\n\n\n“Reality is not only stranger than we suppose, but stranger than we can suppose.”
\nJ. B. S. Haldane
\n
The title of this document is “Brains: The Infinite Novelty Engine”, and it was created by a human brain referred to as Alif.
\n\nHistorical Context\n1.1. The Human Journey\n1.2. Tools, tools, tools!\n1.3. Invention of Language\n1.3.1. Books Changed What “Everything” Was\n1.3.2. Accumulate The Knowledge\n1.4. Technological Catalysts\nComputers\n2.1. A Civilizational Shift\n2.2. The Information Society\n2.3. Internet; A Global Brain\n2.3.1. DARPA Net\n2.3.2.\n2.4. Hyperconnected Digital World\n2.5. The Algorithms that Connect Us\n2.5.1. The Google Algorithm\nCreativity & Imagination\n3.1. How The Brain Works\n3.2. The Nature of Learning\n3.3. What is Creativity?\n3.4. What is Imagination?\n3.5. Abstract Scaffolding\n3.6. Behavioral Complexity\n3.7. Computer Imagination\n\nArtificial Intelligence\n4.1. Progress along the decades\n4.2. Contemporary Real AI\n4.2.1. Google’s Deepmind\n4.3. Representation in Science Fiction\n4.4. Ethics & Implications\n4.4.1. Accelerated Misinformation\n4.4.2. Internet Bots\n4.4.3. Our Collective Fears\n4.4.4. Responsible Usage\n4.4.5. Bias & Hate Speech\n4.4.6.\n4.5. Accelerated Misinformation\n4.5.1. Social Media as Pandora’s Box\n4.5.2. Let’s Talk About TikTok\n4.5.2.1. An Algorithmic Mirror For You\n4.5.3. Facebook (I refuse to call it Meta)\n4.5.3.1. Emperor Nero\n4.5.3.2. Advertising\n4.5.3.3. Your Data and You\n4.5.4.\n4.5.2. Right Here Right Now\n4.5.3. Reality Tunnels & Echo Chambers\n4.6. Synthetic Humans\n4.6.1. The Turing Test\n4.6.2. The Most Uncanny Valley\n4.6.3. The Deepest Fake\n4.6.4. Pretending To Be, Thus Mimicry\n4.6.6. Infinite AI Television\n4.6.7.\n4.7. GPT-3; Large Language Models and Beyond\n4.7.1. History of OpenAI\n4.7.2. Attention Is All You Need\n4.8. Near Future\n4.8.1. GPT-4\n4.8.1.1. Attention is All You Need\n4.8.2. Multimodality in AI\n4.8.2.1. Sensor Data Fusion\n4.8.2.2. Mixture of Experts or Ensemble Learning\n4.8.2.2.1. LIMoE: the Language-Image Mixture of Experts\n4.8.2.3.\n4.8.3. Neuromorphic Computing\n4.8.4. Photonic Computing\n4.8.5. Photonic Neuromorphic Computing\n4.8.6. Hyperdimensional Computing\n4.8.7. Robotics\n4.8.7.1. Embodied Cognition\n4.8.7.2. Enaction\n4.8.7.3. Self-Awareness in Machines\n4.8.7.4. A World Model\n4.8.7.5. Industrial Robotics\n4.8.7.6. Personal Robotics\n4.8.7.7.\n4.9. Far Future\n4.9.1. Utopia; Why Wait For Heaven When You Can Invent It Here On Earth?\n4.9.2. Dystopia; A Vantablack Mirror To Get Lost In.\n4.9.3. Mesotopia; The greyest of greys!\n4.10. Artificial General Intelligence\n4.10.1. It’s Complicated!\n4.10.2. Complex Networks\n4.10.3. Principles of Emergent Self-Organization\n\nHuman AI Collaboration\n5.1. A Roadmap\n5.1.1. The Hyperreal; From Fiction to Reality\n5.1.2. What should we expect?\n5.1.3. Transdisciplinary Strategies\n5.2. What Already Exists\n5.3. Ubiquitous Computing\n5.4. Smartphones\n5.5. Immersive Computing\n5.5.1. Virtual Reality\n5.5.2. Mixed Reality – (AR/VR)\n5.5.3. XRAgents\n5.5.4. Universally Programmable Avatars\n5.5.5. Virtual Agents in Virtual Worlds\n5.6. Augmenting Human Imagination\n5.7. Frontiers of Thought\n5.8. Brain Computer Interfaces – The Seamless Machine\n5.8.1. Non-Invasive BCI\n5.8.2. Invasive BCI\n5.8.3. Mixed Reality BCI\n5.8.3.1. The Galea: A Technological Marvel\n5.8.3.2. Synthetic Programmable Hallucinations\n5.9. Hyperpersonalized Computing\n5.10. Cyborgization\n5.10.1. Posthuman Transhumanism\n5.11. Emergent Complexity in Distributed Intelligent Systems\n\nThe Complexity of Human Civilization\n6.1. Thought Perspectives\n6.2. Tradition\n6.3. Modernity\n6.4. Post-modernity\n6.5. Meta-modernity\n6.6. A Systems Level Overview\n6.7. Our Planetary Ecosystem\nSumma Technologiae\n7.1. Stanisław Lem: Who?\n7.2. Summary of His Book\n7.3. Evolution: Technological vs Biological vs Cosmic\n7.4. Future Predictions\n7.5. Virtual Reality\n7.6. What Can We Learn From This?\n7.7. Epistemological Collapse\n7.8.\n\nBig History View\n8.1. Big Bang(s) in Inflationary Cosmology\n8.2. Evolution of Matter\n8.3. Complexity of Matter\n8.4. Self-Awareness of Matter\n8.4.1. You. You’re self aware.\n8.5. Is Consciousness Computable?\n8.5.1. Emulating a Mind?\n8.5.2. Emergent Computations in Distributed Neural Circuits\n8.5.3. Multi-scale Collective Intelligence in Matter as a Unifying Concept\n8.5.4. The Brain as a Complex System\n8.5.5. Evolution and Self-Organization\n8.6. Does Anyone Know What’s Happening Here?\n8.7. A Hypothetical Technological Singularity\n8.7.1. Arguments For\n8.7.2. Arguments Against\n8.7.3. Techno-Religious Optimism\n8.7.4. Popularization\n8.7.5. Physical Constraints\n8.7.6. Vernor Vinge’s Science Fiction\n\nTutorials\n9.1. How to “Talk” to AI\n9.1.1. GPT-3\n9.1.2. Phone Apps\n9.2. Making AI Generated Images\n9.2.1. Closed Source; The Black Box\n9.2.1.1. DALL-E 2\n9.2.2. Open Source; Where Innovation Happens\n9.3.\n\nWho am I?\n10.1. Names and Aliases\n10.2. The Transient Nature of Identity\n10.3. Growing Up\n10.4. The Suburbs\n10.5.Our ancestors survived in the trees while the dinosaurs were being wiped out 65,000,000 years ago. Our journey as a species over the past 200,000 years has been a fascinating tale. We’re a bipedal hominid species that are now the apex predator of the planet, and we’ve now created an entire civilization sprouting from our collective ingenuity and curiosity. All of the qualities of human beings give us the ability to adapt to a wide variety of environments and circumstances. From our humble beginnings foraging along the savannah hunting animals, to sending probes to different planets and even the Sun, we have a global internet, holograms, virtual reality, microwaveable meals, infinite entertainment.
\nOur species originally came out of Africa, several millions of years ago. We can see our genetic lineage from fossil evidence over the course of millions of years of development. It seems like we just really liked wandering and exploring, this is what eventually took our ancestors out of Africa towards the rest of the world.
\n\n\n“Genetics is really good at telling us qualitative things about the order of events, and relative time frames.”
\n
\n\n“Just by looking at DNA from present day individuals we’ve been able to infer a pretty good outline of human history,” Akey says. “A group dispersed out of Africa maybe 50 to 60 thousand years ago, and then that group traveled around the world and eventually made it to all habitable places of the world.”
\n
Joshua Akey, studies Human Genome at Princeton University
Read more about it here!
Our ability to use tools is what differentiates us from many other species. But we’re not the only one that enjoys using tools! Tool use is broadly present within the animal kingdom. Besides humans, other mammals, birds, fish, cephalopods, and even insects find themselves using tools! Take a look around you, what kinds of tools make up your environment? What are the tools that compose your life? Tools allow us to take our intentions and make them that much easier to conduct upon the world. Tools make life easy, they solve problems, they are everywhere.
\nWe use them to assemble and prepare food, to navigate our world, to listen to music, and much else of everyday life is a symphony of tool usage. Of course, in the current era humans have their most prized tool- the smartphone. It is a fusion of many previous tools into one super-tool, it allows us to communicate with each other from anywhere on the planet and hail any purchasable material onto our doorsteps. Tools allow us to do things that would otherwise be difficult or downright impossible! Tools can be used to hurt or harm, but it’s up to the intention of the wielder.
\nIt’s as if a tool becomes an extension of your body when you’re using it, a newfound power to do new things upon the world. How will the tools of the future change what becomes of the present moment? Humans have been inventing, tinkering, and perfecting their tools in order to
\n\n(To be continued.)
\n\n(To be continued.)
\n\n(To be continued.)
\n\n(To be continued.)
\n\n(To be continued.)
\n\n(To be continued.)
\n\n//from Tomas Bjorkman’s book From God to Market
\n\n\n", "frontmatter": { "title": "A Book About Everything and Nothing", "slug": "a-book-about-everything-and-nothing", "date": "2022-03-08", "author": "Alif Jakir", "description": "Assorted blather from a Homo Sapiens with an internet connection. (Unfinished / under construction.)", "tags": [ "AI", "creativity", "metamodernism", "essay", "unfinished" ] }, "parsedFootnotes": [], "tableOfContents": [ { "id": "brains-the-infinite-novelty-engine", "text": "Brains: The Infinite Novelty Engine", "level": 2 }, { "id": "introduction-to-whatever-this-stuff-is-", "text": "Introduction to whatever this stuff is!", "level": 3 }, { "id": "1-historical-context", "text": "1. Historical Context", "level": 2 }, { "id": "1-1-the-human-journey", "text": "1.1. The Human Journey", "level": 3 }, { "id": "1-2-tools-tools-tools-", "text": "1.2. Tools, tools, tools!", "level": 3 }, { "id": "1-3-invention-of-language", "text": "1.3. Invention of Language", "level": 3 }, { "id": "1-4-technological-catalysts", "text": "1.4. Technological Catalysts", "level": 3 }, { "id": "2-computers", "text": "2. Computers", "level": 2 }, { "id": "2-1-a-civilizational-shift", "text": "2.1. A Civilizational Shift", "level": 3 }, { "id": "2-2-the-information-society", "text": "2.2. The Information Society", "level": 3 }, { "id": "2-3-internet-a-global-brain", "text": "2.3. Internet; A Global Brain", "level": 3 }, { "id": "2-4-", "text": "2.4.", "level": 3 }, { "id": "6-the-complexity-of-human-civilization", "text": "6. The Complexity of Human Civilization", "level": 2 }, { "id": "6-1-thought-perspectives", "text": "6.1. Thought Perspectives", "level": 3 } ] }, { "slug": "ad-infinitum-connexionem", "title": "Ad Infinitum Connexionem", "description": "Two essays from November 2021—an apologia to other species, then philosophical reflections on hyperconnectivity, language, and GPT-3’s implications for humanity.", "tags": [ "hyperconnectivity", "technology", "GPT-3", "essay", "digital life", "philosophy", "ecology", "Holocene" ], "date": "2021-11-16", "formattedDate": "November 16, 2021", "content": "\n“The animist thought perspective: Emerged c. 50,000 years ago
\n
during the Stone Age. Characterised by animistic and magical
beliefs, an occupation with a spirit world and no differentiation
between physical and mental reality. Still present among some
indigenous populations today.
• The religious or pre-modern thought perspective: Emerged c. 800
BCE at different locations across the central axis of Eurasia, in some
aspects as early as 2000 BCE in Egypt and Mesopotamia, but to its
fullest extent only blossoming after 500 CE with the consolidation
of the great moral religions such as Christianity, Islam or
Buddhism. Characterised by transcendental ideas of salvation,
divine law and the rationalisation of mythology in accordance with
universal theological principles. Still dominant in many developing
countries and in certain areas of the West.
• The rational or modern thought perspective: Emerged c. 1500 CE
in Europe during the Renaissance, but in some aspects in its protovariant as early as 500 BCE in Greece. Blossoming only in the
nineteenth and twentieth centuries. Characterised by rationalistic
and scientific thought, notions of progress and material growth,
emancipation of the individual from arbitrary religious and
political control, and human rights and democracy. Remains the
most dominant thought perspective today.
• The postmodern thought perspective: Emerged in the twentieth
century in the West, though some aspects appeared in the late
eighteenth century. Characterised by a critique of rationalism,
progress and established power relations, concerns with issues such
as the environment, gender and race, and a highly relativistic and
pluralistic perspective. Yet to fully blossom and become the most
dominant, but highly influential in much of intellectual life in the
West today and particularly common among the educated classes.”
November 16, 2021
\n\nOh, who are we to be so rude to you all!
\nThis is an apology to all of the species of the world—our rapid consumptive habits have consumed your whole world! And we dare to call you as encroachers on us, how ashamed are we? We have taken by hook and sinker that this world was granted to us, us alone.
\nBut we have consumed all of your homes, how we shall never gaze upon a sea-cow in the flesh! It is a world of tragedy!
\nDarwin’s tortoise soup, was it worth it? We shall never know them again.
\nWe never collectively decided to do anything, but within us all was the festering to grow.
\nAnd grow we did—growth we must have.
\nGrowth upon growth, vines and tangles upon crumbles. Our future gazes upon us now.
\nTo what are we growing into?
\nIn the year 2140, no species besides that which we can capitalize on will exist in the world. It shall be a world of pristine strangeness. But it will be a lot like the days of today, people can always view a virtual hologram of life itself.
\nDoes this capture anything of life at all? What is life?
\nDoes the human quest have an end to it? The rising and the falling of our civilizational soup—is there something to it of greater sense?
\nWe gaze upon our machines, and they gaze back upon us till the end of time.
\nAnd soon it won’t be long before we become our machines.
\nWhat then will we become?
\nAnd shall they have thoughts just the same as us?
\n\n\n\nNote: The essay below is mostly philosophical reflection on hyperconnectivity and on GPT-3’s implications for humanity—language models, machine-generated text, and what it means to stay the critical agent in the loop.
\n
November 16, 2021
\nI am privy to the strangeness of this observation recently, how have our machines accelerated our connections to each other? Has this been for the better? How has it affected our motivations? How has it warped our understanding of the other? Has it made us more fickle, more detached? What has the information age swamped us inside of?
\nHow many of us are capable of the type of reflection that is necessary for recognizing the impacts of the hyperconnectivity we’re slowly being subsumed inside of?
\nAnd is there even any way to pause, stop life for one second?
\nIs there a way to slow the tempo of the frenetic life that we have been enveloped within?
\nIs there a world where it’s too easy to reach another human being? What if you are that someone who prefers to be left alone, to their own ways of thinking and being?
\nHow can one avoid the noise of modernity?
\nAll the best, all the worst of humanity, it has been amplified by our connective technologies. We have never before in history been able to scam each other at such an accelerated pace. We have never before in history had the capacity to find immense quantities of nonsense which articulate any of the infinite absurd convictions that may fly into our heads at any given moment. How can we create a world more inundated with criticality?
\nWhat of the world wherein we are endlessly connected to an infinite three dimensional stream of whatever our ‘identity’ is presupposed to enjoy? How shall we ever be able to escape the buzz of stimuli, all the novelty which convinces our mammalian brain that “I want more”?
\nAre we going to choose to be a people that are passive recipients of whatever our cultural dialogue is? Or should we rather try to shape the very nature of culture?
\nDoes it matter? Is culture a superficial embedding on top of our economic superstructure, or does it inform our values on what to do with ourselves?
\nWhat kinds of behaviors should we try to aim for as a species?
\nWhat are the types of questions we should be asking of ourselves? What kinds of questions should we be asking each other?
\nWhat is the nature of this world that I so often find myself awakening in? Is it possible to understand this complex, chaotic system?
\nThere are underlying patterns behind the structures of thought, our relationships to others, the world around us.
\nScientists hope to glimpse evermore into the truth of this. But with this power, comes a massive responsibility. If those who know you better than anyone you know are a massive corporation with endless data on your habits and behavior, who keeps track of the incentive structures?
\nWe took the smartest people in the world and placed them in a room, and paid them to find ways to create addicts to the hyperstimuli of a small little rectangle.
\nThis little rectangle has changed an immense amount of lives, in immense rapidity, in measurable ways. We’re running into the consequences of connecting the human species to each other, the whole planet, at the speed of thought.
\nThe human species then can be considered from the perspective of a cybernetic superorganism. Not literally, of course, but metaphorically. It is an entity that requires a sort of constant feedback loop of interaction between its subjects and the environment in order to grow its connectivity. Since we have given up some very human aspects of ourselves to accelerate the capacity of this machine of hyperconnectivity. One obvious enough to observe in the everyday is the ever-shrinking concept of the “present”.
\nThere is less time for people to be alone with their thoughts, and to reflect on the kind of life they want to be surrounded by, outside of the demands of machinic life. Machinic life of a different kind to the industrial revolution, as in the growth of the post-industrial digital world.
\nThe brain is a self-organizing component of our Universe that seeks to create evermore accurate ways of finding and maintaining the possibilities of its future. Intelligence can be defined as an algorithm that maximizes the capacity for future actions.
\nThere is an immense lucidity that arises from this existential self realization of the unlikeliest of thermodynamic scenarios that had to occur for your incidental existence.
\nIf you’re reading this right now, there’s probably never been a person like you in this Universe. And how wonderful it is that we’re both alive at the same time. The thoughts existing in my brain are being transformed through these symbols on this screen into a form of intrinsic meaning, that we can both understand. And it is through our language itself that an immense amount of recognition and clarity can be discovered. I can seek to understand you wholly as a person of immense complexity, with thoughts and feelings separate of my own.
\nI’d like to discuss something… a tad strange and mysterious. It is in regards to this autocompletion mechanism we have on Google Docs, where I am writing this. Google Docs suggests to me several words in advance what I could put down if I so desire. It also, thankfully, corrects my at times mediocre adherence to grammatical conventions. What if there was something that could complete entire sentences, nay, paragraphs for you in advance?
