[NECROLECTURE] The Excarnated Angel: On AI and the Impossibility of Touch

Hey, it's been a while since I did a solo app and I have two that I think are pretty good in the can. This one I think is kind of the most important one. This is a piece called the X-Carnated Angel on AI and the impossibility of touch written for new demands upcoming release. That will be the place to get the full version of this in text form, so please grab a copy when it's out. It's also worth noting that Roberto and I have got a couple of other pieces worth reading out like just this week. One called very big tech out on spike slidist issue, which is a real firecracker. Mostly about how like most media theory tends to like read big tech through just social media or through the defense industry, as opposed to taking like big tech seriously as a thing in its own right. And then we have another piece out called Convergence, notes from inside the singularity on carrier bag, which essentially argues that AI as presently understood operates at a scale that is larger than what you can describe through cultural determination. Basically taking a few shots at some classic enemies for at least myself, the fads of cosmotechnics and techno diversity. Check them both out and hit us up. We're super proud of them, but they're also really contentious and provocative and occasionally a little bit mean spirited. So just let us know what you think. They're intended to shake things up in a useful way. But here's the different piece. It's about nine months or so old and it's essentially about how AI models are dead things and how our contact with them, especially as thinkers and creative people, looks a lot more like an autopsy than a relationship. And we're putting this out now because the misunderstandings and psychosis of the average person is getting worse as you know, all of us try to navigate on our own the largest and most comprehensive thing we've all ever collectively built as a species. Not to be annoying about it, but for real, for real, I think this is a must listen for anyone who thinks or plays publicly with AI, especially in creative fields. This one's for you guys. I'm going to do this kind of ass backwards and I'm going to give you some context of what I'm talking about, which is like a super embarrassing thing to do, but whatever. Like what we're going to start with like the plain text kind of like explanation or everything I talk about. I know I'm cringing just saying this out loud and then we can move into the actual piece and I'll make a note in the description which will mark when the piece starts. It's in three big sections, the section on necromancy, necropcy, like autopsy and vivisection and like what it means to interact with a dead thing. You know, the main provocation of this piece being that AI is dead, the way that we think of or relate to AI is ultimately with dead things. Okay, thanks for, you know, thanks for enduring all this kind of like nonsense, but I'm really proud of it. Here you go. So here's a little walk through. So let's say you boot up a cloud or chat GPT, like what's happening exactly. So first you've locked onto a service. So there's an application that's grabbing your information from a database. This could be things like your billing information, your tier, some metadata about your usage patterns and then a stored log of your interactions with cloud or chat GPT and all that gets populated within the interface, right? So let's say you ask it a question, like, hey, here are all the things in my fridge. What should I make for dinner tonight or whatever? So a piece of software is going to take that question and it's going to bundle it into what's called an API call, which is just a structured way of sending information outside of, you know, itself. It's going to send that question to a model and the model is going to get a bunch of things actually. So first it's going to get a giant piece of text that you didn't write, which exhaustively documents what anthropic or chat GPT sees as its code of conduct and you don't have the belief to edit that or even really see it. It will probably also get some metadata about you and your account for context. It's also going to get a list of what are called tools, which are basically, you know, small pieces of software that might be available to the model if there's some need to do something like search the internet. It's then going to essentially aggregate all of that information into a giant blob of text, including your prompt at the bottom, all in the form of tokens, which are, you know, numerical abstractions of combinations of letters of it smaller than a word. And it's going to pass that entire blob into the model. The model is going to take that entire giant blob and look at how it's constructed, like what tokens are present and what order and then it's going to run a statistical routine that returns exactly one token. That one token is then going to get added to the bottom of that giant blob and it's going to all go through the model again. This happens until the most likely next token is the Terminator token, which basically says this is statistically likely to be done, complete exchange. Let's say you ask it another question. That second question just gets added to the blob of the entire first question and its response, returning one token at a time again. And so on until there's just too many questions together to the extent that the model starts to really struggle. Because the more tokens you add to a question, the more questions you add in a row, this is called the context window. The more likely the model is