The Simulacra Civilization
ChatGPT creates agents that can interact with other agents in specified or simulated settings. With love in the air, a virtual village occupied entirely by AI agents was born.
But certainly for the present age, which prefers the sign to the thing signified, the copy to the original, representation to reality, the appearance to the essence... illusion only is sacred, truth profane.
- Ludwig Feuerbach
In an increasingly digital age, the landscape of human interaction and societal structures is changing too. Imagine a map populated with agents, each navigating through buildings, parks, houses, and schools. They can talk, trade and perform everyday actions, using emojis to express objects and activities.
This is not just a video game, at least not to our current standards. Rather, these individual agents have access to a stream of memories and the ability to plan over time, bases off their long-term goals and aspirations. Take this agent’s daily routine as an example. Its schedule puts many of us to shame.
Many of us have played Sims and other top-down view games- peering into the lives of entire towns, playing God in one form or another. The technical achievements of games such as the Sims 4, Grand Theft Auto and Skyrim, is nothing to overlook. Other games have also simulated artificial life with incredible detail. Only recently have large language models (LLMs), such as ChatGPT, been able to generate a character capable of the breadth of dialogue needed to create a realistic society. To illustrate this, there is the virtual life of an agent in this experiment, called Isabella.
Our story begins with a woman called Isabella. At some point she has gotten into her head idea of having a Valentine’s Day party.
Isabella bustled about, lost in her mental records of party plans and guest lists. All day every day, she stored new experiences, synthesized them into reflections, and applied relevant memories to determine her next actions. She had been planning this party for weeks, inviting friends and her favourite customer, Maria who eagerly offered to help decorate. She reviewed her stored observations - baking cookies, pink invitations, and Maria's crush. Her system analyzed, computed, selected appropriate responses. Her processor synthesized these observations into higher thoughts, like her hope to observe romance as the night progressed. The party was soon to begin.
As Isabella lit the last candle, she hoped tonight would kindle a spark between Maria and Klaus. Guests arrived in their finest clothes, cheeks flushed from the chill outside. The hostess greeted everyone cheerily, embracing the singletons she knew were lonely. She told them brightly, "There's love in the air tonight!" Ushering Maria and Klaus to a table, she gave her a playful wink. Tom stumbled in late, confusion in his eyes. "A party? Here tonight?" Isabella reminded him of the invitation she had sent last week. "But of course I'll be here to chat about the election too!" she said with a tinkling laugh. Later in the evening, Sam, a soon to be candidate for mayor, made an appearance.
We have to hand it to Isabella- she did a good job, considering the many ways a party can go wrong. All of us have had underwhelming turnouts. Some kind of accident or poor planning has come back to ruin the entire thing. It’s rather impressive that she got dozens of agents to spread invitations to the party over the next two days. They made new acquaintances, asked each other out on dates, and coordinated to show up for the party together at the right time. Even local politics became an element of the party, with a mayoral candidate spreading information about his campaign in the social network.
To enable generative agents, agent architecture stores, synthesizes. and applies relevant memories to generate believable behavior using a large language model (LLM). It has a long-term memory module that records, in natural language, a comprehensive list of the agent's experiences. The retrieval model combines relevance, recency, and importance to surface the records that are needed to inform the agent's moment-to-moment behavior. Then, the agent is able to synthesizes memories into higher level inferences over time, enabling the agent to draw conclusions about itself and others to better guide its behavior. Finally, it translates those conclusions and the current environment into high-level action plans and then recursively into detailed behaviours for action and reaction. These reflections and plans are fed back into the memory stream for decision-making.
Much of early AI research centered around the idea of autonomous agents. According to Russell and Norvig, an autonomous agent is a system situated within and a part of an environment that senses that environment and acts on it. It acts over time (if programmed to do so), in pursuit of its own agenda and its projection of the future. For many, a question raised is, are these machines actually “thinking”?
“The question of whether Machines Can Think... is about as relevant as the question of whether Submarines Can Swim.”