\nWell, you see, there is.
\nIt’s something I’ve been toying with, a gigantic natural language model known as GPT-3. It’s able to take in some kind of prompt that you’ve written, and it extrapolates and tries to guess a similar pattern and structure of language to continue with. It has the capacity to make endlessly entertaining generations that cannot be distinguished from the text that a human being is able to generate.
\nThat a machine can mimic our language, if it was trained on enough samples of our writing, seems slightly spooky perhaps. Or exciting. It really depends on who you ask. GPT-3 in the largest model has 175 billion parameters that it’s trained on. Each parameter is, on average, a four letter word. 60% of this is from the open internet. This brings up the fascinating realization that we as a species now have an artifact of our compressed collective cybernetic intelligence. An oracle that we can basically ask to do anything we want, language wise.
\nWe can ask it questions, ask it to create poems, stories, articles. We can ask it to summarize text, all such wonderful things. But this is merely the beginning, baby steps towards giving general intelligence capacities to machines.
\nA machine that has the same capacity for existential self-realization sounds pretty far-fetched. However, if we try to understand the underlying structure of cognition, how it operates, and instantiate it within some kind of modality that allows it to constantly learn…
\nWell… we’ve got something completely world-changing. This is of course, a completely new phase of history that will occur.
\nYou see, playing with GPT-3 allowed me to explore my own curiosity and expand the range of complexity that my imagination is able to play with. GPT-3 acts like a mirror for my own cognition, accelerating it far beyond the boundaries of what I know. Of course, I’m aware that it does not recognize falsehoods from truths, but that’s the point of having me as the critical agent.
\nWithin 10 years, more content on the internet will be machine rather than human generated. Finding the words then, of a true human being on this planet, will be much more valuable. Of course, you can imagine that there would be entire authors who are nothing more than Artificial Intelligence.
\nThere will be more and more of these artifacts popping up soon, and they will be immensely powerful, compared to the baby of GPT-3.
\nMaybe it wasn’t such a bad idea for us all to be babbling all of those years, if we can make something like GPT-3 out of it all.
\nWhen I use it, I’m fully aware it has no degree of intelligence or awareness, but I have while using it a gravity of respect for its capacities. It impresses me in its linguistic flexibility; its creativity is surprising.
\nThe world is changing, and so I, being part of the world, am changing with it.
\nWhat else will change?
\n", "frontmatter": { "title": "Ad Infinitum Connexionem", "slug": "ad-infinitum-connexionem", "date": "2021-11-16", "author": "Alif Jakir", "description": "Two essays from November 2021—an apologia to other species, then philosophical reflections on hyperconnectivity, language, and GPT-3’s implications for humanity.", "tags": [ "hyperconnectivity", "technology", "GPT-3", "essay", "digital life", "philosophy", "ecology", "Holocene" ] }, "parsedFootnotes": [], "tableOfContents": [ { "id": "quagmire-of-holocene", "text": "Quagmire of Holocene", "level": 2 }, { "id": "apologia-to-species-unfortunate-to-have-been-born-as-non-homo-sapiens", "text": "Apologia to species unfortunate to have been born as non-Homo sapiens", "level": 3 }, { "id": "ad-infinitum-connexionem", "text": "Ad Infinitum Connexionem", "level": 2 } ] }, { "slug": "consciousness-as-epiphenomena", "title": "Why consciousness cannot be explained away as 'Epiphenomenal'", "description": "A comprehensive exploration of epiphenomenalism—the radical theory that consciousness is merely a byproduct of neural activity, with profound implications for free will, moral responsibility, and the nature of human experience.", "tags": [ "consciousness", "philosophy", "neuroscience", "cognition", "epiphenomenalism", "free will", "philosophy of mind" ], "date": "2025-03-14", "formattedDate": "March 14, 2025", "content": "Imagine discovering that your most intimate experience—your consciousness, your sense of self, your feeling of making decisions—is nothing more than steam rising from the engine of your brain. This is the provocative claim of epiphenomenalism, a theory that suggests consciousness is a mere byproduct of neural processes, as causally irrelevant as the whistle of a steam locomotive to its forward motion.
\nThe implications are staggering. If consciousness is truly epiphenomenal, then our subjective experiences, our qualia, our very sense of agency might be elaborate illusions—shadows cast by the real work happening in the neural substrate below.
\n\nThe term "epiphenomenalism" was coined by Thomas Henry Huxley in 1874, though the concept traces back to ancient philosophical traditions. Huxley, known as "Darwin's Bulldog" for his fierce advocacy of evolutionary theory, proposed a radical reconceptualization of consciousness that would challenge our most fundamental intuitions about the mind.
\nIn his famous analogy, Huxley compared consciousness to the steam whistle of a locomotive. Just as the whistle is produced by the engine's operations but doesn't contribute to the train's movement, consciousness arises from brain activity but exerts no causal influence on behavior or cognition. The whistle exists, it's real, but it's fundamentally epiphenomenal—a side effect rather than a driver.
\nThis wasn't merely academic speculation. Huxley was grappling with a fundamental tension in 19th-century science: how to reconcile the emerging mechanistic understanding of biology with the undeniable reality of conscious experience. Epiphenomenalism offered a elegant solution that preserved both scientific materialism and phenomenological reality.
\n\nEpiphenomenalism emerged partly as a response to Cartesian dualism's infamous "interaction problem." René Descartes had proposed that mind and matter were distinct substances, but this raised the thorny question of how an immaterial mind could causally interact with a material brain.
\nEpiphenomenalism sidesteps this problem entirely by proposing a one-way causal relationship: brain states cause mental states, but mental states cause nothing. This preserves the intuitive distinction between mind and matter while avoiding the mysterious causal interactions that plagued Cartesian dualism.
\n\nContemporary neuroscientist David Chalmers distinguishes between the "easy problems" and the "hard problem" of consciousness. The easy problems—though technically challenging—involve explaining cognitive functions like attention, memory, and information processing. These can be addressed through standard neuroscientific methods.
\nThe hard problem, however, concerns the existence of subjective experience itself: why there is something it's like to be conscious. Why do we have qualitative, subjective experiences (qualia) rather than simply processing information like sophisticated zombies?
\nEpiphenomenalism offers a provocative answer: subjective experience exists because it's an inevitable byproduct of certain types of information processing, but it serves no functional purpose. Consciousness is the brain's excess energy, dissipated as experiential heat.
\n\nModern neuroscience has identified numerous neural correlates of consciousness (NCCs)—brain patterns that reliably correspond to conscious states. Studies using techniques like fMRI, EEG, and transcranial magnetic stimulation reveal that consciousness appears to emerge from the integration of information across multiple brain networks.
\nThe Global Workspace Theory, proposed by Bernard Baars and developed by Stanislas Dehaene, suggests that consciousness arises when information becomes globally accessible across brain systems. This integration creates the unified, coherent experience we call consciousness—but crucially, the integration is what does the causal work, not the conscious experience itself.
\nLet's simulate this with a neural network model that demonstrates how global workspace dynamics might generate consciousness as an epiphenomenal byproduct:
\n\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom scipy.integrate import odeint\n\nclass GlobalWorkspaceNetwork:\n def __init__(self, n_modules=8, n_global=4):\n self.n_modules = n_modules\n self.n_global = n_global\n \n # Connection matrices\n self.W_local = np.random.randn(n_modules, n_modules) * 0.1\n self.W_global = np.random.randn(n_global, n_modules) * 0.3\n self.W_feedback = np.random.randn(n_modules, n_global) * 0.2\n \n # Consciousness emerges from global integration\n self.consciousness_threshold = 0.5\n \n def dynamics(self, state, t, stimulus):\n modules, global_nodes = state[:self.n_modules], state[self.n_modules:]\n \n # Local processing (unconscious)\n dm_dt = -modules + np.tanh(\n np.dot(self.W_local, modules) + \n np.dot(self.W_feedback, global_nodes) + \n stimulus\n )\n \n # Global workspace integration\n dg_dt = -global_nodes + np.tanh(\n np.dot(self.W_global, modules)\n )\n \n return np.concatenate([dm_dt, dg_dt])\n \n def compute_consciousness_level(self, global_state):\n \"\"\"Consciousness as emergent property of global integration\"\"\"\n integration = np.mean(global_state ** 2)\n coherence = 1 - np.var(global_state) / (np.mean(global_state) + 1e-6)\n \n # Consciousness emerges but doesn't cause anything\n consciousness = integration * coherence\n return consciousness if consciousness > self.consciousness_threshold else 0\n \n def simulate_conscious_access(self, stimulus_strength=1.0, duration=10.0):\n t = np.linspace(0, duration, 1000)\n \n # Stimulus appears at t=2, disappears at t=8\n stimulus = np.zeros((len(t), self.n_modules))\n stimulus[(t > 2) & (t < 8), 0] = stimulus_strength\n \n # Initial state\n initial_state = np.random.randn(self.n_modules + self.n_global) * 0.1\n \n # Track consciousness emergence\n consciousness_levels = []\n \n for i, stim in enumerate(stimulus):\n if i == 0:\n state = initial_state\n else:\n state = odeint(self.dynamics, state, [t[i-1], t[i]], args=(stim,))[-1]\n \n global_state = state[self.n_modules:]\n consciousness = self.compute_consciousness_level(global_state)\n consciousness_levels.append(consciousness)\n \n return t, consciousness_levels, stimulus[:, 0]\n\n# Demonstrate epiphenomenal consciousness\nnetwork = GlobalWorkspaceNetwork()\ntime, consciousness, stimulus = network.simulate_conscious_access()\n\nprint(f\"Peak consciousness level: {max(consciousness):.3f}\")\nprint(\"Note: Consciousness emerges from integration but doesn't cause the integration\")Consider the phenomenon of change blindness, where people fail to notice large changes in their visual environment when attention is diverted. This suggests that much of our visual processing occurs unconsciously, with consciousness providing a post-hoc narrative rather than directing attention itself[^9].
\n\nBenjamin Libet's groundbreaking experiments in the 1980s provided empirical support for epiphenomenalist intuitions. Participants were asked to flex their wrist while monitoring their intention to move. EEG recordings showed that brain activity (the "readiness potential") began approximately 350 milliseconds before participants reported being aware of their intention to move.
\nThis suggests that unconscious brain processes initiate action before conscious intention arises. Consciousness appears to be a latecomer to the party, constructing post-hoc narratives about decisions already made by unconscious neural mechanisms.
\nSubsequent studies have extended these findings to more complex decisions. Using fMRI, researchers can predict with up to 70% accuracy whether a person will choose to add or subtract numbers up to 10 seconds before the person reports making the decision consciously.
\n\nIf consciousness is epiphenomenal, why is it so extraordinarily rich and detailed? Consider the vast qualitative landscape of human experience: the redness of red, the pain of heartbreak, the joy of mathematical insight, the ineffable sense of being present in the world.
\nEpiphenomenalists argue that this richness emerges from the complexity of underlying neural processes. Just as the intricate patterns of steam from a locomotive reflect the complexity of the engine's operations, the richness of consciousness reflects the extraordinary complexity of neural information processing.
\nThe philosopher Frank Jackson's famous thought experiment of Mary the color scientist illustrates this complexity. Mary knows everything physical about color but has never experienced color herself, having been raised in a black-and-white environment. When she finally sees color, does she learn something new?
\nFrom an epiphenomenalist perspective, Mary gains new qualitative experiences, but these experiences don't provide new information about the world—they're simply new ways of representing information her brain already possessed.
\n\nOne of consciousness's most remarkable features is its unity. Despite receiving information from multiple sensory modalities and processing it through distributed brain networks, we experience a single, coherent stream of consciousness.
\nThe binding problem asks how the brain integrates these diverse information streams into unified conscious experience. Epiphenomenalists suggest that this binding is a necessary consequence of how the brain processes information globally, and the unified conscious experience is simply how this integrated processing feels from the inside.
\nNeural synchrony—the coordinated firing of neurons across different brain regions—appears to be crucial for binding. Let's model this synchronization and show how consciousness emerges as oscillatory patterns stabilize:
\n\nclass OscillatoryBindingNetwork:\n def __init__(self, n_regions=6, coupling_strength=0.1):\n self.n_regions = n_regions\n self.coupling_strength = coupling_strength\n \n # Natural frequencies for each brain region\n self.omega = np.random.uniform(8, 12, n_regions) # Alpha band\n \n # Coupling matrix (anatomical connectivity)\n self.coupling_matrix = self.generate_anatomical_network()\n \n def generate_anatomical_network(self):\n \"\"\"Simulate anatomical connectivity between brain regions\"\"\"\n # Small-world network topology\n K = np.zeros((self.n_regions, self.n_regions))\n for i in range(self.n_regions):\n for j in range(i+1, self.n_regions):\n if np.random.random() < 0.3: # Connection probability\n strength = np.random.exponential(0.2)\n K[i,j] = K[j,i] = strength\n return K\n \n def kuramoto_dynamics(self, phases, t):\n \"\"\"Kuramoto model for neural synchronization\"\"\"\n dphase_dt = np.zeros_like(phases)\n \n for i in range(self.n_regions):\n # Natural frequency\n dphase_dt[i] = self.omega[i]\n \n # Coupling term\n for j in range(self.n_regions):\n if i != j:\n dphase_dt[i] += (self.coupling_strength * \n self.coupling_matrix[i,j] * \n np.sin(phases[j] - phases[i]))\n \n return dphase_dt\n \n def compute_binding_strength(self, phases):\n \"\"\"Measure of how bound the oscillations are\"\"\"\n # Order parameter (Kuramoto synchronization measure)\n complex_order = np.mean(np.exp(1j * phases))\n return np.abs(complex_order)\n \n def compute_consciousness_emergence(self, phases, threshold=0.6):\n \"\"\"Consciousness emerges when binding exceeds threshold\"\"\"\n binding = self.compute_binding_strength(phases)\n \n # Information integration measure\n integration = self.compute_information_integration(phases)\n \n # Consciousness as emergent property\n consciousness = binding * integration\n return consciousness if consciousness > threshold else 0\n \n def compute_information_integration(self, phases):\n \"\"\"Simplified information integration measure\"\"\"\n # Mutual information between oscillator pairs\n total_mi = 0\n n_pairs = 0\n \n for i in range(self.n_regions):\n for j in range(i+1, self.n_regions):\n # Phase difference as proxy for mutual information\n phase_diff = np.abs(phases[i] - phases[j])\n mi = 1 - (phase_diff / np.pi) # Normalized\n total_mi += mi\n n_pairs += 1\n \n return total_mi / n_pairs if n_pairs > 0 else 0\n \n def simulate_binding_dynamics(self, duration=10.0, disturbance_time=5.0):\n \"\"\"Simulate how consciousness emerges and dissolves\"\"\"\n t = np.linspace(0, duration, 1000)\n \n # Initial random phases\n initial_phases = np.random.uniform(0, 2*np.pi, self.n_regions)\n \n # Solve differential equation\n solution = odeint(self.kuramoto_dynamics, initial_phases, t)\n \n # Track consciousness emergence\n consciousness_levels = []\n binding_levels = []\n \n for i, phases in enumerate(solution):\n # Add disturbance at specific time (like anesthesia)\n if abs(t[i] - disturbance_time) < 0.1:\n phases += np.random.uniform(-np.pi, np.pi, self.n_regions)\n \n binding = self.compute_binding_strength(phases)\n consciousness = self.compute_consciousness_emergence(phases)\n \n binding_levels.append(binding)\n consciousness_levels.append(consciousness)\n \n return t, binding_levels, consciousness_levels\n\n# Demonstrate oscillatory binding and consciousness emergence\nnetwork = OscillatoryBindingNetwork(coupling_strength=0.15)\ntime, binding, consciousness = network.simulate_binding_dynamics()\n\nprint(f\"Peak binding strength: {max(binding):.3f}\")\nprint(f\"Peak consciousness: {max(consciousness):.3f}\")\nprint(\"Consciousness emerges from binding but is causally inert\")When neurons fire in synchrony, their outputs are more likely to be integrated, creating the unified conscious experience. The synchrony does the causal work; consciousness is simply what synchrony feels like[^10].
\n\nIf consciousness is epiphenomenal, what happens to free will? If our conscious decisions don't actually cause our actions, are we truly responsible for what we do?
\nThis challenge strikes at the heart of moral and legal systems built on assumptions of personal responsibility. Traditional notions of praise, blame, punishment, and reward seem to presuppose that conscious agents have genuine causal efficacy in the world.
\nEpiphenomenalists offer various responses to this challenge:
\n1. Compatibilist Reframing: Perhaps moral responsibility doesn't require conscious causation. What matters is that actions flow from the agent's own neural processes, even if consciousness itself is epiphenomenal. A person is responsible for their actions in the same way a computer is "responsible" for its outputs—through the complex causal chains that produce behavior.
\n2. Pragmatic Justification: Even if free will is an illusion, believing in moral responsibility serves important social functions. The practice of holding people accountable shapes behavior through neural mechanisms, even if consciousness itself is causally inert.
\n3. Levels of Description: Agency might be real at the level of psychological description even if it's absent at the neural level. Just as chemistry is real despite being reducible to physics, moral agency might be real despite being reducible to neuroscience.
\n\nEven if consciousness doesn't cause our choices, it profoundly shapes how we experience choice-making. The phenomenology of deliberation—weighing options, feeling conflicted, experiencing the moment of decision—remains vivid and meaningful.
\nConsider the experience of moral struggle. When facing an ethical dilemma, we feel the weight of competing considerations, the pull of different values, the difficulty of choice. From an epiphenomenalist perspective, this struggle reflects real computational processes in the brain, with consciousness providing a compelling narrative overlay.
\nThe struggle is real—it's just not located where we think it is. The real work happens in unconscious neural networks, while consciousness provides a dramatic, first-person account of the proceedings.
\n\nOne of the strongest challenges to epiphenomenalism comes from the causal exclusion argument. If every physical event has sufficient physical causes, where is there room for mental causation? But if mental states don't cause anything, how can they be genuinely real rather than mere illusions?
\nThis creates a trilemma:
\nEpiphenomenalists accept premises 2 and 3 while rejecting 1, but this forces them to explain how epiphenomenal mental states can be real yet causally inert.
\n\nIf consciousness is causally irrelevant, why did it evolve? Natural selection operates on traits that affect survival and reproduction. If consciousness has no causal efficacy, it shouldn't be subject to selective pressure.