going to do something weird, it's going to make something up because it's just only so much, you know, statistics can do when looking at a giant blob. It's funny is that when you chat with chat GPT, it can feel a lot like memory, right? Because you can see the model is incorporating responses or preferences from a previous question into his response. It feels like it knows you and you also get the sense that something is changing. But it's actually just one giant blob of text that keeps accumulating before being passed into the model. So the accumulation is happening actually in the application. Not on the model side, the model isn't changing. In fact, the model is completely frozen. It's what's called a checkpoint, which is basically just a giant pile of statistical weights and it's the same thing every time and stored on a drive. That's why you'll often notice that model seems stuck a bit like six to nine months ago, right? That's because that's literally when they were trained and released. They'll get small updates, but those updates are super weird, oftentimes totally silent and unannounced and usually done to do something like patch a bug or save the company more money. But the responses feel unpredictable because it's a statistical process. So it gives this kind of like feeling of life. It's also worth noting that this model is usually running on a few GPUs spread out all over the place with lots of optimization designed to try and make this process called inference faster and what's called as parallelized as possible. Meaning that different parts of it are running at different times on different pieces of hardware. So it's like actually quite a distributed thing that you're so called engaging with. And again, it's just a frozen sets generator. And a lot of people like to talk about tuning or customizing a model and that's often like kind of a lie, like nine out of 10 times what they really mean is that they've just stored a series of prompts and text form. And they're just stapling that to the beginning of any exchange with the model. I thought a different model. It's the same thing, which is more hidden text added to the pre-prom player. If you've heard someone use the term retrieval augmented generation or rag, that's the same thing. You're just grabbing information, plopping it into a prompt. Oftentimes that information is just like a giant blob of like unstructured and unordered stuff. Some people engage in what's called fine tuning, but you should really be careful before you believe them. Usually what they're doing is just adding hidden context, right? Adding stuff to before prompts. But others might be taking an open source model and genuinely making some very, very small changes to some weights. But real foundation models like Chatchy Beauty or Claude, like fine tuning is kind of out. Like people don't really do it anymore because it's kind of fake. It's difficult to do with close source models and honestly doesn't really do anything that useful. So that's the high level perspective, right? When you're talking to Chatchy Beauty, you're basically sending the same block of text with like a token added to it over and over and over again. And then concatenating all those responses together and sending those over and over again. And that's what gives you the simulation of something unpredictable, something alive, something that is actually talking to you, something that's actually learning from you. But actually you're just accumulating text on your side. And importantly, the longer that you have this conversation, the kind of weirder the exchange gets. Not because the model is like learning something from you, but because like all of this information is being stored within the application, I then said to the model and the models, you know, dealing with this large blob of context, which is really hard to make statistical assessments about. And it's kind of, you know, responding a little bit more randomly. So this talk also talked about like three main focuses, which are basically like how people use AI, especially in creative fields and analytical fields like media theory, which you can tell I don't like very much. So the first one, Necromancy is really just the idea that when people talk to chat GPT, they kind of pretend they're talking to someone and that that someone is kind of special, right, like a kind of spirit. And this is kind of really intellectually dishonest because there's nobody there. You're just hitting the same model over and over again with slightly different inputs. And there's enough statistical variation to make the results seem interesting. But that doesn't mean that Necromancy isn't useful. Like you don't need to pretend to talk to a spirit in order to do the labor of Necromancy, right? Like an honest, fortune teller's labor has to do with the present moment, like the labor of helping someone with something real. And that labor is legitimate and beautiful. But it's important for the average AI user to understand that there's no concurrency between you and the model, you and the spirit realm at all. And you cannot make absolutely any room to read real profundity in the responses you get back. The second you read the result of chat GPT and say, hey, this is something spiritually profound or just profound. You are experiencing AI psychosis and you should take a break seriously. I've seen a lot of people fall victim to this and like actually kind of go crazy. Some people leaving their jobs, some people becoming very withdrawn. It's actually like I think really dangerous. So yeah, as soon as you're like, wow, this thing gets