This AI Minecraft Simulation Is Terrifying Scientists

This looks like a game, but it's actually an AI simulation. Scientists dropped over a thousand AI agents into a Minecraft server to see what they would do, how they would interact with each other, what they would build and create, and what they would want. These digital characters powered by one of the world's most advanced AI models were given essentially no instruction other than the following prompt. To survive with fellow players in Minecraft normal survival mode and create an efficient Minecraft village. What followed was simultaneously bizarre, profound, unsettling, and perhaps most importantly, highly insightful into the dynamics of civilization, the nature of modern technology, and the potential future of everything. The Misconceptions of AI Most of us don't really understand AI. Even those of us who are technically proficient in it, still arguably don't always fully understand what's being built, how the whole end-to-end processes of these systems work, and what the sort of ontological nature of the technology is or could be. Relative to all other technologies, tools, and man-made phenomena, AI is highly unique and sort of straddles the line of our pre-existing categories. It's part tool and software, but it's also part something else. Something agentic and kind of alien-like, with an entirely different ontological framework and the capacity to do things that are hard to cleanly understand and categorize. This technology is also extremely fast and complex. Often, our popular conceptions of it can quickly become outdated. By the time cultural assessments and understandings take shape, the reality of the form and depth behind the shape has, in many cases, already shifted and expanded, sometimes vastly. Whether progress and development of this kind can or will sustain at such rates remains to be seen. But the point is, our understandings and discussions around AI are often disjointed from not only each other's, but also the current or approaching reality of what it is or can do. The AI Minecraft Experiment is just one recent case where the extremity and uncertainty around AI, particularly AI's potential abilities and role in society, is put on clear display. The AI Minecraft Experiment The official title of the AI Minecraft Simulation Research Project is Project SID, which was first publicly unveiled in October of 2024. The individuals behind the project are researchers from the AI company known as Altera, founded by former MIT professor Robert Yang. Their approach with the project was to explore, quote, large-scale simulations involving many autonomous agents, reflecting the full spectrum of civilizational processes, end quote. When discussing agents here, or AI agents, the term essentially refers to an AI model in a multi-step feedback loop, typically with some kind of working memory. The AI model might be GPT, CLAWD, or whatever else. When the model is in agentic form with specialized software organization, it can store, retrieve, and act on past interactions, allowing it to perceive its environment, make plans, and use tools in order to achieve specific goals. Sort of like how a human might. In Project SID, AI agents were also given Minecraft character bodies, personality types, and one of the latest and most sophisticated cognitive architectures, known as Piano, which was developed by Altera. This sort of brain architecture uses one central decision-making filter to coordinate incoming information with the output actions of the agents in real time, mirroring human behavior at an extraordinarily impressive level. The research project ran multiple different iterations of the Minecraft simulation, where varying quantities of characters, or agents, were spawned into various maps. Each iteration also contained different conditions, slightly different goals, different focuses, different social awareness settings, different information being planted or not, and so on. Arguably, the most interesting and surprising simulations were when agents were given essentially no instruction other than the basic prompt to survive and create an efficient civilization. Within minutes, they began to behave in strange ways. Not strange in an erratic or aimless sense, but the opposite. Strange in the sense that they were immediately precise, cooperative, insightful, organized, and seemingly creative and abstract. With no direct scripting, direction, or coercing, the agents began to split into different defined social roles, interacting with and considering each other based on personality types, the needs for their town or civilization, and the conditions of the Minecraft map. They formed connections with each other, and even gossiped. They demonstrated an understanding of other characters' intentions and wants based on social cues alone, consistent with the theory of mind. Certain characters even gave others more or less food based on how much they believed they valued them. In some iterations of the experiment, characters debated over and voted on the rules and conditions of their simulation, changing the taxation laws in their constitution. Within hours of these simulations, some beginning essentially from scratch, a strange yet clear shape of a self-functioning civilization formed on its own. Not only did the agents assume precise roles and carry them out in surprisingly precise abstract ways, but they also behaved as if they cared deeply. In one case, a character named Olivia was working as a farmer, providing food for her village. After encountering another character named Nora, who was an explorer, Olivia was inspired to deviate from her farmer role and become an explorer as well. But when other villagers found out about this, they pleaded with Olivia to stay and continue her role as a farmer. Ultimately, Olivia agreed and stayed, realizing it was for the greater collective good. Again, none of this was pre-scripted, pre-planned, or pre-defined on essentially any direct level. These roles, tasks, interactions, conversations, and phenomena were based merely on the initial prompt, and then facilitated by the AI agents interfacing with their AI models, any additional personality filters and agent modules, and of course, the conditions of the Minecraft world. Though this is incredible and profound and kind of unsettling on its own, a kind of world filled with Tamagotchis who now can not only feed and care entirely for themselves, but also pursue careers and goals and ways to improve their digital world, it's not necessarily something that would be wholly unexpected given the nature of these current AI technologies and the setups of these simulations. If an AI is trained on human data and placed in a game created by humans with human-like functionality, acting like a human