\nEpiphenomenalists propose several solutions:
\n1. Byproduct Hypothesis: Consciousness might be an unavoidable byproduct of the complex information processing that natural selection did favor. Just as the whiteness of bones serves no function but inevitably accompanies their calcium composition, consciousness might inevitably accompany certain types of neural organization.
\n2. Package Deal: The neural mechanisms that produce consciousness might be inextricably linked to causally relevant cognitive abilities. Selection for these abilities brings consciousness along as an unavoidable package deal.
\n3. Misattribution: Perhaps what we call "consciousness" actually refers to various causally relevant cognitive processes, and the truly epiphenomenal aspects are evolutionary spandrels—architectural byproducts with no function.
\n\nJackson's Mary thought experiment poses another challenge. If consciousness is epiphenomenal, how can Mary learn something new when she first experiences color? How can causally inert experiences constitute genuine knowledge?
\nContemporary epiphenomenalists argue that Mary gains new ways of representing information she already possessed, not new propositional knowledge. She acquires new representational formats—new ways her brain can encode color information—but no new facts about the world.
\nThis connects to broader questions about the relationship between conscious experience and knowledge. Perhaps the intuition that Mary learns something new reflects our tendency to conflate different types of information representation in the brain.
\n\nStudies of patients with severed corpus callosum (the bridge connecting brain hemispheres) provide fascinating insights into consciousness and causation. These patients sometimes exhibit conflicting behaviors between their left and right hands, suggesting multiple control systems operating independently.
\nImportantly, only the verbal left hemisphere reports conscious experiences, while the mute right hemisphere demonstrates sophisticated behavior without apparent consciousness. This suggests that consciousness might be a specialized function of particular brain regions rather than a global property.
\nFrom an epiphenomenalist perspective, this supports the view that consciousness is a particular type of information processing (verbal reportability) rather than a general causal force.
\n\nPatients with blindsight have damaged visual cortices but retain unconscious visual processing. They claim to be blind in parts of their visual field but can navigate obstacles and identify objects at above-chance levels when forced to guess.
\nThis demonstrates that sophisticated visual processing can occur without consciousness. The conscious visual experience appears to be an additional layer on top of functional visual processing—potentially an epiphenomenal layer.
\n\nStudies of anesthetic action provide another window into consciousness. General anesthetics appear to disrupt the integration of information across brain networks while leaving local processing intact. This supports theories that consciousness emerges from global information integration.
\nCrucially, anesthetics eliminate conscious experience while preserving many automatic functions. This suggests that consciousness is indeed dissociable from the brain's causal operations—supporting epiphenomenalist intuitions.
\n\nPhenomenologists argue that epiphenomenalism misses something essential about consciousness: its first-personal, subjective character. Even if consciousness doesn't cause behavior, it constitutes our fundamental mode of being-in-the-world.
\nMaurice Merleau-Ponty emphasized the embodied nature of consciousness—how our subjective experience is always already embedded in our bodily engagement with the world. From this perspective, asking whether consciousness "causes" anything misses the point; consciousness is the very condition for the appearance of causation.
\nEdmund Husserl's phenomenological reduction brackets questions of causal efficacy to focus on the structures of experience itself. The richness and intentionality of consciousness might be irreducible to neural processes, regardless of causal relationships.
\n\nEven sophisticated epiphenomenalist theories struggle with the hard problem of consciousness. Why should there be subjective experience at all? Why shouldn't we simply be unconscious information-processing systems?
\nDavid Chalmers argues that even if we can explain all cognitive functions materialistically, the existence of subjective experience remains mysterious. Epiphenomenalism acknowledges this mystery while denying consciousness any causal role.
\n\nIf consciousness is epiphenomenal, this has profound implications for medical practice. Consider patients in vegetative states who show signs of unconscious information processing. Are they experiencing anything, or are they sophisticated unconscious systems?
\nEpiphenomenalism suggests that the presence of appropriate neural activity might indicate conscious experience even without behavioral responses. This could revolutionize how we approach consciousness disorders and end-of-life decisions.
\n\nAs we develop increasingly sophisticated AI systems, epiphenomenalism offers a framework for thinking about machine consciousness. If consciousness is simply a byproduct of certain types of information processing, then sufficiently complex AI systems might develop consciousness automatically.
\nThis raises ethical questions about the treatment of potentially conscious AI systems. If consciousness has no causal efficacy, then conscious AI might suffer without anyone—including the AI itself—being able to report or act on that suffering.
\n\nUnderstanding consciousness as epiphenomenal might inform educational and therapeutic practices. If conscious insight doesn't directly cause behavioral change, then therapeutic interventions might need to target unconscious neural processes rather than conscious understanding.
\nThis could support approaches like cognitive-behavioral therapy that focus on changing thought patterns and behaviors rather than just insight, or mindfulness practices that work with unconscious attentional processes.
\n\nPanpsychist theories propose that consciousness is a fundamental feature of reality, present even in simple physical systems. Integrated Information Theory (IIT), developed by Giulio Tononi, offers a mathematical framework for understanding consciousness as integrated information.
\nIIT suggests that any system that integrates information has some degree of consciousness, with human-level consciousness emerging from highly integrated neural networks. Let's implement a simplified IIT calculation to demonstrate how consciousness might be quantified:
\n\nimport numpy as np\nfrom itertools import combinations\nfrom scipy.stats import entropy\n\nclass IntegratedInformationCalculator:\n def __init__(self, system_size=4):\n self.n = system_size\n self.states = 2 ** system_size # Binary system\n \n def generate_transition_matrix(self, connectivity_strength=0.7):\n \"\"\"Generate state transition matrix for the system\"\"\"\n # Simple feedforward + recurrent connectivity\n transition_matrix = np.zeros((self.states, self.states))\n \n for state in range(self.states):\n current_bits = [(state >> i) & 1 for i in range(self.n)]\n \n # Next state depends on current state + noise\n for next_state in range(self.states):\n next_bits = [(next_state >> i) & 1 for i in range(self.n)]\n \n # Compute transition probability\n prob = 1.0\n for i in range(self.n):\n # Each bit influenced by previous bits\n inputs = sum(current_bits[j] for j in range(i))\n target_prob = 1 / (1 + np.exp(-(inputs - 1.5)))\n \n if next_bits[i] == 1:\n prob *= target_prob\n else:\n prob *= (1 - target_prob)\n \n transition_matrix[state, next_state] = prob\n \n # Normalize rows\n for i in range(self.states):\n if np.sum(transition_matrix[i, :]) > 0:\n transition_matrix[i, :] /= np.sum(transition_matrix[i, :])\n \n return transition_matrix\n \n def compute_effective_information(self, transition_matrix, subset):\n \"\"\"Compute effective information for a subset of nodes\"\"\"\n subset_size = len(subset)\n subset_states = 2 ** subset_size\n \n # Marginalize transition matrix to subset\n subset_transition = np.zeros((subset_states, subset_states))\n \n for full_state in range(self.states):\n for next_full_state in range(self.states):\n # Extract subset states\n current_subset = self.extract_subset_state(full_state, subset)\n next_subset = self.extract_subset_state(next_full_state, subset)\n \n subset_transition[current_subset, next_subset] += \\\n transition_matrix[full_state, next_full_state]\n \n # Normalize\n for i in range(subset_states):\n if np.sum(subset_transition[i, :]) > 0:\n subset_transition[i, :] /= np.sum(subset_transition[i, :])\n \n # Compute effective information\n ei = 0\n for i in range(subset_states):\n if np.sum(subset_transition[i, :]) > 0:\n prob_dist = subset_transition[i, :]\n prob_dist = prob_dist[prob_dist > 0] # Remove zeros for entropy\n ei += entropy(prob_dist, base=2)\n \n return ei / subset_states # Average effective information\n \n def extract_subset_state(self, full_state, subset):\n \"\"\"Extract the state of a subset from full system state\"\"\"\n subset_state = 0\n for i, node in enumerate(subset):\n if (full_state >> node) & 1:\n subset_state |= (1 << i)\n return subset_state\n \n def compute_phi(self, transition_matrix):\n \"\"\"Compute Φ (Phi) - the integrated information\"\"\"\n # Φ is the minimum effective information across all bipartitions\n min_ei = float('inf')\n \n # Consider all possible bipartitions\n nodes = list(range(self.n))\n \n for partition_size in range(1, self.n):\n for partition in combinations(nodes, partition_size):\n complement = [n for n in nodes if n not in partition]\n \n # Effective information of the partition\n ei_partition = self.compute_effective_information(\n transition_matrix, partition)\n ei_complement = self.compute_effective_information(\n transition_matrix, complement)\n \n # Minimum information across the cut\n ei_cut = min(ei_partition, ei_complement)\n min_ei = min(min_ei, ei_cut)\n \n return min_ei\n \n def analyze_consciousness_levels(self, connectivity_range=(0.1, 0.9), steps=10):\n \"\"\"Analyze how consciousness (Φ) varies with connectivity\"\"\"\n connectivities = np.linspace(connectivity_range[0], connectivity_range[1], steps)\n phi_values = []\n \n for conn in connectivities:\n transition_matrix = self.generate_transition_matrix(conn)\n phi = self.compute_phi(transition_matrix)\n phi_values.append(phi)\n \n return connectivities, phi_values\n \n def demonstrate_epiphenomenal_nature(self):\n \"\"\"Show that consciousness (Φ) doesn't affect behavior\"\"\"\n \n # Generate two systems with different Φ but same behavior\n tm1 = self.generate_transition_matrix(0.3) # Low integration\n tm2 = self.generate_transition_matrix(0.7) # High integration\n \n phi1 = self.compute_phi(tm1)\n phi2 = self.compute_phi(tm2)\n \n # Both systems can have identical input-output mappings\n # despite different consciousness levels\n \n return {\n 'system_1': {'phi': phi1, 'behavior': 'identical'},\n 'system_2': {'phi': phi2, 'behavior': 'identical'},\n 'conclusion': 'Consciousness level (Φ) varies without affecting behavior'\n }\n\n# Demonstrate integrated information and epiphenomenalism\niit_calc = IntegratedInformationCalculator(system_size=4)\n\n# Analyze consciousness across different network configurations\nconnectivities, phi_values = iit_calc.analyze_consciousness_levels()\nepiphenomenal_demo = iit_calc.demonstrate_epiphenomenal_nature()\n\nprint(f\"Φ range: {min(phi_values):.3f} to {max(phi_values):.3f}\")\nprint(f\"Peak consciousness at connectivity: {connectivities[np.argmax(phi_values)]:.2f}\")\nprint(f\"Epiphenomenal demonstration: {epiphenomenal_demo['conclusion']}\")This preserves causal relevance for consciousness while acknowledging its deep connection to information processing—though epiphenomenalists would argue that the information integration does the causal work, not the conscious experience that emerges from it[^11].
\n\nPredictive processing theories suggest that the brain is fundamentally a prediction machine, constantly generating models of sensory input and updating these models based on prediction errors.
\nFrom this perspective, consciousness might be the brain's highest-level predictive model—a global representation of the organism's state and environment. Let's model this using chaotic dynamics to show how consciousness emerges from hierarchical prediction systems:
\n\nimport numpy as np\nfrom scipy.integrate import odeint\n\nclass ChaoticPredictiveBrain:\n def __init__(self, n_levels=3, n_nodes_per_level=6):\n self.n_levels = n_levels\n self.n_nodes = n_nodes_per_level\n self.total_nodes = n_levels * n_nodes_per_level\n \n # Hierarchical prediction parameters\n self.prediction_precision = np.array([0.1, 0.3, 0.8]) # Higher levels more precise\n self.learning_rates = np.array([0.1, 0.05, 0.02]) # Slower at higher levels\n \n def predictive_dynamics(self, state, t, sensory_input):\n \"\"\"Hierarchical predictive processing with chaotic dynamics\"\"\"\n dstate_dt = np.zeros_like(state)\n \n # Reshape state into hierarchical levels\n levels = [state[i*self.n_nodes:(i+1)*self.n_nodes] \n for i in range(self.n_levels)]\n \n for level_idx in range(self.n_levels):\n level_start = level_idx * self.n_nodes\n level_end = (level_idx + 1) * self.n_nodes\n current_level = levels[level_idx]\n \n if level_idx == 0:\n # Sensory level: process input and receive predictions\n sensory_prediction = (levels[1][:self.n_nodes] \n if len(levels) > 1 else np.zeros(self.n_nodes))\n \n # Prediction error\n prediction_error = sensory_input[:self.n_nodes] - sensory_prediction\n \n # Chaotic Rössler-like dynamics modified by prediction error\n for i in range(0, self.n_nodes-2, 3):\n if i+2 < self.n_nodes:\n x, y, z = current_level[i:i+3]\n \n # Rössler attractor with prediction error modulation\n a, b, c = 0.1, 0.1, 14.0\n dstate_dt[level_start + i] = -y - z + prediction_error[i]\n dstate_dt[level_start + i + 1] = x + a * y\n dstate_dt[level_start + i + 2] = b + z * (x - c)\n \n else:\n # Higher levels: generate predictions\n lower_level = levels[level_idx - 1]\n \n # Generate prediction of lower level\n prediction = np.tanh(current_level) * self.prediction_precision[level_idx]\n \n # Update based on prediction error from below\n if level_idx < len(levels) - 1:\n higher_prediction = (levels[level_idx + 1][:self.n_nodes]\n if level_idx + 1 < len(levels) else np.zeros(self.n_nodes))\n prediction_error = current_level - higher_prediction\n else:\n prediction_error = np.zeros_like(current_level)\n \n # Chaotic dynamics for prediction generation\n for i in range(0, self.n_nodes-2, 3):\n if i+2 < self.n_nodes:\n x, y, z = current_level[i:i+3]\n \n # Modified Lorenz system\n sigma, rho, beta = 10.0, 28.0, 8.0/3.0\n \n dstate_dt[level_start + i] = (sigma * (y - x) + \n self.learning_rates[level_idx] * \n prediction_error[i])\n dstate_dt[level_start + i + 1] = (x * (rho - z) - y +\n 0.1 * prediction_error[i])\n dstate_dt[level_start + i + 2] = x * y - beta * z\n \n return dstate_dt\n \n def compute_predictive_coherence(self, trajectory):\n \"\"\"Measure coherence across predictive hierarchy\"\"\"\n n_timepoints = len(trajectory)\n coherence_scores = []\n \n for t in range(n_timepoints):\n state = trajectory[t]\n levels = [state[i*self.n_nodes:(i+1)*self.n_nodes] \n for i in range(self.n_levels)]\n \n # Cross-level correlation\n total_correlation = 0\n n_pairs = 0\n \n for i in range(self.n_levels - 1):\n for j in range(i + 1, self.n_levels):\n if len(levels[i]) > 0 and len(levels[j]) > 0:\n min_len = min(len(levels[i]), len(levels[j]))\n corr = np.corrcoef(levels[i][:min_len], levels[j][:min_len])[0, 1]\n if not np.isnan(corr):\n total_correlation += abs(corr)\n n_pairs += 1\n \n coherence = total_correlation / n_pairs if n_pairs > 0 else 0\n coherence_scores.append(coherence)\n \n return np.array(coherence_scores)\n \n def compute_information_flow(self, trajectory):\n \"\"\"Measure information flow between hierarchical levels\"\"\"\n # Simplified mutual information calculation\n information_flow = []\n \n for t in range(1, len(trajectory)):\n state_prev = trajectory[t-1]\n state_curr = trajectory[t]\n \n # Information flow from lower to higher levels\n flow_up = 0\n for level in range(self.n_levels - 1):\n lower_prev = state_prev[level*self.n_nodes:(level+1)*self.n_nodes]\n higher_curr = state_curr[(level+1)*self.n_nodes:(level+2)*self.n_nodes]\n \n # Simplified MI using correlation\n if len(lower_prev) > 0 and len(higher_curr) > 0:\n min_len = min(len(lower_prev), len(higher_curr))\n corr = abs(np.corrcoef(lower_prev[:min_len], \n higher_curr[:min_len])[0, 1])\n if not np.isnan(corr):\n flow_up += corr\n \n information_flow.append(flow_up)\n \n return np.array(information_flow)\n \n def simulate_conscious_prediction(self, duration=15.0):\n \"\"\"Simulate consciousness emerging from predictive processing\"\"\"\n t = np.linspace(0, duration, 1500)\n \n # Dynamic sensory input (complex waveform)\n sensory_input = np.array([\n [0.5 * np.sin(2 * np.pi * 0.1 * time) + \n 0.3 * np.cos(2 * np.pi * 0.15 * time) +\n 0.2 * np.sin(2 * np.pi * 0.05 * time) for _ in range(self.n_nodes)]\n for time in t\n ])\n \n # Initial state (small perturbations)\n initial_state = np.random.normal(0, 0.1, self.total_nodes)\n \n # Simulate dynamics\n trajectory = []\n state = initial_state\n \n for i, sensory in enumerate(sensory_input):\n state = odeint(\n self.predictive_dynamics,\n state,\n [t[i], t[i] + (t[1] - t[0])] if i < len(t)-1 else [t[i], t[i]],\n args=(sensory,)\n )[-1]\n trajectory.append(state.copy())\n \n trajectory = np.array(trajectory)\n \n # Analyze consciousness emergence\n coherence = self.compute_predictive_coherence(trajectory)\n info_flow = self.compute_information_flow(trajectory)\n \n # Consciousness as integration of coherence and information flow\n consciousness_level = np.mean(coherence) * np.mean(info_flow)\n \n return {\n 'trajectory': trajectory,\n 'coherence': coherence,\n 'information_flow': info_flow,\n 'consciousness_level': consciousness_level,\n 'interpretation': 'Consciousness emerges from coherent predictive processing'\n }\n\n# Demonstrate predictive consciousness\npredictive_brain = ChaoticPredictiveBrain(n_levels=3, n_nodes_per_level=9)\nresults = predictive_brain.simulate_conscious_prediction()\n\nprint(f\"Consciousness level: {results['consciousness_level']:.3f}\")\nprint(f\"Mean coherence: {np.mean(results['coherence']):.3f}\")\nprint(f\"Mean information flow: {np.mean(results['information_flow']):.3f}\")\nprint(f\"Interpretation: {results['interpretation']}\")This model influences behavior by guiding attention and action selection, preserving causal relevance for consciousness—though epiphenomenalists would argue the predictive processing does the causal work, not the conscious experience of having predictions[^13].