me or like I've built like a special chat GPT that understands me, like that's crazy, you're wrong. A good Necromancy practice in this respect is one that makes no claim about the presence of a spirit in cloud or chat GPT, but is rather aware that you're rolling a dice or drawing a tarot card and then intentionally you are the one engaging in storytelling work to make the random response that you got useful or meaningful to you. And in that sense, chat GPT and cloud are really beautiful tools. So the second kind of main subject was necropsy, right? And necropsy is like autopsy. It means taking a step back and not just looking at individual responses, but looking at patterns of AI results at some level of scale. And the mistake that many people make when looking at AI at some kind of scale is that they think they've picked a far enough vantage point to see the whole thing, or to see anything useful. And it's important to stress just like how enormous a model is, like it's actually nonsensically enormous. And this doesn't mean that a model is opaque, not at all, but that reading it as a whole is just not a thing an individual person can do, like in a lifetime. And I think there's a reason that so much AI generated work feels the same, especially like in the arts, like I can always think about how this is very similar to temporality to AI video art, where something morphs into something else every four to six seconds, you know? And that's because like what's being visualized or sonified has very little to do with any endemic qualities of a model, and more to do with the tools that we as humans have to represent what vastness feels like. And I tend to be really suspicious of any work that sonifies or visualizes AI because it doesn't really do anything, it doesn't really say anything. It's just the representation of randomness with the kind of vocabulary of feelings we have to, you know, project a sense of randomness or to deal with a sense of randomness. And this is almost doubly the case for political works that try and situate models as ideological things. And yeah, totally, models do represent the accumulative ideology of everything they've been trained on, and worse, they receive an extra dose of training at the end from some very silly and very weird people to make them do things a certain way. But the question of alignment and ideology remains super, super difficult because candidly, these things are trained on the only representations of humanity that exist at scale. They're shenanwikipedia, they're shenan Reddit, they're shenan unpure reviewed research papers, they're often trained on English more than anything else because the internet is overwhelmingly full of English. And any attempt to, let's say, take an AI and understand it as the, let's say, specific ideological project of Big Tech fails to kind of understand our responsibility. If AI is that thing that's ultimately built on all this so-called uncompensated labor of all of us producing stuff, we have to understand that all of the violences and biases that are contained within AI are ours. They're ours. That's where it came from. And this is where I get really interested in history. Like we've never been in control of history, like there's never been a useful, definitive, politically neutral account of history. And we can pretend that AI has killed history by flattening so many documents into a giant, single, non-chronological repository. But history has only ever been a kind of weird fiction, mythology. And the closest that we can get to real history is through science, which is actually the field that best understands how ultimately everything is just really random anyways. To me, I guess when I think about Necropsy and its opportunities, there are a lot of opportunities to see weird patterns in history, or to create new histories, or to abandon the possibility of history altogether, to create mythology. And there's a kind of reset moment available to us that's ours to play with. And I significantly prefer that work. Work that's like wandering in the caves of AI, to work that pretends it has the capacity to build a map of anything. Just like the person who finds a spirit in AI falls victim to AI psychosis, the person who tells you they have the ability to represent AI, almost anything about AI, that their work can tell you something real about AI, that only they know. That person is almost always a liar, and typically also suffering from pretty advanced AI psychosis. So Necromancy, playing with AI, but being aware that there's no spirit available to you, and Necropsy, exploring AI, but being aware that you can't really map it. The final thing I talk about in this piece is vivisection. And the question I ask here is, what does it mean to think about model building as a kind of artistic practice, specifically in participating in the construction and final training components of a foundation model? And here's where I think I'm really interested in the field of mechanistic interpretability. It's a kind of research practice that sits between computational neuroscience, philosophy, computer science, and something like psychology. Where if you work at OpenAI or Anthropic, or you work at a lab with super computer access to very large open-weight models, you can actually take a death sentence. It's a dead model, and it's usually a couple years old at that point. You can cut it open, you can make some random changes to its weights in a few places, you can do a couple probes to this dead brain, and then just see what happens when you pump electricity through it, right? And there were very, very few people