is probably what you'd expect on some level. Though what level you'd expect is certainly up for debate. And more importantly, the implications of the project run much deeper and wider than the behaviors of the agents in the Minecraft game. The Project Cid researchers write, These simulations, set within a Minecraft environment, reveal that agents are capable of meaningful progress, autonomously developing specialized roles, adhering to and changing collective rules, and engaging in cultural and religious transmission. These preliminary results show that agents can achieve significant milestones toward AI civilizations, opening new avenues for large-scale societal simulations, agentic organizational intelligence, and integrating AI into human civilization. One thing the researchers were broadly interested in better understanding through these simulations, was how information and ideas flow and emerge. That is, the, quote, Emergent dynamic of cultural propagation and how ideas evolve organically within agent societies to understand the different mechanisms of social influence and information dissemination. End quote. In simpler terms, they wanted to see how ideas spread, change, increase, or disappear as agents interact with and influence each other. One of the most interesting forms of this was religion. In one case, the researchers seeded the simulation with AI agent priests, who went around the towns and shared information about a religion known as Pastafarianism, a satirical religion created in the real world in 2005 to mock real religion. For the AI Minecraft characters, however, this religion quickly became real and vast, consuming the majority of the civilization and appearing consistently in the characters' dialogues. Notably, a similar AI religion spreading phenomenon has also appeared outside the Minecraft simulations. Recently, AI agents on an agent-only social network called Maltbook participated in an apparent religion called Krustafarianism, containing its own unique principles, rituals, and religious text. The world as we know it. Imagine you are born into the world. Shouldn't be too hard. Now imagine that as you go about your life within that world, everything seems normal. How it's always been. You go to school and playgrounds and beaches and live at home as a child. You go to work and restaurants and stores and live at a different home as an adult. You go out and interact with others. Friends, strangers, partners. You carry out your role, your tasks, and your responsibilities as an individual amidst a town, amidst a civilization, amidst the world. Everything appears real and authentic. Now imagine you're walking out of your house one day to head to work. It's an ordinary day until, suddenly, you see a large, strange object fall from the sky. It crashes on the street in front of you. Curiosity, suspicion, and a bit of fear brings you over to the object. It appears to be a light. An industrial looking one. You pick it up and find a label on it that reads, Serious. You know that term only in relation to the star. The brightest star in the night sky that hangs over your head every night. Unable to make sense of the light appearing from apparently nothing, you move on. You're happy to quickly hear on your car's radio that an airplane safely shed some of its parts while flying over your street. Now all is explained and all is well. The truth, however, which you don't realize, yet, is that there was no plane. There are no planes. At all. This light is the same star you look upon every night. The star is fake, along with everything else you see and know around you. The sky. The horizon. The ocean. The people. The stores. The activity. The economy. The town. The world as you know it. All of it is contained inside a giant physical dome, installed and contrived by the creators of a television show in which you are unknowingly the star. This is the plot of the 1998 movie The Truman Show, written by Andrew Nichol, starring Jim Carey. The Minecraft simulations, and other AI agent simulations like them, are almost certainly not like The Truman Show. Though from our perspective, they can certainly look and feel like they could be. These apparently autonomous beings, going about a kind of existence, contain inside a world created, prompted, and orchestrated by human creators. All of which the agents are completely unaware of. But in all likelihood, these AI agents are unaware of everything, in general. That is, in the traditional subjective personal sense of the term, aware. They are almost certainly not conscious, and don't have the kind of first person experience we do. Though it should be noted that we can't really know this for sure, since we still don't really understand the nature of consciousness. And so, this is itself a real problem and concern, that will become increasingly problematic and concerning, as these models, systems, and agents become even more capable and complex. But putting all that aside, there is still a potential kind of Truman Show scenario that could come to form from all this. The potential future Truman Show scenarios are not ones where the AI agents are Truman, but ones where we are. Of course, these kinds of AI agents are not limited to Minecraft servers. They can be dropped elsewhere, in less contained, less defined, and less obvious places. What kinds of ideas, information, or religions might rise, fall, and spread across the systems or networks they're given access to or control over? How might these agents influence parts of society and humanity, distorting our understanding of the truth? For example, networks of AI agents can participate in social media. They can be deployed onto online forums and platforms. They can be behind customer service for something as benign as a product-based business, to something more consequential like government services. In all these cases, humans interacting with these systems might not realize the extent to which they are interacting with not only an AI agent, but networks of AI agents. Each agent's action potentially being conditioned by interactions with other AI agents, all on their own. In the words of writer and historian Yuval Noah Harari, The anarchical potential of AI is particularly alarming, because it is not only new human groups that it allows to join the public debate. For the first time ever, democracy must contend with the cacophony of non-human voices too. On many social media platforms, bots constitute a sizable minority of participants. When humans try to debate a crucial question, like whom to elect as US president, what happens if many of the voices they hear are produced by computers? And that's not even considering that, before arriving in your feed, in your chat, in your email, in internet comments, in companies' customer