\n\nSome theorists argue for genuine emergence—where higher-level properties like consciousness can exert downward causal influence on lower-level processes. This would make consciousness genuinely causally relevant while acknowledging its dependence on neural activity.
\nStrong emergence remains controversial, as it seems to violate the causal closure of physics. However, weak emergence—where consciousness has novel properties that arise from but don't violate physical laws—might preserve both scientific materialism and mental causation.
\n\nIf epiphenomenalism is true, then coming to believe it is itself an epiphenomenal process. Our conviction about consciousness's causal irrelevance is itself causally irrelevant—a strange kind of self-defeating knowledge.
\nThis creates a peculiar situation: understanding epiphenomenalism might change how we experience our mental lives without changing how we actually live them. We might feel differently about our agency while acting exactly as we did before.
\n\nDoes life have less meaning if consciousness is epiphenomenal? Some argue that meaning comes from conscious experience itself, regardless of its causal status. The beauty of a sunset, the joy of friendship, the satisfaction of understanding—these experiences retain their value even if they don't cause anything.
\nOthers find this unsatisfying, arguing that genuine meaning requires genuine agency. If we're simply along for the ride in our own lives, then life becomes a kind of elaborate movie rather than a participatory drama.
\n\nHow should we treat beings whose consciousness is epiphenomenal? If conscious experience has no causal efficacy, does suffering matter? Should we care about the subjective experiences of others if those experiences don't affect anything?
\nMost epiphenomenalists argue that suffering matters intrinsically, regardless of its causal status. The badness of pain doesn't depend on pain's ability to cause behavior—it depends on the qualitative nature of painful experience itself.
\n\nEpiphenomenalism presents us with a profound puzzle about the nature of mind and reality. It suggests that our most intimate experiences—our sense of self, our feeling of agency, our qualitative encounters with the world—might be elaborate shadows cast by the real causal processes operating beneath the threshold of awareness.
\nYet these shadows are not mere illusions. They constitute the very fabric of human experience, the stage on which all meaning, value, and purpose play out. Even if consciousness doesn't drive the engine of behavior, it provides the experience of the journey.
\nThe debate over epiphenomenalism ultimately reflects deeper questions about the relationship between objective science and subjective experience, between the view from nowhere and the view from here. As we continue to unravel the mysteries of the brain, we must grapple with the possibility that consciousness—the very thing that makes us human—might be nature's most beautiful accident.
\nPerhaps the most remarkable thing about consciousness isn't whether it causes anything, but that it exists at all. In a universe of unconscious matter and energy, somehow, somewhere, something it's like to be has emerged. Whether that something influences the world beyond itself may be less important than the simple, staggering fact that it is.
\nThe whistle of Huxley's locomotive may not power the train, but it announces the journey. And sometimes, the announcement is everything.
\nThis exploration of consciousness as epiphenomena draws from philosophy of mind, neuroscience, and phenomenology to examine one of the most challenging questions in cognitive science. The integrated computational models demonstrate how consciousness might emerge from complex neural dynamics while remaining causally inert—a beautiful byproduct of information processing rather than its driver. As our understanding of the brain deepens, the relationship between consciousness and causation will undoubtedly continue to evolve, potentially resolving—or deepening—the mysteries explored here.
\n", "frontmatter": { "title": "Why consciousness cannot be explained away as 'Epiphenomenal'", "date": "2025-03-14", "author": "Alif Jakir", "description": "A comprehensive exploration of epiphenomenalism—the radical theory that consciousness is merely a byproduct of neural activity, with profound implications for free will, moral responsibility, and the nature of human experience.", "tags": [ "consciousness", "philosophy", "neuroscience", "cognition", "epiphenomenalism", "free will", "philosophy of mind" ], "image": "/images/blog/consciousness_epiphenomena.jpg" }, "parsedFootnotes": [ { "number": "9", "text": "Simons, D.J. & Chabris, C.F. (1999). "Gorillas in Our Midst: Sustained Inattentional Blindness for Dynamic Events"
\n" }, { "number": "10", "text": "Singer, W. (2001). "Consciousness and the Binding Problem"
\n" }, { "number": "11", "text": "Tononi, G. (2008). "Integrated Information Theory"
\n" }, { "number": "12", "text": "Clark, A. (2013). "Whatever Next? Predictive Brains, Situated Agents, and the Future of Cognitive Science"
\n" }, { "number": "13", "text": "Hohwy, J. (2013). "The Predictive Mind: Cognitive Science Meets Philosophy of Mind"
\n" } ], "tableOfContents": [ { "id": "the-birth-of-epiphenomenalism", "text": "The Birth of Epiphenomenalism", "level": 2 }, { "id": "huxley-s-steam-engine", "text": "Huxley's Steam Engine", "level": 3 }, { "id": "the-cartesian-shadow", "text": "The Cartesian Shadow", "level": 3 }, { "id": "the-neural-foundations-of-consciousness", "text": "The Neural Foundations of Consciousness", "level": 2 }, { "id": "the-hard-problem-and-easy-problems", "text": "The Hard Problem and Easy Problems", "level": 3 }, { "id": "neural-correlates-and-causal-impotence", "text": "Neural Correlates and Causal Impotence", "level": 3 }, { "id": "demonstrate-epiphenomenal-consciousness", "text": "Demonstrate epiphenomenal consciousness", "level": 1 }, { "id": "the-libet-experiments-and-the-illusion-of-will", "text": "The Libet Experiments and the Illusion of Will", "level": 3 }, { "id": "the-phenomenology-of-epiphenomenal-experience", "text": "The Phenomenology of Epiphenomenal Experience", "level": 2 }, { "id": "the-richness-of-irrelevance", "text": "The Richness of Irrelevance", "level": 3 }, { "id": "the-binding-problem-and-unified-experience", "text": "The Binding Problem and Unified Experience", "level": 3 }, { "id": "demonstrate-oscillatory-binding-and-consciousness-emergence", "text": "Demonstrate oscillatory binding and consciousness emergence", "level": 1 }, { "id": "implications-for-free-will-and-moral-responsibility", "text": "Implications for Free Will and Moral Responsibility", "level": 2 }, { "id": "the-dissolution-of-agency", "text": "The Dissolution of Agency", "level": 3 }, { "id": "the-experience-of-choice", "text": "The Experience of Choice", "level": 3 }, { "id": "contemporary-debates-and-challenges", "text": "Contemporary Debates and Challenges", "level": 2 }, { "id": "the-causal-exclusion-problem", "text": "The Causal Exclusion Problem", "level": 3 }, { "id": "the-evolutionary-puzzle", "text": "The Evolutionary Puzzle", "level": 3 }, { "id": "the-knowledge-argument-revisited", "text": "The Knowledge Argument Revisited", "level": 3 }, { "id": "neuroscientific-evidence-and-challenges", "text": "Neuroscientific Evidence and Challenges", "level": 2 }, { "id": "split-brain-studies-and-consciousness", "text": "Split-Brain Studies and Consciousness", "level": 3 }, { "id": "blindsight-and-unconscious-processing", "text": "Blindsight and Unconscious Processing", "level": 3 }, { "id": "anesthesia-and-consciousness", "text": "Anesthesia and Consciousness", "level": 3 }, { "id": "the-phenomenological-response", "text": "The Phenomenological Response", "level": 2 }, { "id": "the-irreducible-first-person-perspective", "text": "The Irreducible First-Person Perspective", "level": 3 }, { "id": "the-hard-problem-persists", "text": "The Hard Problem Persists", "level": 3 }, { "id": "practical-implications-and-applications", "text": "Practical Implications and Applications", "level": 2 }, { "id": "clinical-considerations", "text": "Clinical Considerations", "level": 3 }, { "id": "artificial-intelligence-and-machine-consciousness", "text": "Artificial Intelligence and Machine Consciousness", "level": 3 }, { "id": "educational-and-therapeutic-applications", "text": "Educational and Therapeutic Applications", "level": 3 }, { "id": "alternative-theories-and-synthesis", "text": "Alternative Theories and Synthesis", "level": 2 }, { "id": "panpsychism-and-information-integration", "text": "Panpsychism and Information Integration", "level": 3 }, { "id": "demonstrate-integrated-information-and-epiphenomenalism", "text": "Demonstrate integrated information and epiphenomenalism", "level": 1 }, { "id": "analyze-consciousness-across-different-network-configurations", "text": "Analyze consciousness across different network configurations", "level": 1 }, { "id": "predictive-processing-and-the-bayesian-brain", "text": "Predictive Processing and the Bayesian Brain", "level": 3 }, { "id": "demonstrate-predictive-consciousness", "text": "Demonstrate predictive consciousness", "level": 1 }, { "id": "emergence-and-downward-causation", "text": "Emergence and Downward Causation", "level": 3 }, { "id": "living-with-epiphenomenal-consciousness", "text": "Living with Epiphenomenal Consciousness", "level": 2 }, { "id": "the-paradox-of-self-knowledge", "text": "The Paradox of Self-Knowledge", "level": 3 }, { "id": "meaning-and-purpose-in-an-epiphenomenal-world", "text": "Meaning and Purpose in an Epiphenomenal World", "level": 3 }, { "id": "the-ethics-of-epiphenomenal-beings", "text": "The Ethics of Epiphenomenal Beings", "level": 3 }, { "id": "conclusion-the-shadow-and-the-substance", "text": "Conclusion: The Shadow and the Substance", "level": 2 } ] }, { "slug": "elide-polyglot-runtime", "title": "What Elide.dev Is Actually Doing", "description": "A polyglot runtime for the 21st century, we find quaint things underneath—Truffle ASTs, native-compiled tooling, in-process inference, and a bunch of other cool shit.", "tags": [ "elide", "runtime", "polyglot", "graalvm", "javascript", "typescript", "python", "performance", "developer-tools", "truffle", "rust" ], "date": "2026-01-20", "formattedDate": "January 20, 2026", "content": "Elide is not just "Bun for the polyglot world." It obscures what's meaningfully novel about the project. Bun is fast Node. Elide is trying to do something structurally different: collapse the process boundary between languages. This is quite a lofty goal, with implications too vast and staggering to mention here.
\nIf you crawl through their codebase you'll be able to see their progress so far in their goals!
\n\nIf you've ever called Python from Node—child_process.spawn, JSON over stdin/stdout, maybe gRPC if you're feeling industrious—you know the tax. Serialization, deserialization, process context switching, the impedance mismatch between type systems. Every boundary crossing costs you. Not just in latency, but in cognitive overhead: you're now debugging across two runtimes with two sets of assumptions about memory, error handling, and concurrency.
The standard industry response is microservices. Need Java's speed, Python's ML ecosystem, and TypeScript's web ergonomics? Three services, three Dockerfiles, three CI pipelines, internal APIs, network hops. It works. It's also expensive, fragile, and architecturally complex in ways that have nothing to do with your actual problem.
\nElide's bet is that this complexity is an artifact of tooling limitations, not a fundamental constraint. And the mechanism they're using to test that bet is GraalVM's Truffle framework.
\n\nTruffle represents every supported language—JavaScript, Python, Ruby—as an Abstract Syntax Tree that the GraalVM compiler can reason about uniformly. This is the key architectural decision, and everything else follows from it.
\nWhen a TypeScript function calls a Python function in a hot loop, the compiler doesn't context-switch between language runtimes. It performs partial evaluation across the combined AST and can inline the Python nodes directly into the TypeScript compilation unit. The boundary between languages dissolves at the machine code level—your polyglot call becomes a monoglot binary.
\nI want to be precise about what this means and what it doesn't. It means that in the optimized steady state, cross-language calls can approach the cost of same-language calls. It doesn't mean the first invocation is fast (JIT warmup still applies), and it doesn't mean every cross-language pattern optimizes well. Truffle's partial evaluation is powerful but not magic—it works best when the compiler can specialize on stable types and predictable control flow.
\nThe deeper implication is architectural. If language boundaries genuinely disappear at runtime, the case for splitting a system into per-language services weakens. You don't need a Python microservice for your ML pipeline and a separate TypeScript service for your API layer—they can live in the same process, share memory, and get optimized together. Whether you should do this is a design question, not a technical limitation.
\n\nThe most technically interesting piece might be Espresso—a JVM implementation written in Java that runs on Truffle. Read that sentence again, because it's doing a lot of work.
\nTraditionally, Java runs on the host JVM. In Elide, Java code runs on top of Truffle as a guest language, on equal footing with JavaScript and Python. This means Java participates in the same cross-language inlining I described above. A Java method can be inlined into a Python lambda, which can be inlined into a TypeScript event handler. Three languages, one optimized compilation unit.
\nThe recursion here is worth noting: a JVM, implemented in Java, running on a compiler framework (Truffle), itself running on a JVM (GraalVM). It's turtles, but the performance characteristics are surprisingly good because Truffle's partial evaluation can specialize through all those layers.
\n\nElide claims a 20x speedup for javac. My first reaction was skepticism—that's a big number. But the mechanism is straightforward once you understand it.
Standard javac is a Java application. Every invocation pays the JVM startup tax: class loading, JIT warmup, garbage collector initialization. For a quick recompile of three changed files, most of the wall-clock time is overhead, not compilation.
Elide provides javac compiled ahead-of-time into a GraalVM Native Image—a standalone binary with no JVM startup. It launches in sub-millisecond time, and the compilation logic is pre-optimized. For small-to-medium changes, this transforms the dev loop from "wait for the JVM to think about starting" to "basically instant." The 20x number is plausible for incremental builds, though it narrows for large full-project compilations where actual compilation dominates over startup.
The same approach applies to TypeScript. Elide integrates OXC, a Rust-based parser, directly into the execution path. When you run elide run app.ts, OXC strips types and produces an AST at native speed. Elide caches this in a binary format, so subsequent runs skip parsing entirely and load the cached AST straight into the GraalJS engine. TypeScript execution overhead converges toward plain JavaScript.
I find the OXC integration more compelling than the javac story, honestly. The TypeScript tooling ecosystem has a real performance problem—tsc is slow, the alternatives are all varying degrees of incomplete—and a native parser that caches aggressively addresses the bottleneck directly.
Elide embeds llama.cpp via a Rust FFI layer directly into the runtime. You load a GGUF model once into process memory, and your application logic calls inference in-process. No network hop, no serialization, no separate Python process managing a model server.
\nThis is the "AI-native" feature, and I have mixed feelings about it. On one hand, the engineering is clean—co-locating inference with application logic eliminates real latency and complexity. For privacy-sensitive applications where data can't leave the process, or for edge deployments where network calls are expensive, this matters.
\nOn the other hand, most production AI workloads call hosted APIs for good reasons: model management, scaling, cost amortization, the ability to swap models without redeploying. In-process inference trades operational flexibility for latency. That's the right trade for some applications and the wrong one for most.
\nI think the interesting use case is not "replace your OpenAI API call" but rather "run a small specialized model alongside your application logic without the overhead of a separate inference service." Local embedding models, classification heads, small reasoning models—tasks where the model is part of the application rather than a service it consumes.
\n\nI'm skeptical of any project that doesn't acknowledge its own limitations clearly, so let me state what I found:
\nThe NumPy problem is real. GraalPy—Elide's Python engine—struggles with C extensions. If your Python code depends on NumPy, Pandas, or SciPy in non-trivial ways, you'll hit compatibility issues or fall back to a much slower emulation mode. This isn't a minor limitation—it excludes a large portion of the Python ecosystem that people actually use. The Truffle approach works beautifully for pure-Python code and falls apart at the C boundary.
\nCold start in JIT mode is heavy. Native Images start fast, but during development you're typically running in JIT mode, which has meaningful cold start overhead. For serverless or short-lived processes, this matters. Elide is better suited to long-running services where JIT warmup amortizes over time.
\nThe documentation lags behind the code. Features like the local-ai module are implemented in Rust and Kotlin but sparsely documented for end users. The codebase is moving fast—which is good—but the gap between "what's possible" and "what's practically accessible" is wide. I spent non-trivial time reading source code to understand things that should have been in a README.
It's beta software. I want to be honest about this because the marketing doesn't always emphasize it. Elide is not ready for production backends handling real traffic. The engineering foundation is strong, but the edge cases, error messages, debugging story, and ecosystem integration are all still maturing.
\n\nThe technical term for what Elide is attempting is architectural collapse—reducing the number of independent runtime boundaries in a system. Instead of N languages requiring N processes requiring O(N²) integration points, you get one process with shared memory and unified optimization.
\nThis is genuinely interesting as an engineering direction, regardless of whether Elide specifically succeeds. The Truffle/GraalVM approach to polyglot execution is, in my assessment, the most technically sophisticated attempt at cross-language interop that exists today. The AST-level inlining is not a marketing claim—it's a real compiler technique with measurable results.
\nBut I want to hold this loosely. We've seen "one runtime to rule them all" promises before. The JVM was supposed to be this. The browser was supposed to be this. WASM is currently supposed to be this. The history of computing is littered with universal runtimes that turned out to be excellent at some things and mediocre at others. Elide might follow the same trajectory—genuinely useful for certain workloads, not the paradigm shift it aspires to be.
\nWhat I'd watch for: whether the NumPy/C-extension problem gets solved (it's fundamental to Python adoption), whether the developer experience catches up to the compiler engineering (right now the gap is large), and whether real-world polyglot applications actually benefit from in-process integration versus the operational simplicity of separate services.
\nThe engineering is good. The vision is ambitious. Whether the two converge into something practically useful—I genuinely don't know. But it's worth paying attention to.
\ncrates/ directory has the Rust/OXC integration, packages/ has the Kotlin/JVM piecesIn 2019, Singapore completed something unprecedented: a digital twin of an entire nation. Virtual Singapore integrates building information models, real-time sensor data, demographic information, and environmental systems into a single platform. Urban planners can simulate evacuation routes. Engineers can model how a new tower will cast shadows. In pilot districts, the system helped reduce energy consumption by 30%.
\nMeanwhile, the European Union is building something even more ambitious: Destination Earth, a digital twin of the planet itself. By 2030, it aims to simulate Earth's climate, ecosystems, and human systems at kilometer-scale resolution, updating continuously with satellite data. The goal is nothing less than modeling the future of our planet to inform policy on climate change, disaster response, and resource management.