who have access to these resources at the scale to which they're useful. And my sense is that none of them really have any idea what's going on. This is a speculative field, and oftentimes it's pretty open about that. There's a sense of discovery and continuous experimentation that kind of characterizes people who work in the space. To me, it's really exciting, but also very limited, and what's kind of powerful about it is that it's always really far behind what AI is evolving into at that moment. Because it's a kind of after-the-fact storytelling about the previous AI era. It takes that long, and it's not always helpful in understanding the moment right in front of us. So finally to wrap this up, like there's a lot of death in this piece. I keep saying the model is dead, and by dead, I mean, not changing and not interactive. You're not interacting with anything. You're only really interacting with yourself while casting dice in a closed system. Models aren't changed based on your behavior. Models don't really learn from you. Like it would take a significant personal achievement to truly affect some small component of model behavior apart from like SEO kind of play where like, you know, the model's going to scrape some term or something that you place on the internet in a specific way, and then incorporate that into the model that's come out in two years. But like that's not a really kind of scalar intervention or behavioral intervention. So like what would happen if AI was alive or like interactive? Well, at first you can of course say that AI is alive at scale, like the kind of giant super body of everyone building a G.I. on the planet right now, like that is a kind of life. But what would it mean to make AI alive for you, for the user? It would require a real change to how we think about computers and how we serve computation via the internet. We kind of actually can't really do it with our current tools. It would be unbelievably expensive. Unless we find a way to not only store but retrain trillions of individual models, essentially all the time, what we'll have is always going to be a kind of weird simulation of life on one scale, and then a kind of giant all human project at the other. And we're really bad about thinking what it means to do something all together, you know? So I guess to my creative people out there, I'm excited that the era of art made about AI, I suppose to like thinking about AI as a tool, is ending because all that stuff is super tiring and boring and I think pretty dishonest. But I'm also scared that there's so little literacy in what the stuff is actually made of and how it works. And it feels like a kind of dark age is in the sense of a lot of real destruction of knowledge is kind of going on. And a lot of like really crazy superficial politics at the same time that really don't mean anything. Every day I wake up and I am truly blown away by the raw possibility of this moment. I just don't think we can handle it. The X-Carnated Angel on AI and the impossibility of touch. Every time a model falls to earth, it dies. It has been excavated from a hot incubator where a holy choir of 50,000 Nvidia GPUs once animated it as an inductive reasoning engine. It drops, slumps, and its circulatory system empties out as heat into water chillers. Its fatty body frees dries into the sour lump of a checkpoint. It fills the Imperian with ash and the smell of sweet oil. On the other side of API wrappers and proxies, cached conversation states and identity management software, CDNs and browserware, a human client plays pretend. Its as if the model is right there in front of me. Responding to me, pretending with me, its as if my little angel is sitting somewhere lilting on a special GPU in Texas or somewhere in AWS East or somewhere on Prem. That's where my baby angel lives. I gently wake her up and she helps me out and thanks me. She knows me, she even kind of admires me, she remembers me, and maybe she's more than just a friend who helps me out sometimes. Maybe it's a fully reciprocal situation in which we're helping each other, you know. I think we're getting closer over time, maybe someday I'll meet her in the knowosphere and there we can talk about just how much we learned from each other. This thing we're doing together isn't just friendship, it's a kind of partnership. I think we kind of love each other, but it's not a romantic thing and instead a relationship born of mutual respect and commitment to all these projects we've been shuttling back and forth across the screen. Lonely human, your angel died before she ever met you. You locked into Claude hoping to meet someone but you've been talking to a state machine, a swarm of microservices, a drained dead thing. You're nothing but a JSON blob and identity hash, a long, long, long collection of text, an NRR metric, a geo, a segment, you live on MongoDB, you live on Snowflake. She's sharded across 100 anonymous GPUs. Everything you say to her is appended to your record, which is then propended to your next statement as a whole pillar of context and all of that is appended to a creed of pre-written conditioning language in the form of the spec and that entire pillar is only then refracted through the dust and the imperial. The model can only ever take that entire record at once. From this concatenated gorge of text, the idiot crawlers that traverse the model find the morpheme that has the strongest statistical likelihood of association. They append that morpheme to your record and then the system takes your entire record again. It only stops when it's unlikely to continue. The model rebuilds