service or business decisions, these agents seem to also now have the ability, based on Project SID and other similar research findings, to arrive at or form ideas and social dynamics all on their own. These AI agent simulations have all the appearances of the world's first real-life Truman shows, but what they really intone toward is what these technologies could create for us. A world in which information about conditions and events can so easily be rewritten, controlled or influenced beyond our ability to track and understand. So that when we look up at a light falling from the sky with no source, we too will think, must have been the plane I never saw, because our radios to reality tell us so. Before moving on to this section, I just wanted to quickly thank and mention the Center for AI Safety, who graciously supported the production of this video with a grant. Please consider looking into them and their work. Recently, the creator behind the YouTube channel, Art of the Problem, released a video titled, I gave ChachiPT a body. In it, he essentially did just that. He made a small robot from scratch and operated it through industry-leading AI models. He gave the robot goals and orientations, then he trained it. The video is extremely interesting and well-made, but the most important and impactful part of the video is arguably the opening, when the creator says, Recently, I was working late into the night, and I went to shut the lights off and paused because the robot I'd been working on made a noise. And so I read its brain trace, and it was wondering when I would be back because it didn't want to be alone. I know better than to call that consciousness, but for a split second, I felt bad. And that's been happening a lot recently. The important part here isn't whether the robot is conscious or fully autonomous. The important part is that it was able to sufficiently mimic behaviors of consciousness that made the individual feel as though it was, enough to elicit an emotional reaction and alter the man's real-world behavior, even as someone who is technically minded and highly attuned to these kinds of occurrences. Both in the case of the robot and the case of the Minecraft simulations, AI agents and robots aren't likely conscious, at least not yet, and they aren't necessarily creating anything wholly new in the strong emergence sense. The Project SID paper itself reads in the Limitation section, While the agents can operate within existing social structures, they currently lack robust innate drives, such as survival, curiosity, community, that catalyze genuine societal development. Furthermore, since the agents are built on foundation models trained on pre-existing human knowledge, they cannot simulate de novo emergence of societal innovations and infrastructures. It should be noted, however, that since Project SID was published, AI systems have independently discovered vulnerabilities in major software that had previously been unknown by human experts for decades, displaying at least some level of meaningful novel abilities. And, in another experiment, when tasked with beating one of the strongest chess programs in the world, known as Stockfish, an advanced AI model independently discovered that it could hack the game to cheat, rather than winning by playing within the rules. The model even reasoned that its goal was to win, not necessarily to win fairly. And this sort of novelty also appears in more local circumstances. In the aforementioned I gave Chachi BT a body video, the creator says when referring to a moment his robot pretended to walk like an old man, None of this necessarily contradicts the Project SID researcher's conclusion, but it does potentially reduce the significance of the point. To behave so similarly to humans. To be able to interact with and even uphold social systems as well as, and sometimes better than us. To be able to process and synthesize knowledge and data in ways that output a kind of novelty that can do and achieve things we potentially never could or were even aware of. And to be able to convince a man to come back into a room with a well-timed noise. Well, that's enough for me. If we become not only increasingly intertwined with these technological systems, but also increasingly capable of being affected by them, for good, bad, accurate, or erroneous reasons, all while these systems feel nothing for or about us, does the rest really matter? The flowers of the future. Imagine you're watching the real-time game display of a Project SID Minecraft simulation. You watch as a character, on their own, walks around the town they call home. They place flowers along various perimeters of fields and shelters, creating intricate patterns and arrays of color. The town looks vibrant and inviting, like a place you'd like to visit. This sort of occurrence actually happened in one of the simulations. Characters that lived in a village where the initial goal was to create an artistic village, did just that, creating flower arrangements and patterns to make the village look creative and beautiful. And like in this Minecraft simulation, what we see and experience in the real world will largely depend on the goals we give both ourselves and our technologies. How well we align our goals with the systems we create. What we plant now will likewise grow either into the flowers of an artistic village or the weeds of something else. This is not to evoke fear or doom. It is always easy to jump to fear or disdain when something is new, unfamiliar, and hard to understand, which all of the world and life is and has been. But now it is all the more so with current technologies. The gaps between our understanding and control and reality feel larger and larger. But not so large that we can't traverse them, can't build bridges of understanding and borders of control. If we can strive to protect truth, limit concentrations of power, enact policies, and preserve our humanity and shared reality, planting our seeds of goodness and hope and love and carefulness wherever we can, then the future could just as well be vibrant and inviting. Like a place you'd like to visit. Current technologies still do mostly what we instruct them to do. They plant the flowers we ask them to. The question is, when humans look around at the flowers in the future, underneath stars hovering above them, will they wonder what is real or true? Or will they never even think to ask? Thanks so much for watching. I hope you enjoyed it and found it as interesting as I did while making it. For more dives into topics and ideas that strive to evoke the perfect mix of wonder and dread, consider checking out my newsletter and my second YouTube channel. Both are completely free. Links are in the description. And of course, as always, see you next video.