiction encouraged this view, which is as far from the (still impressive, yet grounded) reality of supercomputing as the Earth is from some distant galaxy. The distance between marketing hype and actually existing machines is like that: vast and unbridgeable, except in dreams.
Which brings me to this Verge story, posted on 24 January, 2022:
Social media conglomerate Meta is the latest tech company to build an “AI supercomputer” — a high-speed computer designed specifically to train machine learning systems. The company says its new AI Research SuperCluster, or RSC, is already among the fastest machines of its type and, when complete in mid-2022, will be the world’s fastest.
“Meta has developed what we believe is the world’s fastest AI supercomputer,” said Meta CEO Mark Zuckerberg in a statement. “We’re calling it RSC for AI Research SuperCluster and it’ll be complete later this year.”
Verge: https://www.theverge.com/2022/1/24/22898651/meta-artificial-intelligence-ai-supercomputer-rsc-2022
The phrase, “AI supercomputer” is obviously designed to sell the idea that this supercomputer, unlike others, is optimized for AI. And to give the devil his due, the fact it’s reportedly composed of NVIDIA game processing units, which, since the mid 2000’s have found extensive use powering tasks such as building large language models, gives some amount of credibility to the claim.
Some, but not as much as it might seem. Consider this hyperventilating article:
This is clearly the tone Meta (and others) is hoping to cultivate via the use of ‘AI supercomputer’ as a descriptor. The assumption is that if enough computational power is thrown at the task of building machine learning models, those models will, in some not sharply defined way, reach unprecedented heights of…well, one isn’t sure.
Are ever larger machine learning models a sure indicator of remarkable progress? Two papers, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” and “No News is Good News: A Critique of the One Billion Word Benchmark” suggest the answer is no. These papers are focused on Natural Language Processing (NLP) models and it’s suggested that Meta will be building models for its Second Life warmed over ‘Metaverse’ effort. Even so, there appears to be a point at which ever larger models fail to produce hoped for results.
Supercomputers: Our Old Drinking Buddies
The category, ‘supercomputer’, created to describe a class of tightly integrated, high performance computational platforms, has existed for over 60 years. The first supercomputers were developed for nuclear research (weapons and energy) at Lawrence Livermore Labs in the US at the height of the Cold War (maybe we should call it Cold War Classic) and have also been applied to demanding tasks such as modeling the Earth’s climate. It’s a venerable technology with clearly defined parameters such as the use of symmetric multiprocessing. In all these decades, no supercomputer has managed to exhibit intelligence or plot our demise, except in fiction.
Adding ‘AI’ to the mix doesn’t change that reality since ever larger statistical pattern matching techniques do not cognition make. Oh and Meta’s claim is that these types of supercomputing data centers will, in addition to serving as development platforms, also host the haunted cartoon castle they call the “Metaverse’.
Considering this statement from Intel we have reason to doubt this too.
This blog is dedicated to analyzing the field of ‘artificial intelligence’ specifically, and the tech industry generally from a materialist perspective.
By materialist perspective, I mean a point of view informed, in large measure by historical materialism but also, by a deep analysis of the nuts and bolts – the supply chains, extractive industrial activity, data centers and actually existing capabilities undergirding what the tech industry presents as magic.
There is however, another layer operating which is captured by, among other disciplines, psychoanalytic theory.
To explore this approach, I’ve been studying (‘reading’ is too light a word for it) ‘The Psychoanalysis of Artificial Intelligence‘ by Dr. Isabel Millar.
In this series of video meditations, I consider aspects of the book. This is an ongoing project and I’ll be updating this post as new videos are added.
Introduction:
Part Two:
Part Three:
Computational Impacts 