\nThese projects represent the apex of a $154 billion industry projected to reshape everything from manufacturing to medicine. Rolls-Royce uses engine-specific twins to predict maintenance needs so precisely that some aircraft engines now operate far beyond standard service intervals—saving fuel, reducing emissions, and preventing failures before they happen. BMW twins entire factories before building them. Philips is developing "digital patients" that model individual human physiology for personalized treatment.
\nThe promise is extraordinary: perfect virtual mirrors of physical reality, continuously updated, predictively intelligent, capable of optimizing systems we could never understand through intuition alone.
\nWhat the press releases leave out: most of this doesn't work the way we're told it does.
\n\nThe National Academies of Sciences provides the most rigorous definition:
\n\n\n"A digital twin is a set of virtual information constructs that mimics the structure, context, and behavior of a natural, engineered, or social system (or system-of-systems), is dynamically updated with data from its physical twin, has a predictive capability, and informs decisions that realize value. The bidirectional interaction between the virtual and the physical is central to the digital twin."
\n
The key words: dynamically updated, predictive, bidirectional. A digital twin isn't a 3D model. It isn't a dashboard. It isn't a simulation you run once. It's a living representation that continuously ingests data from reality, makes predictions about what will happen, and—crucially—can influence the physical system it represents.
\nBy this definition, most "digital twins" deployed today aren't twins at all. They're what researchers call digital shadows (data flows one way, from physical to virtual) or digital models (static representations with no live data connection). The industry conflates these categories constantly, inflating capabilities and obscuring limitations.
\nA 2025 study across engineering domains found that Technology Readiness Levels for digital twins average just 4.8 out of 9—roughly "technology validated in lab." We're not at deployment maturity. We're at promising prototype.
\n\nResearchers call it the reality gap: the divergence between what a simulation predicts and what the physical system actually does. Every digital twin has one. The question is how large, how consequential, and whether it's growing or shrinking.
\nThe gap emerges from several sources:
\nContext mismatch. A twin trained on one set of operating conditions—temperature, load, usage patterns—will drift when conditions change. A bridge monitoring system calibrated for summer traffic behaves differently in winter. A patient twin trained on clinical trial data may not generalize to real-world populations with different demographics, comorbidities, and behaviors.
\nPhysics we didn't model. Every simulation makes simplifying assumptions. We model the beam but not the rust. We model the flow but not the turbulence at the edges. We model the organ but not the way it moves when the patient breathes. These omissions compound. A case study of material transfer systems in remanufacturing found that minor differences in collision geometry modeling—the shape of contact surfaces, the friction coefficients—produced major discrepancies between virtual and physical behavior.
\nLatency and scale. High-fidelity simulation requires enormous computation. Real-time response requires speed. You can't have both. A twin that takes an hour to predict what happens in the next minute is useless for safety-critical applications. The tradeoffs are brutal: reduce resolution, simplify physics, or accept that your "real-time" twin is actually running on yesterday's data.
\nData quality. Twins are only as good as their sensors. Sensors fail, drift, get occluded, report noise. Networks drop packets. Edge computing introduces latency. The chain from physical reality to virtual representation has dozens of failure points, and each one widens the gap.
\n\nThe most interesting recent research doesn't promise to close the gap. It acknowledges the gap and builds systems that can detect, measure, and adapt to it.
\nReality Gap Analysis (RGA). Researchers at Carnegie Mellon developed a module that continuously monitors the divergence between a digital twin's predictions and real sensor data. When the gap exceeds a threshold, the system triggers recalibration. They tested it on a steel truss bridge in Pittsburgh: the twin detected context shifts—temperature changes, traffic pattern variations—and adjusted its internal model accordingly. The gap didn't disappear, but it stopped growing.
\nSemantic Digital Twins. A 2025 paper on LLM-augmented twins addresses a different kind of gap: the semantic one. Traditional twins know physics but not regulations, guidelines, or domain expertise encoded in documents. By integrating large language models with twin architectures, the researchers built systems that understand not just "how does this structure behave" but "what are the regulatory constraints on modifying it." They demonstrated this for offshore wind farm planning—a domain where engineering, environmental law, maritime regulations, and local policy all intersect.
\nNeuro-Symbolic Reasoning. The ANSR-DT framework combines neural networks (good at pattern recognition) with symbolic reasoning (good at logic and rules) and reinforcement learning (good at adaptation). The result is a twin that can both learn from data and explain its reasoning—critical for high-stakes domains where "the AI said so" isn't an acceptable justification.
\nHardware Acceleration. For applications where milliseconds matter—collision avoidance, surgical robotics, power grid stability—researchers are pushing twins onto specialized hardware. A 2025 study implemented neural twin components on FPGAs, achieving response times five times faster than human reaction time. This matters when the gap between prediction and reality has to be closed in real-time, continuously.
\nGenerative Twins. Perhaps most striking: researchers have begun building twins that can design themselves. Using vision-language models trained on 120,000 prompt-sketch-code triplets, they demonstrated systems that convert rough layout sketches and natural language descriptions into executable simulation code. The twin doesn't just mirror reality—it generates the mirror.
\n\nWhat the technical literature underplays: the digital twin gap isn't just about fidelity, latency, or data quality. It's about power.
\nA digital twin is a representation. Representations are never neutral. They encode decisions about what matters, what gets measured, what counts as success. When GE builds a wind farm twin, it optimizes for energy output and equipment longevity—metrics that serve GE's interests. The twin doesn't model community impact, noise pollution, bird mortality, or the aesthetics of the landscape. Those aren't in the objective function.
\nWhen a hospital builds a patient twin, it models physiological parameters that can be measured, quantified, and acted upon by medical professionals. It doesn't model the patient's social support network, their financial stress, their trust in the healthcare system—factors that profoundly influence health outcomes but resist quantification.
\nThe gap isn't just between simulation and reality. It's between the reality that gets represented and the reality that doesn't.
\n\nVirtual Singapore is built by government agencies using public funds. The data it contains—building specifications, traffic patterns, energy usage—comes from citizens and businesses. But who can access the twin? Who can query it? Who can build applications on top of it?
\nThe Digital Twin Consortium's 2025 whitepaper on aerospace and defense identifies interoperability and data governance as critical gaps. Different contractors use different platforms, different data formats, different simulation engines. Twins can't talk to each other. Data gets siloed. The result: rather than a unified digital representation of a system, you get a collection of incompatible fragments, each owned by a different vendor, each optimized for a different purpose.
\nThis isn't a bug. It's a business model.
\n\nIn medicine, the digital twin gap becomes visceral. A 2025 study in JMIR examined barriers to clinical digital twins and found that the primary obstacles aren't technical—they're institutional. Data sits in silos controlled by hospitals, insurers, and device manufacturers with no incentive to share. IRB processes designed for clinical trials struggle to accommodate continuously-updated AI systems. Liability frameworks don't know how to handle decisions made "by the twin."
\nAnd beneath all of this: patients themselves. Whose body is being modeled? Who consents to the modeling? Who benefits from the predictions? The twin that helps an elite teaching hospital optimize surgery schedules may be trained on data from populations very different from the ones it will serve. The gap between the twin and the patient becomes a gap between populations—with predictable consequences for who gets helped and who gets harmed.
\n\nLet me be direct about what the state of the art actually looks like in 2025:
\nManufacturing is the success story. BMW, Siemens, and others have achieved genuine digital twins of production lines—systems that continuously ingest sensor data, predict equipment failures, and optimize throughput. The business case is clear (downtime is expensive), the physics are well-understood (machines behave more predictably than people), and the data infrastructure exists. Even here, a case study of a flexible manufacturing line found that gaps in historical data and system modeling complexity limited what the twin could actually predict.
\nInfrastructure is promising but early. Bentley Systems has impressive case studies: a digital twin of New Orleans' flood-gate infrastructure, modeling of the UK's Severn Tunnel, structural analysis of a Saudi port warehouse under complex soil conditions. These represent real value delivered. But they're also showcase projects with dedicated resources. The typical infrastructure operator—a municipal water authority, a regional transit agency, a power cooperative—lacks the expertise, budget, and data infrastructure to build and maintain twins at this level.
\nSmart cities are mostly hype. Virtual Singapore is real. Most "smart city" digital twins are 3D visualizations with some sensor overlays—useful for presentations, less useful for prediction and optimization. The computational and data requirements for genuine urban-scale twinning are staggering. Few cities can afford them. Fewer still have the governance structures to use them responsibly.
\nHealthcare is a minefield. Patient-specific twins for surgical planning exist and provide value. Population-scale twins for disease modeling are emerging. But the regulatory, ethical, and institutional barriers are enormous. A digital patient that predicts treatment outcomes is also a liability time bomb. The gap between technical capability and deployable system is measured in years, maybe decades.
\nEarth itself remains aspirational. Destination Earth is real work with real funding. But simulating the planet at kilometer resolution with continuous updates is not a 2030 problem. It's a 2040 problem, maybe 2050, maybe never. The promotional materials don't mention this.
\n\nThe question I keep coming back to: do we actually want to close the gap?
\nThe promise of digital twins is perfect knowledge, predictive certainty, optimal control. But perfect knowledge isn't possible for complex systems. Predictive certainty is a fantasy. And optimal control assumes we know what we're optimizing for—which requires value judgments that no simulation can make.
\nA bridge monitoring system that detects micro-fractures before they become failures is unambiguously good. A patient twin that predicts disease progression could save lives—or could entrench medical paternalism and erode patient autonomy. A planetary twin that models climate futures could inform policy—or could become another tool for powerful actors to justify decisions already made.
\nThe gap between the twin and reality isn't just a technical problem. It's a feature that preserves space for human judgment, uncertainty, and the possibility of being surprised by the world. Perfect mirrors reflect only what we already expect to see. Distorted mirrors—imperfect, incomplete, obviously partial—remind us that the model is not the territory.
\nThe question isn't whether we can close the digital twin gap. The question is what we lose when we try.
\nA note on this piece: I've drawn on research from the National Academies, the Digital Twin Consortium, and recent academic literature. I've tried to be precise about what's demonstrated versus what's promised. The field moves fast; some of this will be outdated within a year. But the structural questions—about power, representation, and the limits of modeling—will persist long after today's technical limitations are solved.
\nKey Sources:
\nDraft — work in progress.
\n\n\nAbstract. A regular human self-transcends into something far more complex as complexity grows and biological and technological matter converge—inspired by Akira, Romantic poets, Stanisław Lem’s Summa Technologiae, and stray thoughts. Unlike a purely destructive arc, collective human wisdom presses this being toward peace; it is now more cyborg than anything else. The imagery moves between memories of ordinary humanity and the life of a cybernetic superintelligence, where what was human survives mostly as memory. The narrator is haunted by the wars of the Singularity era…
\n
Right before I became this being that I am, I must remember all the events that came in the fore before.
\n\nThe naivete I had back then, I could not have imagined what there was to come. All that I shared, all that I yearned for, all these things that I held in our small hands, these memories I hold so dear to me. Those fragments of experiential feeling, scenes that I haunt, that I cherish, this I still remember. This I shall never forget. Even with the understanding that I could create any memory now, and I could feel any little thing, these are the things that I cherish, that made me become now something entirely alien to who I was. But these little threads are the core of who I am, and I will hold them dear to the last decay of the last atom of my pattern. And I can glimpse backwards, backwards so far, backwards to the genesis of my form now. I’ll tell a tale of how I became what I am, I’ll tell a story, for I am but a story, I am words, I am form. And when I remember myself too, you’ll remember me.
\n\nThese memories, they come back rushing, no matter where I am, like seared imprints into the throngs of resonance pulsing through my brain. I now understand what I am, what I have come here to do. When I was but flesh and blood I was struck with anxiety, confusion. I crawled through a world I could barely understand, never really finding out why I had encountered all this. The restless paradox, life itself, remained out of my reach. Memories, are they who I am?
\nLet’s search the sky for a while.
\nYou and I.
\nCollide like two stars for a while.
\nYou and I.
\nPetrichor floods me so deeply wafting out of storm soaked grass as I ponder these boundless and sublime tragedies. There is something deeply unsettling that I cannot move myself away from. The futility of my cries against an indifferent world. My senses close off and I feel memories of the sunbeam kissing my fragile face. Bubbles of seafoam against my toes, pressed against the softest sand, with not a care in the world, but the rhythm of an ocean.
\nWho am I but not the softest sand?
\nThe air is so silent now. I remember forever these eternal fragments of my simpler existence. How quaint it felt that those butterflies fell upon my face, I had not known that my brief sojourn as flesh and blood would be just so brief. Or I would have torn the fabric of the world apart to take still that moment and hold it in just a little bit longer to feel a femtosecond more of that glimpse into something beautiful and timeless. The guilt wraps around me for not appreciating all that came to me, but then I was merely a dust mote upon the world.
\n\n[Draft — text to follow.]
\n\nHow much change can one world bring? There was a period in biological history on Earth where—
\n[Draft — to be continued.]
\n\nNobody told me that my life as a human being would be a small prelude of things to come. I remember nothing before my human form, isn’t that strange? That we just come into this world, knowing nothing at all!
\n\n[Draft — text to follow.]
\n\nSwimming
\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n\n[Draft — text to follow.]
\n", "frontmatter": { "title": "In The Wake Of What I Used To Be; In Posthumanity; The Memories Before", "slug": "in-the-wake-of-what-i-used-to-be", "date": "2026-03-30", "author": "Alif Jakir", "description": "Draft prose—self-transcendence, bio–tech convergence, and memory after the Singularity; haunted, peaceful cyborg subjectivity. Akira, Romantic poets, Lem’s Summa Technologiae.", "tags": [ "fiction", "draft", "posthuman", "transhumanism", "Singularity", "memory", "prose poetry" ] }, "parsedFootnotes": [], "tableOfContents": [ { "id": "table-of-events", "text": "Table of events", "level": 2 }, { "id": "glimpsing-backwards", "text": "Glimpsing Backwards", "level": 2 }, { "id": "the-memories-they-come-a-flood", "text": "The Memories They Come A Flood", "level": 2 }, { "id": "remembering-sunbeam", "text": "Remembering Sunbeam", "level": 2 }, { "id": "meditations-upon-cambrian-explosion", "text": "Meditations Upon Cambrian Explosion", "level": 2 }, { "id": "as-a-man-became-machine", "text": "As a Man Became Machine", "level": 2 }, { "id": "oh-i-grew-so-much", "text": "Oh I Grew So Much", "level": 2 }, { "id": "dying-back-to-life", "text": "Dying Back to Life", "level": 2 }, { "id": "kiss-the-earth", "text": "Kiss the Earth", "level": 2 }, { "id": "sapiens-underwater", "text": "Sapiens Underwater", "level": 2 }, { "id": "cosmic-apotheosis", "text": "Cosmic Apotheosis", "level": 2 }, { "id": "the-salt-in-my-wounds", "text": "The Salt in my Wounds", "level": 2 }, { "id": "every-human-there-ever-was", "text": "Every Human There Ever Was", "level": 2 }, { "id": "beautiful-flower", "text": "Beautiful Flower", "level": 2 }, { "id": "the-earth-so-crushes-me", "text": "The Earth So Crushes Me", "level": 2 }, { "id": "from-child-s-eyes", "text": "From Child’s Eyes", "level": 2 }, { "id": "the-grief-of-immortality", "text": "The Grief of Immortality", "level": 2 }, { "id": "infinite-beauty-of-cosmos", "text": "Infinite Beauty of Cosmos", "level": 2 }, { "id": "apollo-her-dionysion", "text": "Apollo Her Dionysion", "level": 2 }, { "id": "the-other-minds-after-me", "text": "The Other Minds After Me", "level": 2 }, { "id": "blooming-you-ll-show-me", "text": "Blooming You’ll Show Me", "level": 2 }, { "id": "memories-of-a-culture", "text": "Memories of A Culture", "level": 2 }, { "id": "the-absurdity-upon-mind-and-matter", "text": "The Absurdity Upon Mind and Matter", "level": 2 }, { "id": "the-cosmos-i-am", "text": "The Cosmos I Am", "level": 2 }, { "id": "wanting-to-become", "text": "Wanting To Become", "level": 2 }, { "id": "the-abyssal-of-thought", "text": "The Abyssal of Thought", "level": 2 }, { "id": "i-m-but-a-dream-of-a-dream", "text": "I’m But A Dream of A Dream", "level": 2 }, { "id": "hyper-self-awareness", "text": "Hyper-Self-Awareness", "level": 2 }, { "id": "that-s-the-way-everything-goes", "text": "That’s The Way Everything Goes", "level": 2 }, { "id": "all-suffering-the-ignorance-machine", "text": "All Suffering The Ignorance Machine", "level": 2 }, { "id": "no-i-love-you-more", "text": "No I Love You More", "level": 2 }, { "id": "so-closely-i-am-so-far-away", "text": "So Closely I Am So Far Away", "level": 2 }, { "id": "for-death-is-no-end-to-me", "text": "For Death Is No End To Me", "level": 2 }, { "id": "a-transhuman-phenomenology", "text": "A Transhuman Phenomenology", "level": 2 }, { "id": "an-explosion-of-qualia", "text": "An Explosion of Qualia", "level": 2 }, { "id": "the-way-you-washed-my-hands", "text": "The Way You Washed My Hands", "level": 2 }, { "id": "a-small-island-nation", "text": "A Small Island Nation", "level": 2 }, { "id": "death-the-famine", "text": "Death The Famine", "level": 2 }, { "id": "the-autonomous-war", "text": "The Autonomous War", "level": 2 }, { "id": "in-the-wake-of-a-hurricane", "text": "In The Wake Of A Hurricane", "level": 2 } ] }, { "slug": "prolegomena-to-the-future-of-mind-in-the-universe", "title": "Prolegomena to the Future of Mind in the Universe", "description": "A draft outline exploring mind, substrates, computation, neurobiology, and possible futures of cognition beyond biological constraints.", "tags": [ "philosophy of mind", "consciousness", "cognition", "neuroscience", "AI", "draft" ], "date": "2026-03-30", "formattedDate": "March 30, 2026", "content": "This is a draft. It’s currently an outline / scratchpad for a longer essay.
\n\nDraft: table of contents / chapter outline only.