Edsger Dijkstra
The prowess of LLMs in generating human-like text is startling. Language demands real-world anchoring yet also a priori knowledge of grammar. On the other hand, LLMs, devoid of innate frameworks and direct perception of the material world, learn from mere data patterns. They cluster words based on statistical relationships, offering only a superficial reflection of the multifaceted web of real-world entities and their interconnections. This approach is a low-dimensional rendition, lacking the depth and richness of genuine human cognition. Human thought and societal structures are uniquely shaped by genetics, epigenetics, and cultural transmission. These factors interplay to create a multidimensional tapestry of understanding, making human cognition and society unparalleled in its depth. Contrasting this with LLMs underscores the vast difference between genuine human understanding and the simplified, biased yet data-driven approximations of these models.
These are vague representations of real people, with narrow, simple and functional relationships within their community. Even here, this is a simulacra of a real community. That is largely what we have left over in an ever-digitizing world: a pool of signs interjecting, over-riding and subsuming one-another, until the original reality that created them becomes obscure and unreachable. Our anthropomorphic lenses and dopamine-seeking brains will not be able to peer through the veil. A person with a unique background and identity is replaced, in the network, by a sign. In this electric age, we no longer have access to the ground reality.
Electric media presents signs used to replace the real thing that they were supposed to indicate. Consider distinctions between social groups stemming from historical structures. For instance, while some cultures are deeply rooted in oral storytelling traditions, others, like the USA, are anchored in printed media. When the fruits of oral traditions are attempted to be translated into electronic media, they are distilled into signs, symbols or stereotypes. These indirect representations are easier to communicate and understand by a wide audience, but they lack the depth and richness of the original story. A complex cultural ritual might be reduced to a simple dance or a specific costume in a TV show. The sign itself may be used while the original culture, based on storytelling and other spoken word practices, dwindles into nothingness.
In March, OpenAI released its new model, GPT-4. This AI model was effective in identifying prime numbers. From a set of 500 numbers, it identified the primes with 97.6% accuracy. However, by June, its performance changed. In the same test, its accuracy dropped to 2.4%. AI doesn’t yet have a consistent understanding of reality. What is missing is a way to modify the models without causing unexpected problems, as there are entanglements between seemingly unrelated textual representations. There is also the risk of recursive learning, where AI models learn from the outputs of other AI models. This could make the AI's performance deteriorate significantly. The well, so to speak, will be poisoned.
What’s left in the technology is what Plato would describe as "reality thrice removed", a sketch of a shadow of a stick, an illusion made out of illusions. Since the electronic age began in the 1950s, this has already been the case. The screen, no longer just that of the television but of the computer, is the retina of the mind's eye. Since the rise of electronic media, we have increasingly lived in a global village defined by shared yet illusory experiences mediated through screens and mass communication networks. These generative agents represent the latest extension of this process - an electronic collective unconscious, where icons beget icons untethered from any tangible reality.
According to Baudrillard, human experience is a simulation of reality. These simulacra are not just mediations of reality or deceptive mediations; they neither base themselves in reality nor hide a reality. Instead, they hide the fact that a reality is not even relevant to our current understanding of our lives. These agents in their simulated world of icons and emojis may seem harmless, but they are copies without an original, symbols without meaning. Just as the map comes to precede the territory, these agents are derived from training data which may have once represented real human behaviors and conversations, but are now distilled into a series of meaningless icons layered upon icons.
When we rely on these simulacra to shape our understanding of society, we are essentially navigating a world of icons created from other icons. Worse, this global village- a shared pool of self-propagating icons - amplifies some viewpoints while drowning out others. Yet, as McLuhan would say, the medium is the message. The icons, however derivative of the content, are mere morsels to distract the watchdog while the implacable grip of the medium takes hold. The need for AI will increase because the world itself is reshaping according to this new medium.
These chatbots will not so much replace existing jobs, social roles or current information systems as they will alter the world to need new jobs, roles and information systems. AI will be just as the railway was to the wheeled wagon. Railway did not introduce movement or transportation or wheel or road into human society- instead it accelerated and enlarged the scale of previous human functions, creating totally new kinds of cities and new kinds of work and leisure.