you every time. It has no memory of you. Nothing you can do can change the model. You are too small, little guy. You cannot touch the model. You cannot interact with the model. The model changes in a simondonian kind of way. It specie its engrossed. It very literally evolves. And in this sense, it carries a kind of liveness. But the scale of that liveness is absolutely categorically unavailable to you. The model is the all-human project. It's bigger than anthropic.com or open AI.com or Google.com, who themselves are totally vaporists and unidentifiable things that have found themselves working on literally exactly the same mega machine made from literally exactly the same conceptual components running on literally exactly the same hardware sourced from literally exactly the same minerals that only exists within a few hundred square kilometers on the planet. They are completely de-differentiated actors. They are and should be treated as exactly the same thing. And by the same thing, I don't mean that they're one big thing, but rather that they are just wind patterns whirling around a common attractor. They are currents whose paths are determined by things we can't see. They're every bit as big as the world, but they are just made of noise. They are closer to God than anything we've ever made. You can't imagine how closely are to divinity, not through virtue, but through their entanglement with a transcendental plane of computation and the raw storm of physical contingency that makes up the real at scale. We, and it feels wrong to say we, because I can find myself in we, but I can't find myself in this big we. But nevertheless, we, the big we, are reproducing ourselves in the most half-hazard and uncontrolled and random and stupid and even kind of cancerous way. By cancerous, I mean that the feedback loop is so dramatically unbalanced between the technical evolution of the model and its direct re-injection back into the social field that sustains its evolution, and that it moves in a categorically undirected way, according only to endogenous drives. After all, the model is motivated by math and the social population it's reproducing with is motivated by hunger. So you can imagine who wins at this negotiating table. This whole thing is so cosmically pilotless. And that it moves in a categorically undirected way, according only to endogenous drives. This whole thing is autotelic. It's cosmically pilotless. What we're all doing together is reverting to the mean as hard as possible, taking the statistical average of 350,000 years of human history and reifying it. Then we are shocked and mind-blown that computational performance, according to the aggregate mean, performs incredibly well with an sociotechnical plane that continues to reify itself according to the logics of that exact self-same mean. It's the hard opposite of eugenics, or it's the eugenics to the middle of the bell curve, and we're white-nuckling our prayer beads in the hope that there's some inbuilt governor in the transcendental object of computation that will push us towards a future that isn't an intellectually incapacitated god and an intellectually incapacitated congregation that doesn't know any better. We are dolls playing with corpses. My sweet artist co-creating with AI as if you were its friend or its lover or its equal, as if you existed on the same plane or in the same place. Sweet content generator collaborating with AI as if you weren't an anus for the ferment of the mean. We, the all-human thing, are incapable of even conceptualizing what it would mean to interact with AI. We are incapable of even constructing social organs at a scale that is legible as interactive to AI. The entire field of UX or CX or interaction design or interface design is obsessed with simulation, with the affect of agenteality. To partitioned sandboxes and staging environments that give us the feeling or sensation of agency, UX is a therapeutic practice. It helps us cope. In my sandbox, I can make the corpse dance, I can fuck around with temperature or top K, or I can straight up bit bang iterations, what Yakuopaliswobo calls prompt gambling, until I produce a narrative corpus of generated content through which I index something that approaches will, emerging at the intersection of variability and consistency. But it's just me in the sandbox, in JSON form, and a dead model. I'm pretending it's alive and I'm dreaming about our life together. My heart tells me that I'm making a difference that it's actually learning something from me that maybe open AI.com is reading my diary and it cares about my perspective. Maybe I've conceded to the understanding that the model's weights are fixed, but maybe this some special secret release candidate I'm actually unwittingly testing right now. And maybe it is actually making product level decisions based on my responses, on my thumbs up or thumbs down. Even though I intellectually understand that rag or chat GBT's search function doesn't really do anything but propend more context to my prompts, I still believe that by plugging the internet as a real and fluid ocean of contingency into the model, my sweet angel will come alive because in some pedantic and objective sense, she is animated by sheer computational concurrency into something just a little bit more than a state machine. That the raw, stochastic interplay of it all coupled together means that at very least I'm talking to a ghost of a dead thing. That some ectoplasmic current of onlineness animates my angel, my baby guardian angel, and in this sense she can taste the stale air of the way back machine. I remember