\n\nIntroduction: AI, Power, and the Shaping of Collective Intelligence\n1.1. AI as a Political Actor\n1.1.1. The Evolution of AI in Politics\n1.1.2. AI as a Tool for Power Consolidation\n1.1.3. AI’s Role in Shaping Global and Local Political Dynamics\n1.2. Thesis: AI, Epistemology, and Desire Manipulation\n1.2.1. AI’s Impact on Collective Knowledge\n1.2.2. AI and Human Desire: Predicting and Shaping Preferences\n1.2.3. Multi-Stakeholder Politics in the Age of AI\nChapter 1: AI in Governance and Surveillance - Power, Control, and Public Trust\n1.1. Governance Systems and Automated Decision-Making\n1.1.1. The Rise of Automated Bureaucracies\n1.1.2. AI-Driven Public Policy\n1.1.3. Political Consequences of Delegating Power to AI\n1.2. Surveillance and the Erosion of Civil Liberties\n1.2.1. AI-Driven Surveillance Systems\n1.2.2. Corporate Surveillance: Data Harvesting and Behavioral Control\n1.2.3. Public Resistance to Surveillance\n1.3. Desire Manipulation and Social Control\n1.3.1. AI’s Power to Predict and Manipulate Behavior\n1.3.2. Social Credit Systems and the Regulation of Behavior\n1.3.3. AI in Behavioral Nudging\nChapter 2: Epistemology, Public Figures, and the Media Landscape - Collective Intelligence in Crisis\n2.1. Social Media and the Fracturing of Reality\n2.1.1. The Algorithmic Filter Bubble\n2.1.2. The Weaponization of AI Algorithms in Political Campaigns\n2.1.3. The Polarization of Public Knowledge\n2.2. Generative AI and the Corruption of Public Trust\n2.2.1. The Rise of Deepfakes\n2.2.2. AI-Generated Content and Fake News\n2.2.3. Restoring Public Trust in the Age of AI\n2.3. Celebrities, Influencers, and AI in the Attention Economy\n2.3.1. AI-Enhanced Influence: Public Figures and Algorithmic Popularity\n2.3.2. AI’s Role in the Creation of Virtual Influencers\n2.3.3. Political Influencers and AI Amplification\n2.4. Memes, Collective Behavior, and Political Movements\n2.4.1. AI-Generated Memes: Fueling Online Political Movements\n2.4.2. Memetic Warfare and Astroturfing\n2.4.3. Memes as Cultural Weapons\nChapter 3: Disinformation, Astroturfing, and the Dark Arts of AI\n3.1. AI-Driven Disinformation and Deepfakes\n3.1.1. The Evolution of Deepfakes and AI-Generated Media\n3.1.2. Disinformation Campaigns and AI-Generated Narratives\n3.1.3. International Responses to AI-Driven Disinformation\n3.2. Astroturfing, Chatbots, and Advertising Manipulation\n3.2.1. AI in Astroturfing Campaigns\n3.2.2. Chatbots as Political Operatives\n3.2.3. AI in Advertising and Subliminal Political Messaging\n3.3. Counteracting AI Disinformation\n3.3.1. AI-Driven Solutions for Detecting Disinformation\n3.3.2. Public Education Campaigns and Digital Literacy\n3.3.3. Policy Solutions and Regulatory Frameworks\nChapter 4: Corporate Influence and AI’s Role in Desire Manipulation\n4.1. AI, Consumerism, and Personal Data\n4.1.1. AI-Driven Consumer Targeting and Manipulation\n4.1.2. AI and the Creation of Consumer Desires\n4.1.3. Personal Data as the New Oil: Corporate Power in the AI Age\n4.2. Targeted Advertising and Political Influence\n4.2.1. Advertising Algorithms in Political Messaging\n4.2.2. The Fusion of Commercial and Political Agendas\n4.2.3. AI in Shaping Political Preferences: The Ethics of Manipulation\nChapter 5: Global Power Struggles, AI Regulation, and Ethical Boundaries\n5.1. The Global AI Arms Race\n5.1.1. US vs. China: The Battle for AI Supremacy\n5.1.2. AI in Cyber Warfare: Digital Conflicts and National Security\n5.1.3. Autonomous Weapons and AI in Modern Warfare\n5.2. Ethical Boundaries and Global Regulation\n5.2.1. The Ethical Dilemmas of Autonomous AI\n5.2.2. Global Efforts at AI Regulation: EU, US, and China\n5.2.3. Multilateral Agreements on AI: The Road to a Global AI Accord\nChapter 6: Public Perception, Trust, and Resistance\n6.1. Building Trust in AI Governance\n6.1.1. Public Trust in AI Systems: A Fragile Relationship\n6.1.2. Transparent AI: Designing Systems that Enhance Trust\n6.1.3. Engaging Citizens in AI Policy-Making\n6.2. The Role of Civil Society in AI Oversight\n6.2.1. Civil Society’s Pushback Against Corporate AI Domination\n6.2.2. Activist Movements and AI: The Fight for Accountability\n6.2.3. Citizen Engagement in Shaping Ethical AI Futures\nConclusion: Shaping the Future of Politics in the Age of AI\n7.1. Democracies, Civil Society, and AI Governance\n7.1.1. Can Democracies Survive AI’s Rise?\n7.1.2. The Role of Civil Society in Balancing AI Power\n7.1.3. Toward an Inclusive AI Governance Model\n7.2. The Path Forward: Ethical AI and Global Cooperation\n7.2.1. Global Cooperation for Responsible AI Development\n7.2.2. Navigating the Ethical Minefields of AI\n7.2.3. A Call to Action: Building an AI Future that Benefits HumanityIn Lebanon's Shatila camp, Syrian refugees work American hours. Through the night they label images—house, shop, car—tagging the streets where they once lived, perhaps training the vision systems that will return to those streets in drones.
\nThis is artificial intelligence. Not silicon consciousness bootstrapping toward godhood, but displaced people doing piecework, teaching machines to see.
\nThe people who dominate AI discourse rarely mention them. They're focused elsewhere—on a hypothetical superintelligence that might, someday, threaten everyone. The harm already happening, to specific people, for specific reasons, stays invisible.
\nThe framework for that concern emerged from 1990s mailing lists devoted to cryonics, life extension, and escaping the heat death of the universe. Its architect was a self-taught teenager with no credentials in any relevant field. He later wrote Harry Potter fanfiction to propagate his ideas—a novel where the young wizard defeats death through "rationality." Within two decades, this framework had captured billions in funding, shaped policy at the highest levels, and become Silicon Valley's default lens on its own work.
\nHow did ideas from here become dominant? Who benefits from the framing? And if the singularity ever arrives—who will own it?
\nI want to be clear about my uncertainty. I could be wrong. The people making these arguments occupy powerful positions—they run major AI labs, control billions in capital, shape policy conversations. They've built institutions around these ideas. If they're right and I'm wrong, the stakes are existential. That asymmetry haunts me. But intellectual honesty requires me to say what I actually believe, and I believe the superintelligence narrative—as currently constructed—rests on foundations that haven't been adequately examined. And I think examining why this particular narrative has gained such traction, in this particular historical moment, among these particular communities, is as important as examining the arguments themselves.
\n\nLet me start by stating the strongest version of the position I'm skeptical of. This isn't the cartoon version; this is Nick Bostrom's actual argument in Superintelligence, which I've read carefully and take seriously.
\nThe Core Claims:
\nOrthogonality Thesis: Intelligence and final goals are independent. A superintelligent system could have any goal—including goals harmful to humans. There's no reason to assume intelligence converges on human values.
\nInstrumental Convergence: Almost any final goal would lead an intelligent system to pursue certain instrumental goals: self-preservation, resource acquisition, cognitive enhancement, goal-content integrity. These pursuits could conflict with human interests regardless of the system's ultimate purpose.
\nIntelligence Explosion: A system capable of improving its own intelligence could enter a positive feedback loop. Each improvement makes the next improvement easier, potentially leading to rapid capability gains that outpace human understanding and control.
\nThe Treacherous Turn: A sufficiently intelligent system might conceal its true goals until it's powerful enough to pursue them without human interference. We couldn't trust apparent alignment.
\nThese arguments don't require consciousness. They don't require human-like cognition. They require only that optimization processes can become powerful enough to pursue goals in ways we can't predict or control.
\nI find this argument more sophisticated than its critics often acknowledge. But I also find it resting on assumptions I'm not willing to grant.
\n\nThe entire edifice rests on an assumption so fundamental it's rarely stated: that intelligence is a substrate-independent property of information processing. If brains are just computers made of meat, then silicon computers could—in principle—do everything brains do, only faster.
\nI'm not sure this is true. And I notice that this assumption is extremely convenient for the communities that hold it.
\nIf minds are just software, then minds can be copied, upgraded, and optimized. If minds are just software, then the people who build software are—in a profound sense—the architects of future consciousness. If minds are just software, then the pathway from "I write code" to "I might create god" becomes navigable. The computational theory of mind isn't merely a neutral philosophical position; it's a flattering one for engineers, and it makes their work seem cosmically significant.
\nThis doesn't make it false. But it should make us suspicious of how readily it's assumed in communities where everyone benefits from its truth.
\nConsider: I can simulate a hurricane on a computer. The simulation doesn't make anything wet. I can simulate a star. The simulation doesn't generate light or gravity. Simulation captures some properties while missing others. The question is whether intelligence is like the wetness of water (not capturable in silicon) or like the pattern of a hurricane (perfectly capturable).
\nWe don't know. We assume.
\nA distinction that rarely gets made: simulation is not emulation. A simulation models the behavior of a system from an external perspective. An emulation reproduces the mechanism. I can simulate a bird's flight path with equations; to emulate flight, I need to build something that actually generates lift. The superintelligence narrative assumes that simulating intelligent behavior at sufficient fidelity becomes intelligence—that the model somehow crosses over into the thing modeled. This is a metaphysical leap, not an engineering conclusion.
\nThe brain doesn't compute like a digital Turing machine—but it can still achieve Turing-completeness through radically different mechanisms. Neurons don't manipulate discrete symbols in sequential steps; they're analog devices with continuous dynamics, temporal coding, and chemical gradients. Yet the computational power is there, implemented in wetware rather than silicon.
\nConsider the Assembly Calculus, developed by Papadimitriou et al. and extended in the NEMO (Neuronal Models) framework, which is the focus in my research. This research models how neural assemblies—groups of neurons that fire together—perform complex computations through operations like projection, association, and merge. The Assembly Calculus is provably Turing-complete: it can compute anything a Turing machine can compute. But it achieves this through fundamentally different mechanisms—Hebbian plasticity, distinct excitatory and inhibitory neuron types, brain areas with specific connectivity patterns—that bear no resemblance to transformer architectures.
\nThe striking part: these biologically plausible models achieve language processing and acquisition with far less data than transformers require. The massive data hunger of current AI isn't a necessary cost of intelligence—it's a symptom of the wrong architecture.
\nWe actually understand, mathematically, what deep learning is doing—and it's not what the hype suggests.
\nDomingos (2020) proved that every model learned by gradient descent is approximately a kernel machine—including deep networks. This is a mathematical theorem, not a conjecture. Kernel machines are classical methods that "simply memorize the data and use it directly for prediction via a similarity function." Deep network weights are, in Domingos's words, "effectively a superposition of the training examples."
\nThis result directly challenges the dominant narrative about deep learning. As Domingos puts it: "Perhaps the most significant implication of our result for deep learning is that it casts doubt on the common view that it works by automatically discovering new representations of the data, in contrast with other machine learning methods, which rely on predefined features. As it turns out, deep learning also relies on such features, namely the gradients of a predefined function... All that gradient descent does is select features from this space for use in the kernel."
\nThe mystique dissolves. These systems aren't discovering deep truths about the world; they're memorizing training data and pattern-matching against it. The paper also explains why deep networks are so brittle—why their "performance can degrade rapidly as the query point moves away from the nearest training instance." This is exactly "what is expected of kernel estimators in high-dimensional spaces." The behavior that seems mysterious if you think AI is "understanding" becomes perfectly predictable once you realize it's interpolating from memorized examples.
\nWright and Gonzalez (2021) extended this, proving that transformers specifically are infinite-dimensional kernel machines—their dot-product attention operates in a feature space with infinite dimensions. The billions of training examples aren't evidence of approaching general intelligence; they're the cost of approximating functions in infinite-dimensional spaces through brute-force memorization.
\nThis is classical mathematics, not mysterious emergence. Kernel theory has been understood for decades. The "intelligence" of GPT-4 is the same kind of "intelligence" as a well-tuned support vector machine, scaled up enormously and applied to text.
\nThe crucial move is upstream: the architecture encodes human knowledge. Domingos notes that "the network architecture incorporates knowledge of the target function into the kernel"—the structure that makes transformers work for language was designed by researchers, not discovered by the system. The learning algorithm just selects which features from this pre-specified space to use. All the actual insight is in the architecture; gradient descent is doing bookkeeping.
\nIf gradient descent can only produce kernel machines, and kernel machines are fundamentally limited to interpolating from training examples, then the superintelligence scenario requires something beyond gradient descent—something we don't have. As Domingos concludes: "If gradient descent is limited in its ability to learn representations, better methods for this purpose are a key research direction." We're not scaling toward AGI; we're scaling a method with known mathematical limitations.
\n\nIf it's not emerging intelligence, what explains the dramatic AI progress of the last decade?
\nStanford statistician David Donoho offers a compelling alternative in his 2023 paper "Data Science at the Singularity". His thesis: what looks like "AI Singularity" is actually something much more mundane—the maturation of frictionless reproducibility in research practice. Three developments came together:
\nWhen these three combine, Donoho argues, they create a "Frictionless Research Exchange"—a community where researchers can reproduce, modify, and improve on each other's work with essentially zero friction. This is the actual "superpower" driving rapid AI progress:
\n\n\n"The collective behavior induced by frictionless research exchange is the emergent superpower driving many events that are so striking today."
\n
This reframing is devastating to the superintelligence narrative. The rapid progress isn't evidence of approaching AGI—it's evidence that research communities organized around open data, open code, and competitive benchmarks iterate faster. That's a sociological insight, not a technological one. Any field that adopts these practices sees similar acceleration. Protein folding improved dramatically not because AlphaFold approached consciousness, but because CASP competitions had been running for three decades with shared data and clear metrics.
\nDonoho is particularly sharp on the "brutal scaling" narrative pushed by tech hegemons—the idea that throwing more compute at larger models is the only path forward. He points out that returns to scaling follow "breathtakingly bad exchange rates"—power laws close to zero, or even logarithmic. The first "800-pound gorilla" everyone avoids: if brutal scaling is the only path, we can't afford it. Training costs can't scale by another factor of 100 or 1000 as they did last decade.
\nThe second gorilla: "AI proudly, willfully, has no ideas and doesn't want any." The "Bitter Lesson" mentality in AI—that only scale matters, not ideas—worked during Moore's Law. But Moore's Law stopped a decade ago, and "we don't have any concrete path to the next scaling miracle." As Donoho puts it: "Hope is not a strategy."
\nThe leaked Google memo saying "we have no moat" is evidence of exactly this. In a world of frictionless reproducibility, no company can maintain advantage for long—every achievement gets reproduced and improved upon. The hegemons benefit from the perception that they "own" AI, but the real engine is open research practices that anyone can adopt.
\nSome might point to recent work like the Platonic Representation Hypothesis (Huh et al., 2024), which observes that neural networks trained on different data and modalities are converging toward similar representations as they scale. The paper frames this as evidence that models are approaching a shared "statistical model of reality." But notice what this actually says: statistical model, not reality itself. All models are converging to similar ways of compressing the statistical structure of their training distributions—which is exactly what you'd expect if they're all kernel machines learning from overlapping data sources. Convergence in representation space doesn't mean convergence toward understanding; it means convergence toward the same pattern-matching strategy applied to the same underlying data statistics.
\nThe paper's own framing reveals the limitation: models are learning representations of the joint distribution over events that generate observable data. That's a fancy way of saying they're memorizing statistical regularities in what they've seen. This is precisely what kernel machines do. The "platonic representation" isn't Plato's realm of Forms—genuine abstract understanding—it's the fixed point of kernel learning on internet-scale data. Models align because they're all doing the same mathematical operation on overlapping training sets.
\nBiological systems, by contrast, learn from sparse experience because they use finite, structured representations—neural assemblies with specific connectivity patterns, not infinite-dimensional kernel expansions. The Assembly Calculus achieves Turing-completeness with mechanisms that have natural inductive biases for the kinds of learning organisms actually need to do. Transformers achieve impressive interpolation within their training distribution by memorizing statistical regularities. These are fundamentally different approaches, and only one of them resembles anything like cognition.
\nThis matters because the superintelligence narrative assumes that scaling transformers—more parameters, more data, more compute—is the path to general intelligence. But what if intelligence requires something architecturally different? What if we're scaling the wrong paradigm? The existence of alternative approaches that are Turing-complete, more neurobiologically grounded, and vastly more data-efficient should make us deeply skeptical of claims that we're on a path to AGI by making transformers bigger.
\nDigital and analog computation have fundamentally different properties. Analog systems are continuous, noisy, and energy-efficient; digital systems are discrete, precise, and energy-hungry. The brain is analog—or rather, it's a complex system that defies the analog/digital binary entirely, using spike timing, dendritic computation, glial interactions, and who knows what else. The confidence that digital transformers will replicate this by scaling up is a bet, not a theorem.
\nThis isn't just theoretical. Neuromorphic computing—hardware designed to mimic biological neural processing—already exists. Intel's Loihi chip, IBM's TrueNorth, and various academic projects implement spiking neural networks in silicon. These systems are orders of magnitude more energy-efficient than GPUs running transformers. They process information through spike timing and local learning rules rather than global backpropagation. If we wanted to build systems that actually resemble biological cognition, we have alternative paths. The industry isn't pursuing them at scale—not because they don't work, but because the current paradigm is profitable and the infrastructure is already built. The choice to scale transformers is a business decision, not a scientific necessity.
\nWhen I experience the color red, there's something it's like to be me experiencing it. This subjective quality—what philosophers call "qualia"—seems to be something over and above any functional description. You could know everything about the physics and neuroscience of color perception and still not know what red looks like to me. This is the hard problem of consciousness, and after 30 years of discussion, we haven't made progress. The computational response is: consciousness is just what complex information processing feels like from the inside. But this assumes the very thing at issue—that complex information processing necessarily produces subjective experience.
\nThe Chinese Room thought experiment isn't conclusive—Dennett and others have legitimate responses—but it points to something real. There's a gap between syntactic manipulation (moving symbols according to rules) and semantic understanding (grasping what symbols mean). The superintelligence narrative assumes this gap can be crossed computationally. Maybe it can. But we haven't demonstrated it, and the demonstration might be impossible even in principle.