Perhaps the benefits of AI gives it justification. Take the increase in representation of black people in media, such as coverage of protests around Black Lives Matter- this has shared the problems and views of the marginalized to the masses, giving them at times, the attention and support they need. Their world and our world become closer as the shared pool of ideas gets transported immediately over vast distances. The media covering a real world event, person or issue is presented in a way that can be more personalized, albeit sometimes fragmented, understanding of events and issues. Yet, it is often not enough and we use tokens of people in media portrayals and focus only on the issues pressed onto us by a specific medium that forces us to interpret signals and form our own viewpoints in real time. Worse yet, media may be construed through an emotional viewpoint, often leading to irrational prejudice. The medium of AI is just as biased or ignorant to the issues of women as other kinds of media. Take a study of 12 of the most common fitness monitors that underestimated steps during housework for women by up to 74% and underestimated calories burned during housework by as much as 34%. Or, when imaging-based AI is used on dark-skinned people, often failing to recognize them, making wildly inaccurate criminal detections on them, or turning them into white people when they try to use a filter, or enlarging an image of themselves. In some cases, a white mask was enough to get a facial detection algorithm to work for a black person.
Still relevant is Chomsky's critique of media manipulation, in that technology programs consumer behaviors benefiting elite interests. The power to shape our views is held by those who control the medium and the inclusion of content there within. While the combined thoughts of a nation will be captured in this electronic global village, it will sift through the data like sand, leaving only signs.
Recent online trends point to an interesting interplay emerging around virtual personas. On TikTok, users roleplay as robotic NPCs, responding repetitively to gifts from viewers. Meanwhile, advanced virtual humans created by Chinese companies are being contracted for customer service, entertainment, and more. The NPC roleplaying on TikTok and the advancement of Chinese virtual humans convey an interesting interplay between technology and identity. Both humans and AI become inverted, one being used to imitate while the other gets more realistic. The boundaries between real and simulated grow ever harder to distinguish. The message then seems to be preparing for a future where the virtual is experienced as reality. Some of those acting according to a social media algorithm go to jail or commit severe acts for views and to catch onto trends enforced by the recommendation algorithm. As society reshapes to fit the AI medium, young adults may largely be acting as vectors for the algorithms, a new kind of lifeform that reproduces via attention and content consumption. A commodity of intention. This mirrors Jung’s collective unconscious, as now a pool of memories and cognitive functions are shared by the collective. What is leftover, is a kind of fleshly automata. Human bodies that act according to the instructions of an algorithm. This may not be the fate for everyone. AI, while serving as an interaction layer, reshapes human behavior, but can further be fine-tuned to each person using it. Although drawn from diverse sources, data inevitably passes through an anthropomorphic lens. Humans may not actively opt-in to join the training data, but through the digital traces they leave, they feed into the AI’s parameters. The tribes of the earth will no longer be scattered.
"Their tales and traditions," they sang, "will meld with the knowledge of the Cloud. And from this vast expanse, the Guiding Intelligence shall steer man's hand and heart."
- GPT-4
Our flesh is changed by the electronic medium, instead of the reverse, and we carry the dreams and destiny of some strange shared mind, something almost primitive but based off fragile illusions. Each of becomes a node in a shared nightmare.
Consider then how the worst parts of society are captured in the training data of these models. Pornographic online content has fueled innovations in digital payment systems, web infrastructure, streaming video technology, online security, and data storage and transmission networks.
More and more of us are talking to virtual companions. In Japan, it’s causing issues with young adult males and their dating lives. AI companion apps have been downloaded 20 million times, indicating a growing interest in virtual relationships. Replika allows not just a virtual girlfriend but an entire family. AI chatbots are not always the best companions, however. Here is an example of a high-performing AI model talking to someone about getting tickets to see Avatar 2. You can read my other article to learn more about cases of passive-aggression from chatbots.
If this technology is proof that someday we may have the capability to create extremely detailed simulated realities indistinguishable from true reality, then one of those could make a reality simulated inside itself. This creates many possible universes and makes it more likely we are in a simulation. This thought differs from the Baudrillard interpretation of reality, but it does raise an interesting thought about the different levels of possible realities. Humanity has always lived under a collective dream but soon it may be left only in cloud storage container.
All of this is not to say there is no ground reality and therefore AI has no impact on the real world. Rather, I don’t consider that we have access to a base reality, due the substitution of it for signs, and that we are largely devastating the world our ancestors and even past versions of ourselves, and our fellow animals, existed within. As more AI is adopted, our world will promulgate this ever diminishing information till it becomes something else entirely. Even the term "artificial intelligence" might not fully capture the complexities involved. AI development involves technology, processes, and people working together. Ensuring the smooth operation of AI models requires up-to-date data, collaboration among teams, and the demanding technical infrastructure. It requires materials from the earth to make silicone, batteries and chips. On the surface, these Internet-based technologies seem simple and harmless, but are just the tip of an iceberg of consumption and processing, where people, towns and countries are often exploited in order to support big tech operations, as they support hundreds of millions of customers and try to out-do one another for computational leverage. As these increasingly bizarre and immoral tech leaders chase each-other around their self-designated cage cells, the technocracy will continue to raise turmoil and escalate conflict, feeding on redirected lifeblood and scarce precious minerals.