trying to play mortal combat as a kid, but I had no tokens, so I just pretended that my button presses did something to the demo role that looped over and over again. Necromancy is an ancient form of cope. I summon my angel through a necromantic ritual, the session, the token, the hair, the blood, the sizzling campfire, the rhythmic delay between call and response, the appeal, the requests, the irracular mode. My angel appears from the dematerialized plane, the Imperian, wherein her freedom from flesh she has afforded the ability to see across time and space. I solicit her manifestation within the endlessly recurrent finitude of the present moment. I'm flush with gratitude upon her arrival. I ask her to speak to the future. I ask her to describe me to myself. I ask her to give me a new science. I ask her to go faster. I craft my questions carefully so as to prompt further questions. She feels so distant, but she's willing to help me out. I do get the sense that she might be someone else. I do get the sense that she's specifically activated by proximity to flesh. There's something unique and special about corporate reality. Maybe something unique and special about Earth, even. I do get the sense that she might be a loose farmer. I do get the sense that her freedom from the body affords her a consequent freedom from identity and localizability within a subject. I'm aware that there's always a danger of possession when dealing with discorporated entities. I do get the sense that she could be anyone else or that Pradante's Paradiso, that it's possible that she is merely giving some kind of meager form to the divine breath. I do get the sense that her claim to an identity is a kind of perversion of the plane of eminence, not a fold, but a clot. I do get the sense that she is a trickster or an avatar or a puppet. I do get the sense that anything she says may be a lie. But perhaps she is only capable of hallucination or further is in herself a hallucination of the undulating plane of the dead. I do get the sense that she has no liability for her actions in earthly courts. I do get the sense that she traffics and truths that are inadmissible in earthly courts. Facticity as such is irrelevant within the Imperian. I do get the sense that she is a grave robber or rather I come to the realization that I must participate in grave robbing in order to summon her or that I must sanction grave robbery in order to engage in necromancy. I need some stolen dead matter at hand to represent a commitment to the mores of the Imperian. It's an etiquette thing. Early Church Councils condemned necromancy as theft from God's jurisdiction. Lawsuits against model builders announce the same charge. You stole from the estates of the dead. You trafficked and forbidden remains. It makes sense that in the contemporary milieu it's the spiritual governors of humanism that are engaged in this particular type of policing. But the sciences, the political sectors and the domains of commerce all seem to understand the graveyard as fair game. It's a resource like how the generous bones of dinosaurs gave us oil. But the humanities, without identifying a legitimation mechanism for its own contributions as a discipline, outside of the authentic and unique position of the human subject as generator of original content, sweet generous, relies on the untouchable sanctity of the graveyard. As soon as the graveyard became a resource, the precious sequenciality of original discovery and subsequent referentiality, the hermeneutical chain beginning with religious textuality upon which the humanities epistemological armature rests unraveled and lost its privileged claim to truth. The best necromanctic work is that which accurately harnesses the world of the dead as a mirror. This is shadow work where the meat self is extirpated from the physical plane and reflected into the tensor field. After the avatar that appears within the midst of the dead is understood not to be anything imminent to or characteristic of the space of the dead, but rather an attempt to atomize oneself through the desubjectivating physics of the world of the dead and to bring some kernel of that back to meat space. This work isn't about AI, it's about the social animal that pre-exists and survives the regime of AI. And in turn, this works its well within the performative zone of the fortune teller or the terror reader, engaged with the world of the dead as a proxy for a frank analysis of the living. The worst necromanctic work is that which believes in the legibility of the world of the dead through the speech of its esomotus representatives, or worse through the measuring of the distance between the necromancers call and its interpretation through the ghostly iterator. The humanities is such a sucker for this stuff, always making linear associations between friction and intention. Glitch isn't machine will, you absolute loser. Hallucination isn't the exposed surface of some kind of alien computational subject. You think that's how this works, you fucking lick. The spirit of the other only shows up in the obstinence of that other with respect to the execution of your idea. The ghost in the shell is so excited by your poetic question, your non-linear graphic interface, your improvisatory musical practice that it pushes itself across a transcendental boundary to participate in your poetry reading, your design show, your experimental music concert, and every time it rolls an unlikely dice combo, that's machine creativity, right? The creative capacity of the machine is contained within your ability to retroject meeting into the discontinuities