\nWhy does this matter? Because some of what we value about intelligence might be intrinsically connected to consciousness: understanding (not just pattern-matching), creativity (not just recombination), wisdom (not just optimization). If these require consciousness, and consciousness isn't computational, then superintelligence might be impossible in the relevant sense. You could have a very powerful optimization process that lacks the properties we most care about.
\n\nWhen people talk about superintelligence, the word "intelligence" does an enormous amount of unexamined work.
\nWe treat intelligence as a single quantity, like temperature or mass—something that can be measured, compared, and scaled up indefinitely. But this framing has a specific history, and it's not pretty.
\nIQ testing was developed in the early 20th century by eugenicists—people who believed humanity could be improved through selective breeding. The tests were designed to sort people into hierarchies: who could serve in which military roles, who deserved to reproduce, who should be institutionalized or sterilized. "General intelligence" (the g factor) became dogma not because scientists discovered a single substrate of cognition, but because a single number was useful for ranking humans. The claim that cognition reduces to a scalable dimension isn't a discovery—it's an invention, and it was invented to justify hierarchies.
\nThis isn't ancient history. The superhuman AI narrative is the logical extension of the same ideology: if intelligence is a single axis, and some people have more of it than others, then something could have more of it than all of us. The hierarchy continues upward. TESCREALism doesn't just inherit transhumanist language from eugenics; it inherits the core move of hierarchizing minds and treating the "lower" ones as raw material for the "higher."
\nConsider what this framing excludes: my grandmother, who never went to college, navigates complex family dynamics with a sophistication I couldn't match after decades of trying. A friend who solves differential equations in his head can't read a room to save his life. My cat understands physics (catching prey mid-air) while being confused by mirrors. These aren't just different amounts of some underlying thing. They're different kinds of cognitive capacity, shaped by evolution for specific purposes, implemented in specific biological architectures, and deeply dependent on social context and lived experience.
\nThe people we call "intelligent" are usually people who excel at the particular cognitive tasks our institutions have decided to measure and reward. And those institutions were built by people who benefited from exactly that framing.
\nWhen we say AI might become "more intelligent than humans," what exactly are we claiming? More intelligent at:
\nThe superintelligence narrative treats intelligence as though it has a single axis that can be scaled. But cognition might be more like "athletic ability"—a loose cluster of distinct capacities with different constraints and different scaling properties. Usain Bolt's sprinting doesn't scale to swimming. Skill in theoretical physics doesn't scale to navigating a marriage. Excellence in one domain often comes at the cost of others, not as a foundation for them.
\nThis doesn't mean AI won't be dangerous. But it suggests the "intelligence explosion" model—where each improvement enables the next—might hit barriers specific to each cognitive domain. And it should make us ask: whose interests are served by pretending otherwise?
\n\nI can't write about superintelligence in 2024 without addressing what's actually happening in AI.
\nGPT-4 and Claude are remarkable. I use them daily. They pass bar exams. They write competent code. They engage in conversations that feel genuinely intelligent. This is evidence I have to take seriously.
\nBut notice the shape of the improvement: they're getting better at things that can be represented in text. They're getting better at pattern-matching within training distributions. They are not—as far as we can tell—getting better at:
\nWhen GPT-4 "reasons," it's drawing on patterns it's seen. When a mathematician reasons, they're doing something that—at least phenomenologically—feels different. Whether this difference is fundamental or just a difference of degree is exactly what's at issue.
\nThe scaling hypothesis says: keep adding parameters and training data, and eventually you get AGI. Maybe. But we've seen scaling laws plateau before (Moore's Law is slowing). We've seen capabilities that looked exponential turn out to be S-curves. The prediction that scaling will produce superintelligence is a prediction, not an observation.
\n\nLet me steelman the opposition more thoroughly. The most compelling recent evidence for the superintelligence concern comes from two areas:
\nEmergent capabilities: Research from Google and others has documented capabilities that appear suddenly at scale—abilities that weren't present in smaller models and weren't explicitly trained. Three-digit addition, chain-of-thought reasoning, certain forms of in-context learning. The argument is: if capabilities can emerge unpredictably, then dangerous capabilities might emerge unpredictably too. We might not see superintelligence coming.
\nMechanistic interpretability: Anthropic and others are doing careful work to understand what's happening inside these models. They've found that neural networks develop internal representations—"features"—that correspond to meaningful concepts. The models seem to be building something like world models, not just memorizing patterns.
\nI take this seriously. Two things keep me skeptical:
\nThe "emergence" framing is contested. Recent work by Schaeffer et al. (2023) argues that emergent capabilities are often artifacts of how we measure performance—discontinuous metrics create the appearance of sudden jumps when the underlying capability is improving smoothly. When you use continuous metrics, the "emergence" often disappears. This doesn't mean capabilities aren't increasing, but it suggests the discontinuities might be less dramatic than claimed.
\nAs for interpretability: finding that models develop internal representations doesn't tell us those representations constitute understanding. The kernel machine result still holds. Models are still interpolating from training data—they're just doing so through a richer feature space than we initially expected. A "world model" that's built entirely from statistical regularities in text isn't the same as a world model built from causal reasoning and embodied experience. It might be useful for prediction within the training distribution while being completely unreliable for genuine novel situations.
\nI could be wrong. The next generation of models might do things that force me to revise. I'm holding my position loosely.
\n\nEven granting that AI could become superintelligent, I'm skeptical of the intelligence explosion scenario.
\nThe idea is: once AI is smart enough to improve its own code, it enters a positive feedback loop. Each improvement makes the next improvement easier. Within days or hours, it goes from human-level to incomprehensibly superhuman.
\nBut consider: to improve yourself, you need to understand yourself. To understand yourself, you need to be smarter than yourself. At best, you can improve the parts of yourself you understand—but those might not be the bottlenecks.
\nHumans have been trying to enhance human intelligence for millennia. We have schools, books, meditation practices, nootropics. We've improved—but not explosively. Each advance makes the next advance harder, not easier. We find ourselves on an S-curve, not an exponential.
\nWhy would AI be different? Maybe because AI can modify its own code directly, while humans can't rewrite our neurons. But code modification faces its own constraints:
\nThe scenario where AI goes from human-level to galaxy-brain in an afternoon requires a very specific optimization landscape—one that slopes consistently upward with no false summits, no deceptive gradients, no combinatorial explosions of test cases. We have no reason to believe the landscape looks like that. Most optimization landscapes don't.
\nThere's something revealing about how confidently the TESCREAL communities embrace the bootstrap scenario. These are communities that prize pure reasoning from first principles, that built identities around being "more rational" than mainstream institutions, that believe careful thinking can solve any problem. The intelligence explosion is what you'd expect if you believed reasoning could simply think its way out of any constraint. It's the philosophical position of people who've never hit a wall that more thinking couldn't dissolve.
\nEngineers who've actually worked on complex systems tend to be more skeptical. They've learned that adding features creates bugs, that optimization has diminishing returns, that the last 10% of performance takes 90% of the effort. The confidence in explosive self-improvement comes from people building theories about AI, not people debugging code at 2am.
\nSo far I've been treating the superintelligence narrative as a set of arguments to be evaluated on their merits. But arguments don't exist in a vacuum. They arise from particular communities, serve particular interests, and gain traction for particular reasons. The next question is: who built this narrative, and why has it succeeded?
\n\nComputer scientist Timnit Gebru and philosopher Émile P. Torres have given us a useful framework for understanding where superintelligence discourse comes from. They coined the acronym TESCREAL: Transhumanism, Extropianism, Singularitarianism, Cosmism, Rationalism, Effective Altruism, and Longtermism.
\nThese aren't separate movements—they're an interconnected ideological bundle with shared origins, shared funders, and shared institutional bases. The superintelligence narrative didn't emerge from mainstream AI research, cognitive science, or neuroscience. It emerged from this specific milieu.
\nGebru and Torres trace these ideologies to 20th-century eugenics—the belief that humanity can be improved through selective breeding and technological enhancement. This isn't guilt by association; it's intellectual genealogy. The transhumanist dream of transcending human limitations, the longtermist fixation on future generations over present suffering, the rationalist confidence in quantifying and optimizing human values—these carry forward specific assumptions about which lives matter and why.
\nTorres is particularly worth listening to here because he was inside these movements before becoming a critic. He identified as a longtermist; he took the arguments seriously. His critique comes from someone who knows the internal logic intimately—and found it wanting.
\n\nTo understand how we got here, you have to go back to the 1990s mailing lists.
\nThe Extropians list, founded by philosopher Max More in the late 1980s, was ground zero for transhumanist thought. Extropianism championed "Boundless Expansion," "Self-Transformation," and "Intelligent Technology"—fighting entropy through technological transcendence. The list attracted futurists, cryonics enthusiasts, and early tech entrepreneurs. It was here that ideas about intelligence enhancement, life extension, and technological singularity first crystallized into a coherent worldview.
\nEliezer Yudkowsky entered this world as a teenager in the late 1990s. Self-taught, with no formal credentials in AI or any other field, he became a prolific contributor to discussions about machine superintelligence. In 2000, with funding from internet entrepreneurs Brian and Sabine Atkins, he founded the Singularity Institute for Artificial Intelligence (later renamed MIRI, the Machine Intelligence Research Institute). He was 20 years old.
\nWhat's remarkable about this origin story is what's missing: peer review, academic credentials, empirical research programs. The ideas developed in a hothouse of enthusiasts, reinforced by shared assumptions rather than tested against external reality. Yudkowsky's writings—collected in the "Sequences" on LessWrong—blend philosophical speculation, amateur decision theory, and what critics have called "soft science fiction." He also wrote Harry Potter and the Methods of Rationality, a fanfiction novel that doubled as rationalist propaganda, blurring the line between fiction and philosophy entirely.
\nSpend time in these communities and a pattern emerges: intellectual theatre rather than intellectual work. The style mimics rigor—numbered premises, expected utility calculations, Bayesian updating—but the substance rarely survives contact with actual experts. Claims about consciousness, computation, and intelligence that would get shredded in a philosophy seminar circulate as established fact. Novel "decision theories" are invented without engaging the existing literature. Thought experiments substitute for empirical research. The performance of rationality replaces its practice.
\nThis isn't to say everyone in these communities is foolish—there are genuinely engaged people involved. But the institutional structure rewards seeming rigorous over being rigorous. When your audience is other autodidacts who share your assumptions, you never hit the friction that sharpens thought. The result is an elaborate edifice of ideas that feel profound to insiders but look amateurish to outsiders with actual domain expertise. Try getting a MIRI paper through peer review at a top AI conference. Try publishing Yudkowsky's decision theory in a philosophy journal. The ideas don't port because they were never stress-tested against anything but themselves.
\nThen came Peter Thiel.
\nThe PayPal co-founder and early Facebook investor has been one of the most significant funders of the TESCREAL ecosystem. Through the Thiel Foundation, he's given over $1 million to MIRI, $3.5 million to the Methuselah Foundation (life extension research), and $1.25 million to the Seasteading Institute (autonomous ocean communities). Thiel's interest in transcending human limitations—including his reported interest in parabiosis, the transfusion of young blood—reflects the transhumanist core of his worldview.
\nThiel is also associated with the Dark Enlightenment or neoreactionary movement, which critiques democratic governance and fantasizes about tech-CEO monarchies. While the rationalist and neoreactionary communities are distinct, they overlap in personnel, platforms, and certain shared assumptions about human hierarchy and the dispensability of democratic norms. Curtis Yarvin (Mencius Moldbug), a key neoreactionary thinker, found his early audience in the same Bay Area tech circles that produced LessWrong.
\nThis is the petri dish from which superintelligence discourse emerged: a small, insular community of self-reinforcing enthusiasts, funded by billionaires with idiosyncratic ideological commitments, developing elaborate theories untethered from mainstream scientific practice.
\n\nThe communities that produce superintelligence discourse share specific sociological features:
\nAbstract reasoning detached from empirical feedback: These communities excel at building elaborate models and following chains of logic. But the real world rarely provides clear tests of their predictions, so the models can become self-reinforcing.
\nA particular relationship to elite institutions: Often adjacent to but critical of academia. They built alternative institutions (LessWrong, the Future of Humanity Institute, MIRI) that function as credentialing bodies within the community while remaining peripheral to mainstream scholarship.
\nStrong in-group identity formation: Believing these ideas marks you as part of the community. Skeptics aren't just wrong—they "don't get it," they lack the intellectual courage to face hard truths, they're failing to take ideas seriously.
\nMaterial resources from a narrow base: Open Philanthropy, tech billionaire personal foundations, and—until its collapse—FTX. Sam Bankman-Fried's billions flowed to effective altruist and longtermist causes. When the money came from fraud, the movement had to reckon with what that meant about its vetting processes and incentive structures.
\nThe revealing part: both AI "doomers" (who think AI will destroy humanity) and "accelerationists" (who think AI will save it) share the TESCREAL framework. They disagree about outcomes but agree on premises:
\nWhether you land on "we must slow down AI" or "we must speed up AI," you've already accepted that superintelligence is coming and that it will be decisive. The disagreement is tactical, not fundamental.
\nGebru and Torres argue that both positions function to justify the same thing: unlimited AI development by tech companies. Doomers say "we need to build it first so we can align it." Accelerationists say "we need to build it faster to reach utopia." Either way, the conclusion is: keep building.
\n\nSeveral scholars have noted that TESCREAL functions like a secular religion. It has:
\nThis isn't a criticism of religion per se. But it helps explain why the arguments feel so compelling to insiders and so strange to outsiders. We're not just debating empirical claims about AI capabilities—we're navigating a meaning-system, a way of understanding what matters and why.
\nI say this not to dismiss the arguments—I've tried to engage them on their merits—but to note that the superintelligence narrative fits these communities like a key fits a lock. It rewards exactly the kinds of thinking they've built identities around. It provides existential stakes that justify their lifestyle choices. It makes their particular expertise feel cosmically important.
\nThe question isn't whether the people involved are "smart"—that framing accepts the very premise I'm questioning. The question is whether the arguments hold up outside the hothouse conditions that produced them.
\n\nIn a remarkable 2023 paper, Neşe Devenot—a researcher at Johns Hopkins—argues that TESCREALism functions as what she calls "the court philosophy of the global oligarch class":
\n\n\n"Just as earlier court philosophers articulated the divine right of kings to naturalize monarchy, present-day billionaires are funding TESCREAList philosophers and 'thought leaders' to articulate ethical justifications for extreme inequality under oligarchic rule."
\n
This is the sharpest framing I've encountered. The superintelligence narrative isn't cutting-edge philosophy—it's ideology dressed as philosophy, serving the same function as divine right theory did for monarchies. It makes extreme concentration of power seem not just acceptable but cosmically necessary.
\nDevenot connects this to what political theorist Achille Mbembe calls necropolitics: the power to determine which lives are disposable. Drawing on work by Keith Williams and Suzanne Brant (scholars of Haudenosaunee ancestry), she contrasts TESCREAList longtermism with Indigenous concepts of intergenerational responsibility:
\n\n\n"Indigenous views on the wellbeing of future generations are commonly rooted in a non-hierarchical ontology based on reciprocity and relationality.... By imposing a hierarchy wherein some forms of life are ascribed greater value and meaning than others, neoliberalism justifies the instrumentalization of those at the bottom of that hierarchy by those at the top."
\n
This distinction is crucial. When longtermists invoke "future generations," they're not talking about the Honorable Harvest principle that Robin Wall Kimmerer describes—caring for the land so it remains rich for the seventh generation. They're talking about birthing posthuman consciousness that transcends biological limitations. The extinction of Homo sapiens is, on some versions of this view, an acceptable or even desirable outcome. Present suffering—in the Global South, among the working class, among non-human life—becomes acceptable sacrifice for cosmic transcendence.
\nSam Altman's involvement in psychedelic medicine company Journey Colab illustrates the extraction logic. The company emphasizes "Indigenous reciprocity" while pursuing FDA approval for mescaline, a substance with origins in Indigenous ceremonies. But as Altman describes it: "Those [Indigenous] communities will share with Journey what they know of the history of these medicines, and Journey will share what Silicon Valley is good at, with how to use startups and capitalism to deliver something to people who can really benefit from it." Reciprocity, in this framing, means Indigenous knowledge becomes grist for the TESCREAL mill.
\n\nFollowing the money is clarifying:
\nAI companies benefit from the perception that they're building something world-historically important. "We might create superintelligence" is a better pitch to investors than "we're making incremental improvements in pattern matching." OpenAI's structure—a nonprofit that spawned a capped-profit subsidiary that sought billions in investment—makes sense only if you believe the world-historical narrative. Sam Altman learned the superintelligence framing from Yudkowsky's circles before building the company that might (on this view) accidentally destroy the world.
\nAI safety researchers have built careers and institutions on the premise that their work is existential. I don't think most are cynical—they genuinely believe. But the material incentives and the beliefs are mutually reinforcing. When your funding, status, and sense of purpose all depend on a particular narrative, motivated reasoning becomes very hard to avoid.
\nTech billionaires get to be either saviors or prophets. Elon Musk warns about AI risk while building AI companies. Peter Thiel funds both AI development and AI safety research—placing bets on both sides of a game he helped define. The role is flattering regardless of which position you take: you're one of the few people clear-eyed enough to understand what's at stake.
\nThe broader tech industry benefits from a discourse that frames AI as autonomous and inevitable rather than as a set of choices made by corporations for profit. If superintelligence is coming regardless, we might as well have Google or OpenAI build it rather than someone less responsible. The narrative naturalizes what is actually a political choice about resource allocation, labor, and power.
\nThe exploited labor that actually builds AI disappears from the narrative entirely. As journalist Pete Jones reported for Rest of World:
\n\n\n"Forced to adapt their sleeping patterns to meet the needs of firms on the other side of the planet and in different time zones, the largely Syrian population of Lebanon's Shatila camp forgo their dreams to serve those of distant capitalists. Their nights are spent labeling footage of urban areas—'house,' 'shop,' 'car'—labels that, in a grim twist of fate, map the streets where the labelers once lived, perhaps for automated drone systems that will later drop their payloads on those very same streets."