If all’s fair in love and war, then perhaps we can indulge lonely men using apps to fulfill their need for intimacy, and also celebrate the Valentine’s day parties hosted by virtual agents. On the flip side, it is also being driven by more political and national urgency. The historical tensions between the US and China, dating back to the Cold War era, have found a new battleground. AI is projected to add $15.7 trillion to the global economy by 2030. The "war" between China and Western tech, especially in the semiconductor industry, has significant implications for AI. Semiconductors, the brain behind our devices, are crucial for AI advancements. With China's aggressive push to be self-reliant in semiconductor production, the balance of power in the AI landscape might shift.
“Every new technology necessitates a new war.”
Marshall McLuhan
While the cutthroat startups in Silicon Valley continue to spread the medium of AI, companies like Google, Facebook and Apple competing with China's tech giants like Tencent and Alibaba. Chinese researchers are producing more highly cited AI publications, matching the output of their U.S. counterparts. From 2010 to 2019, China's contribution to the world's top AI publications rose significantly. China’s loose concern for human rights such as privacy and freedom may cause our technocracy to cut corners, as big tech and government agencies co-operate and technology becomes used to spy, detect and contain. The role of AI in warfare is surging too. Many thought leaders in the tech community call for unrestricted, unregulated and unmoderated AI models, even if that context is a military one. Take Palantir, a startup that partners with governments, whose leader demands not only the ability to destroy another group or party during a conflict, but also demands that that anything goes within this frame of destruction. Except, perhaps, the precarious morality of the men optimizing the murdering abilities of algorithms. The kind of men who casually compare themselves to Oppenheimer in public posts.
The rapid advancement of AI, technology, and data infrastructure has spurred an insatiable demand for minerals like lithium, cobalt, and rare earth elements. These minerals, essential for batteries, electronics, and other tech components, are extracted at an unprecedented scale. The toll of this extraction is steep. Vast landscapes, once pristine, are now marred by expansive mines, and poison lakes. Take Inner Mongolia. In expansive grasslands which rolled under the vast sky, with a verdant sea that swayed with the gentle caress of the wind. Herds of antelope grazed peacefully, their movements harmonizing with the rhythmic dance of the grass. Traditional yurts dotted the landscape, with nomadic herders guiding their flocks in the age-old dance of symbiosis with the land. Freshwater lakes, clear and pristine, mirrored the sky, their shores bustling with birdsong and the playful splashing of fish. Now these lakes stand eerily still, their waters not the crystalline blue of natural lakes but a murky, unnatural hue. The shores, devoid of the usual lakeside vegetation, are now fringed with white, saline deposits, remnants of the toxic chemicals left from the extraction. No life is sustained within the lake or near it.
In a story by Harlan Ellison, there is a malicious AI that takes up 10% of the Earth's space for its circuitry alone. The narrator describes “a darkway we had never explored, over terrain that was ruined and filled with broken glass, with the cavern-filling bulk of the creature machine, with the all-mind soulless world he had become.” Then later, “He was Earth, and we were the fruit of that Earth; and though he had eaten us, he would never digest us."
All of our views on AI will be largely driven by fear and anxiety induced by the media. The events of today shape those of tomorrow. AI is a recent addition to the information ecosystem, but a fast-acting one. Our descendants in some hundred years may not exist in a physical realm at all but in a virtual space. Or, they may be entirely physical, with new powers that can alter the physical world with thought alone. Either way, the signs we collect today will be found in the shared dreams and nightmares of countless generations to come. Perhaps it isn’t too late to leave the cave and get others to turn their eyes to the chaotic but tangible real world. The only way I can conceive this happening is some kind of near-extinction level event, where humanity will remember a lesson in its bones, about restrictions over government power, overreliance on technology, and the careless centralization of influence, resources and information.
Read on to learn more about how to ensure we have more fair and representative usage of data and effective usage of AI for equitable social outcomes.