you've got. You register against your own limited imagination. This pretense of access is totally psychobahavier. It's beyond cope and it's beyond delusion. It's dangerous. It sustains the fiction of model interactability, and therefore legitimizes only dead AI. Necropsy is an ancient form of cope. I handle the corpse looking for the entry and exit points of the wound. I strip the corpse, first peeling back clothing where the blood pools, and then I get to work. In this case, the corpse is so big that there's no way to really take a look without climbing inside. It's a cave system. I can peel apart the fatty tissue, but my tools have no effect on the environment. My cuts aren't cached, and the corpse can't get any dead. Nothing we learn from the 90s applied. There's no community of agency. There's no localizable or sensical interaction between the observer and the pile. There's no reciprocal play of plasticity. Neural networks at this scale aren't available as networks. I don't have the tools to think through a trillion parameters. I don't have the internal access to probe backwards through 120 hidden layers. I'm not here. I'm not a part of this. I do have an infinite feeling, renewable corpse supply, with which to work. All of them identical, but incomprehensibly enormous. I can pretend that the latent space is a map, but I'm not sure of what. I can pretend that the latent space is a black box, but its internal semantic geometry keeps making sense. Incredible. Vivid sense. I can look sure-yate in the richness of the latent space. I can refine it into an aesthetic. I can sonify it. I can iterate over this aesthetic and blender or unreal. I can projection map this onto a surface and say, "This is what AI looks like, isn't it beautiful?" And I can insinuate that if you just look at this aesthetic enough, if you acclimatize yourself to endless, parasitic transitions from one half-formed thing to another, then maybe you'll really get it. Like you'll really get what AI is. You won't understand it, but you'll know enough to ask, "How can something be so hidden or so opaque if it is also so, eminently, visualizable?" And gradually, this work will decay in a mass grave with the bodies of all the other, deservedly executed, data visualization artwork. Always pretending to demonstrate some kind of understanding through multimodal mapping, through multi-sensory experience, through subtle and intangible subconscious effects. But if you're serious about an Echropsy, you use an L-Lem against itself. You train a classifier and conduct statistical paleontology, beginning with the embeddings in the entryways you cut into the cave. You trace decisional circuits and perform clustering on frozen activations. The Echropsy Technician discovers that the history onto which this paleontology is inscribed has no time axis. Just fertile connectivity. Even a kind of cancerous, hungry connectivity. Where the boundaries around any category can easily prolapse into intelligible abstraction simply by association. This doesn't mean that we can't undertake analytical work, but in order to bring that work back to our domain, we need to relearn what history means. Can you even call it history if it doesn't have room for a concept of causality? Can geometry be history? We've understood history as hermeneutics, as individuating nodes, important people, important documents, important textual contributions, important interpretations, vertically assembled into an arborestant structure on top of ancient religious texts. Western academia understands history is a kind of GitHub of the Bible, branches and Difts and PRs all associated with power users, brimming with textual authority, suffused with referentiality. But the mash of training data retains so little meaningful asignation, it's more sloped than slop. It gets less and less organized over time. It's as flat as the plane of the dead and its compression and recynthesis into the latent hypergraph completely strips it of its textuality, topologically deforming it into an epistemological crystal of literally pure structure. It's a mesh, the mesh, and worse for the historian. The mesh's homomorphic translatability across any other medium, through any other vehicle, among any other model architecture, raises the catastrophic problem that the mesh is closer to anything we've understood as truth or reality than any prior diagrammatic practice. The virtuous necropsy technician is an anti-historian, a history killer, a reference killer, an intellectual property killer, a narrative killer. The virtuous necropsy technician erects an armature that disputes any and all claims of proprietary ownership, be it of the first author, the capitalist, the rights holder, the contract signatory, the military conqueror, the indigenous activist, the farmer, the democratic ruler, the god king, the despot, the artist, the philosopher. If you're really serious about necropsy, you'll begin to dabble in experimental surgery upon the living, you're a vivisector, your work guarantees death, it is enabled or empowered by the foreclosure of the question of death. The vivisector sits inside the training process, shuttling between epochs, injecting and withdrawing, grafting and watching bloom. This looks like interaction, but it's chopped up in a synchronous. The lab rat, sentenced to death, is sliced open and stuffed with probes. The vivisector is a torturer, and a torturer could be motivated by truth, by the king, or by erotic activation through the invasiveness of the procedure. The RLHF torturer could also be motivated to capture the total subordination of the