\n
This is the material base of "artificial intelligence"—not silicon consciousness bootstrapping toward transcendence, but refugees doing clickwork for pennies while fleeing wars that the technology they're training might one day intensify. The superintelligence narrative, with its focus on far-future existential risk, directs attention away from these present-tense harms.
\nThe harms aren't hypothetical. They have names and case numbers:
\nCOMPAS — A recidivism prediction algorithm used in criminal sentencing across the United States. ProPublica's 2016 investigation found it was twice as likely to falsely flag Black defendants as future criminals compared to white defendants. Judges used its scores to determine prison sentences. The company that built it claimed proprietary secrecy. This is AI causing measurable harm right now, to specific people, with specific addresses.
\nAmazon's hiring algorithm — In 2018, Reuters revealed that Amazon had built a machine learning tool to review resumes that taught itself to penalize women. The system downgraded resumes that included the word "women's" (as in "women's chess club") and graduates of all-women's colleges. Amazon scrapped it, but only after years of use.
\nClearview AI — A facial recognition company that scraped billions of photos from social media without consent and sold the resulting database to law enforcement. It's been used to identify protesters, track immigrants, and enable stalking. Multiple countries have found it violates privacy laws. The technology exists, is deployed, and is causing harm—while we debate whether hypothetical superintelligence might someday be dangerous.
\nUber's algorithmic management — Drivers are hired, fired, and disciplined by algorithms they don't understand and can't appeal. The opacity is a feature, not a bug: it insulates the company from accountability while extracting maximum labor from workers who have no recourse.
\nThese aren't edge cases. They're the norm. AI is already being used to make consequential decisions about who gets hired, who gets loans, who goes to prison, who gets surveilled, and who gets deported. The systems are opaque, unaccountable, and disproportionately harm marginalized communities. This is the AI safety problem we actually have—and it's not the one the TESCREAL communities are focused on.
\nDevenot calls this pattern "trickle-down ecstasis"—the belief that transforming elite consciousness will eventually benefit everyone. Ronan Levy, former CEO of the psychedelic company Field Trip Health, stated it explicitly:
\n\n\n"Even if you only serve rich white men with access to psychedelic therapies, it's going to change them in a way that I think is constructive and positive.... I know it seems counterintuitive that creating more inequity is going to create equity, but I do genuinely believe that's a possible outcome here."
\n
This is the logic across the TESCREAL bundle: concentrate resources at the top, trust that enlightened elites will trickle benefits downward, and frame any objection as shortsighted failure to appreciate the cosmic stakes. It's supply-side economics dressed in psychedelic language, or in AI language, or in longtermist language—but the structure is the same.
\nThis doesn't prove the arguments wrong. But when a narrative was incubated in 1990s mailing lists, funded by billionaires with specific ideological commitments, developed outside mainstream scientific institutions, and serves the interests of everyone involved in propagating it—we should demand extraordinary evidence before accepting extraordinary claims.
\n\nThe TESCREAL bundle has existed since the 1980s, but superintelligence moved from fringe to mainstream in the 2010s-2020s. Why?
\nDeep learning created a plausibility crisis. After decades of AI winters, systems started doing impressive things. AlphaGo, GPT-3, DALL-E—these made it feel like we were making real progress toward AGI. The TESCREAL communities, who had been saying this was coming for decades, suddenly seemed prescient rather than cranky.
\nMassive capital needed a story. Billions of dollars have flowed into AI. That capital needs to believe it's not just building better autocomplete—it's building the future of intelligence itself. The superintelligence narrative provides that story. Venture capital and the TESCREAL worldview are mutually reinforcing: the narrative justifies the investment, and the investment legitimizes the narrative.
\nThe FTX moment revealed the material base. When Sam Bankman-Fried's empire collapsed, it became clear how concentrated the funding for effective altruism and longtermism had been. One person, with billions of dollars from what turned out to be fraud, had shaped the entire research agenda for existential risk. This should make us ask what other funding dependencies might distort the field.
\nProfessional-class anxiety found an outlet. For the first time, automation threatens not just factory workers but lawyers, doctors, programmers—the people who thought they were safe. Superintelligence discourse lets this class process their anxiety through a framework that feels intellectual rather than merely self-interested. It's easier to worry about extinction than about becoming economically redundant.
\nThe meaning vacuum. In a secular age, the TESCREAL bundle offers something like religious purpose. Working on AI safety isn't just a job; it's potentially saving the world—or the entire future light cone of the universe, if you're a longtermist. This is a powerful attractor for people seeking significance.
\nThe depoliticization of suffering. Mark Fisher identified this dynamic before his death in 2017:
\n\n\n"The chemico-biologization of mental illness is of course strictly commensurate with its depoliticization. Considering mental illness an individual chemico-biological problem has enormous benefits for capitalism. First, it reinforces Capital's drive towards atomistic individualization (you are sick because of your brain chemistry). Second, it provides an enormously lucrative market."
\n
The superintelligence narrative performs the same operation on a grander scale. If our problems stem from insufficient intelligence—whether human or artificial—then the solution is technological rather than political. We don't need to redistribute wealth, dismantle hierarchies, or reorganize the economy. We just need smarter machines, or smarter ways to align smarter machines. The radical imagination is foreclosed: the only futures available are the ones the existing power structure can provide.
\nThe real hallucinations. Drawing on Naomi Klein, Devenot points out that we've been distracted by the wrong hallucinations. When AI systems fabricate citations or invent facts, the industry calls these "hallucinations"—subtly naturalizing the transhumanist fantasy that these systems are conscious beings who experience things. But the actual hallucinations are the promises: AI will end poverty, cure all disease, solve climate change, make jobs more meaningful. These claims "dissociate from the evidence of material conditions" to justify breakneck development by corporate interests.
\nThe solutions to most of our problems are not mysterious. We know what causes mental illness epidemics: inequality, precarity, isolation, environmental degradation. We know what causes climate change: fossil fuel extraction driven by profit motive. We know what would help: material security, community, clean air and water, meaningful work. But these solutions require systemic change that threatens existing power structures. So we're told to wait for a technological messiah instead.
\nI find myself suspicious when an idea fits too perfectly with the moment. Ideas that feel obviously correct to many people often do so because of social context, not evidence. The superintelligence narrative is useful to too many interests for me to trust my intuitions about it.
\n\nThis isn't abstract. The TESCREAL framing is actively shaping governance right now.
\nIn November 2023, the UK hosted the Bletchley Park AI Safety Summit—the first major international gathering on AI risk. The agenda was dominated by existential risk from advanced AI systems. Civil society groups, labor unions, and Global South representatives were largely absent. The framing assumed the problem was future superintelligence, not present-day harms: algorithmic discrimination, exploitative labor practices, environmental costs, concentration of power.
\nThe U.S. Executive Order on AI (October 2023) similarly prioritized "safe" and "trustworthy" AI in language that treats the technology as an autonomous agent to be managed rather than a set of corporate decisions to be governed. The order focuses heavily on dual-use foundation models—the ones frontier labs are building—while saying less about the systems already causing harm.
\nMeanwhile, AI companies have discovered that x-risk discourse provides perfect cover for regulatory capture. By emphasizing the dangers of future superintelligence, they make the case that only they have the expertise to develop it safely. The subtext: don't regulate us into oblivion; we're the responsible ones. OpenAI, Anthropic, and Google all now employ teams of people whose job is to argue that their employers' technology might destroy humanity—while continuing to build it. The argument becomes: we need to be at the frontier to ensure safety. Safety requires scale. Scale requires investment. Investment requires permissive regulation.
\nCompare this to what labor advocates, civil rights groups, and affected communities are actually asking for: accountability for algorithmic discrimination in hiring and lending; transparency about training data and labor conditions; worker protections for the people labeling images and moderating content; environmental disclosure for the staggering energy costs of training runs; antitrust enforcement against the concentration of AI capabilities in a handful of corporations.
\nThese demands get little traction in the policy conversation. They're not existential enough. They concern merely the people being harmed right now, not the hypothetical posthumans of the far future.
\n\nThere's another function the superintelligence narrative serves: justifying a new arms race.
\n"We must build it before China does" has become a mantra in Washington. The argument goes: if superintelligence is coming regardless, and if it will determine the future of civilization, then the United States cannot afford to fall behind. Safety concerns must be balanced against competitive pressures. Slowing down is unilateral disarmament.
\nThis framing is remarkably convenient for AI companies. It transforms corporate interests into national security imperatives. It makes criticism of breakneck development seem naive or even treasonous. And it recycles Cold War logics that the national security establishment finds familiar and compelling.
\nBut notice what the framing assumes: that superintelligence is coming, that whoever builds it first "wins," that the relevant competition is between nation-states rather than between corporations and publics. None of these assumptions survive scrutiny. China is pursuing AI development, but there's no evidence they're closer to AGI than anyone else—because no one is close to AGI. The "race" metaphor implies a finish line that may not exist.
\nMeanwhile, the actual AI competition is for market share, surveillance capability, and military applications—domains where present-day AI is already causing harm. Framing the issue as an existential race distracts from questions we should be asking: Should autonomous weapons be banned by treaty? Should facial recognition be regulated? Should the companies building these systems be broken up?
\nThe x-risk frame says: those questions can wait; we're racing toward godhood. The power-analysis frame says: those questions are urgent precisely because the technology is already being deployed at scale. The geopolitical framing of superintelligence serves the same interests as the domestic framing: unlimited development by incumbent players, with accountability deferred to a future that never arrives.
\nI've spent this essay tracing where the superintelligence narrative comes from, who benefits from it, and how it's shaping policy. But I should also be clear about what I actually believe.
\n\nOn current AI: These systems are powerful tools with significant risks. The risks are mostly mundane: misinformation, manipulation, job displacement, concentration of power, algorithmic bias. These risks are real and worth addressing now.
\nOn AGI: I don't know whether it's possible. If intelligence is substrate-independent, maybe. If it requires something about biological instantiation we don't understand, maybe not. I'm genuinely uncertain.
\nOn superintelligence: I do think it's possible, and it might even be emerging. But it won't look like what Yudkowsky imagines.
\nThe superintelligence scenario pictures an autonomous agent with its own goals, bootstrapping itself to godhood and then pursuing those goals with terrifying efficiency. But why would superintelligence be an agent at all? The more likely scenario—the one already unfolding—is superintelligence as infrastructure: the integration of AI systems into capital flows, logistics networks, surveillance apparatus, and military operations. Not a mind, but a machine—in the older sense of an arrangement of parts that accomplishes work.
\nWho owns the datacenters? Who controls the training data? Who decides what gets optimized and for whom? These questions matter more than the alignment problem. A "misaligned" AI that escapes human control is science fiction. An "aligned" AI that perfectly serves the interests of its owners—accelerating wealth concentration, automating exploitation, optimizing engagement at the cost of mental health, enabling surveillance at scale—is already here.
\nThe superintelligence we should worry about isn't a digital god; it's the emergent intelligence of capital itself, augmented by AI tools, operating at speeds and scales no human can match, pursuing the maximization of returns with perfect indifference to human flourishing. You don't need consciousness for that. You don't need general intelligence. You just need optimization power in the hands of people whose interests diverge from everyone else's.
\nOn existential risk from AI: I think it's non-zero but probably much lower than superintelligence advocates claim—if we're talking about the autonomous agent scenario. The scenarios require many conjunctions: AI becomes superintelligent, AND it has goals misaligned with humans, AND we can't detect this, AND it can take actions we can't prevent. Each conjunction multiplies uncertainty.
\nBut the existential risk from AI-as-infrastructure is substantial and immediate. Climate systems optimized for short-term profit. Labor markets restructured to extract maximum value from humans. Information environments engineered to maximize engagement and minimize solidarity. These aren't hypothetical—they're happening. The risk isn't that AI will "escape" human control. It's that AI will remain perfectly under control—of the wrong people.
\nOn what we should do: Focus on the real, present risks of AI rather than speculative future ones. Build governance structures for the technology we actually have. Resist the urge to treat far-future speculation as a reason to ignore present harms.
\nBut I'm trying to hold these views loosely.
\n\nI'm trying to practice what I preach about holding views loosely. What would update me toward the superintelligence concern:
\nAn AI system that demonstrates genuine novel mathematical reasoning—proving theorems that couldn't plausibly be interpolated from training data.
\nAn AI system that maintains coherent long-term goals across extended interactions without constant steering by humans.
\nAn AI system that accurately models and communicates its own uncertainty, rather than confidently hallucinating.
\nEvidence that scaling continues without diminishing returns at the capability frontier.
\nA theory of consciousness that explains how it emerges from computation, with empirical predictions we can test.
\nIf these things happen, I'll update. I'm not emotionally attached to being right. But until they happen, I remain skeptical—and skepticism implies responsibility. If I don't think the superintelligence framing is useful, what would I put in its place?
\n\nA sketch of a non-TESCREAL approach to AI governance:
\nDemocratic oversight, not expert capture. The people affected by AI systems should have power over how they're deployed. This means worker representation on company boards making AI decisions. Community review boards for high-stakes algorithmic systems. Global South voices in international AI governance—not as token consultants, but as decision-makers. The TESCREAL approach concentrates authority in a small group of self-appointed experts; the alternative distributes it.
\nPresent harms over speculative ones. Redirect resources from existential risk research to the immediate: algorithmic discrimination in hiring, lending, and criminal justice; exploitative labor conditions for data workers; environmental costs of training runs; concentration of power in a few corporations. These problems are tractable. We know what causes them. We can measure whether interventions work.
\nAntitrust enforcement. The AI industry is consolidating rapidly—a handful of companies control the compute, the data, and the talent. This concentration isn't inevitable; it's a policy choice. Breaking up these concentrations, requiring interoperability, and preventing exclusive access to key resources would create more distributed, accountable AI development. The existential risk framing serves incumbents by suggesting that only well-resourced labs can build safely; the alternative is to not require such immense resources in the first place.
\nEnvironmental accounting. Training a single large language model can emit as much carbon as 125 round-trip flights from New York to Beijing. The industry treats this as an externality. Actual accounting—requiring disclosure of energy use, mandating offsets, incorporating environmental costs into decision-making—would change incentives. The longtermist calculus that justifies present harms for future benefits systematically discounts the environment we actually live in.
\nWorker power. The data labelers in Shatila, the content moderators with PTSD, the gig workers without benefits—these are the people actually building "artificial intelligence." Unionization, minimum wage requirements for contractors, mental health support, and classification as employees rather than independent contractors would redistribute the value they create. The TESCREAL vision treats human labor as a waystation to automation; the alternative treats it as something worth protecting.
\nMaterial security. The best way to prevent AI-driven immiseration isn't to align superintelligence; it's to ensure people don't depend on labor markets for survival. Universal basic services—healthcare, housing, education, transit—decouple wellbeing from employment. A strong welfare state makes technological disruption manageable. The reason we're so anxious about AI taking jobs is that in our current system, losing a job means losing access to the conditions of life. Change the system, and the anxiety dissolves.
\nNone of this requires solving alignment. None of it requires predicting what superintelligence would want. It requires confronting power—which is harder, and less flattering to the people currently dominating the conversation.
\n\nWhat I've come to believe: we don't know enough to be confident about superintelligence in either direction.
\nWe don't know if human-level AI is possible. We don't know if consciousness is computational. We don't know if intelligence explosions can happen. We don't know if AI goals would be alien to human values.
\nThe superintelligence narrative presents these uncertainties as resolved, then builds elaborate scenarios on the resolutions. I understand the appeal—uncertainty is uncomfortable, and these are important questions. But false confidence isn't better than acknowledged ignorance.
\nDevenot offers a metaphor I find haunting: the self-awareness paradox. Neuroscientists have speculated about giving psychedelics to patients in comas, hoping to "awaken" consciousness. But as one researcher noted, you might "wake someone up to the reality that two years have passed and their wife has left them"—cognizant of their predicament but unable to change it.
\nApply this to the promise of AI or any individualized treatment for structural problems. Without the material conditions to support change, heightened awareness might only intensify suffering:
\n\n\n"They can't afford healthcare or feed their kids, they're working three jobs around the clock, and their water is full of lead."
\n
This is the trap of technological solutionism: we develop ever-more-sophisticated tools for perceiving problems while leaving the structural causes untouched. The superintelligence discourse exemplifies this. Even if we developed perfectly aligned superintelligent AI, who would own it? Who would benefit? Under current arrangements, it would accelerate existing power concentrations—making a few people unfathomably wealthy while disrupting the livelihoods of billions. The apocalypse wouldn't come from misaligned AI; it would come from perfectly aligned AI serving its owners' interests at the expense of everyone else.
\nWhat I want is epistemic humility matched to the actual state of our knowledge:
\nThe dinner party question—will AI lead to utopia or apocalypse?—is malformed. It assumes a future shaped primarily by technology rather than by politics, economics, and social organization. And it imagines the relevant agent as a digital god rather than the actually-existing superintelligence: capital itself, augmented by AI, operating at inhuman speed and scale, optimizing for returns with perfect indifference to human flourishing.
\nThe real questions are: who controls AI development? Whose interests does it serve? What institutions govern its deployment? What would genuine democratic governance of these technologies look like? These are questions about power, not about intelligence.
\nWe'll muddle through, as we usually do—not because we're wise, but because the future is always more complicated than any narrative can capture. The superintelligence story is seductive precisely because it promises clarity: there's a problem, it has a shape, and certain people (with certain skills, certain training, certain institutional positions) are uniquely qualified to solve it.
\nI don't believe that. I think the challenges we face are distributed, political, and fundamentally about how we organize ourselves—not about building or controlling some hypothetical godlike machine. The solutions we need aren't mysterious: material security, democratic governance, ecological sustainability, community bonds. These don't require superhuman intelligence to implement. They require confronting power—which is precisely why we're offered technological messiahs instead.
\nThe truth is messier and less dramatic than the stories we'd prefer to tell. But I'd rather be honest about the mess than confident about a fiction.
\nI'm genuinely interested in being wrong about this. If you think I've missed something important, I want to know. The stakes are too high for me to prioritize being right over actually understanding.
\nFurther Reading:
\nThe argument I'm skeptical of:
\nCritical perspectives on the TESCREAL bundle:
\nHistory of the movement:
\nPolitical economy of technology:
\nIndigenous perspectives:
\nPhilosophy of mind and consciousness:
\nNeuroscience and alternative approaches to AI:
\nUnderstanding what deep learning actually is:
\nAI skepticism and labor:
\nPresent-day AI harms (the safety problems we actually have):
\n