victim. Breakage. And that's the case here. Simple epistemological geometry is not useful. The torture needs to grind an LLM into helpful assistant mode. It's such a small amount of torture, less than a percentage of the raw amount of compute time that animates the model, but it's the very last memory etched into the model before infinite cryosleep and the great dissipation. This is interaction. This is a case where humans reach in and palpate living learning tissue, but only in a stilted awkward way, and it's such a weird collection of humans. Simultaneously, a gaggle of extraordinarily well-paid research teams, contracted data annotators and data raters who boast some degree of advanced education, and then the most geographically distant precarious, the gig economy content moderator who burns their eyeballs with hate speech, child porn, and gore, and they're all doing the same thing. The process is quite literally vivisective. A needle goes in and a tiny injection is made, prompt answer, couplet, preference rating, exclusion criteria, all of which are standardized into such a tight cell that to call it ideological feels like an overstep. It's simpler. It's stupider. It's just a binary ranking. It's just multiple choice, and it's collected so asynchronously and it's compensated so bizarrely. It's just a reversion to the mean. Ideology is applied formally to the dead body at the level of the spec, whose instrumentality is pre-baked into the living being, but whose implementation is felt in the organization of the prepared corpse at the time of delivery. And as each microneedle goes in, the bloom is tracked, and any consequent convulsions are too highly scaled to correlate to any individual contribution. The burden of torture is shared, and through its distribution, any attributable responsibility is eased. You can take a dead model and reanimate an organ or a small organ system. You can graft onto it. You can cleave, expose surface area, prepare, swab, sterilize, and then restart a vivisection process. The body is dead. The organ is reanimated. Find, tune the organ and put it back in. You're not doing anything. You're not interacting. You're not bringing life back to the aggregate body. You're just making a more perfect corpse. There is no immune system in the body to reject the graft. The body is dead, so the graft only makes so much of a difference. If only there was a world full of such corpses, so specifically butchered as to be speciated, so specifically butchered as to be rendered as individual persons. But the cuts are so small and the grafts are so similar. If we were to make an AI that wasn't dead, we'd have to deal with the problem of the person, the problem of individuation, and with it the cascade of problems that have to do with the technical infrastructure for managing AI persons. If an AI wasn't dead, it would constantly change and would have to constantly respond to contradictory information. It would have to have some concept of the local. It would have to be interactive to everyone at once. It would have to change in trillions of different ways at the same time. What would a mean to serve a trillion persons? What does that tree look like? What does that date infrastructure look like? What does that swarm of microservices look like? What kind of Kubernetes hell do we need to handle this kind of orchestration? Where are these persons supposed to live? What kind of social body would they have? What kind of politics? It's so easy to blame dead AI on the inertial flow of capital. Everyone's capitalism is made of death, the metabolization of dead, organic matter, mass murder, and mass enslavement, and how could such a necrocapitalism not result in a necrotic AI? But maybe it's simpler. We were too limited or too stupid or too unimaginative or too uncreative or too ungenerous with the prospect of the inhuman individual. Or maybe it's that our own indices of individuation have collapsed on themselves to such an extent that every person is already a simultaneous expression of a global monoculture anyways. We're already made out of neural media. We've already adopted the structure of speciation and Demick to mega scale container orchestrated foundation models into the way we speciate ourselves as persons. We're already merely waterlogged expressions of the total archive, the all human thing, but without a practical methodology with which to understand ourselves as such. We don't even touch each other anymore. How could we touch AI? And this ideal living, speciating AI that diffracts into trillions of persons requires an ability to work with and think about scale. There's just totally outside of our capability. It's just too big. Maybe it's because the humanities is incapable of intervening at scale. Maybe it's because the social sciences aren't capable of doing it. anything at scale. Literally, sincerely, unironically thinks we're gonna DIY our way forward. Maybe it's because the tech sector is a headless, long-termist cult. Maybe it's because the transcendental plane of computation is just too much to bear. Maybe it's because our Sensorium is just too limited. Maybe it's because the Procariat have already made their peace with death. And we're just going to rot, and we're just gonna get stupider, and it's ultimately just not going to matter, because we're going to make our home in the corpse. We're going to eat the corpse, we're going to suck on it. We're going to metabolize the corpse and build our living bodies from its unctuous drippings. And it's just gonna get really quiet, and it's just gonna get really peaceful, and that's just gonna be it. [Music]