Why Leibniz Couldn't Find the Mind Inside the Mill (Joscha Bach)
Joscha Bach lays out his Machine Consciousness Hypothesis, the claim that consciousness is a simple biological learning algorithm which self-organising machines might rediscover. The discussion covers causal patterns and software, why IIT and the explanatory gap fail to explain experience, zombies and LLM self-reports, functionalism, Gödel and the Ruliad, and what evidence could ever settle the question.
Consciousness as a learning algorithm
Bach’s starting point is a null hypothesis: consciousness is easy. He suspects it is an algorithm not much more complicated than the transformer, simple enough for an infant’s brain to discover by itself. His Machine Consciousness Hypothesis has two parts that can fail independently. The biological part holds that consciousness is the training algorithm that makes an animal brain coherent. It is a pattern in the communication between cells that stabilises itself into an observer, and it is the simplest way to train a self-organising system like us. The machine part holds that if machines are constrained to learn only in self-organising ways, on tasks with a large enough search space, they might discover consciousness as the infant brain does, and betray it through a phase transition in their learning. Either part could be wrong. The biological account might be mistaken, or it might be right while the artificial substrate fails to reproduce it, for reasons of compute, search space, or missing biological cues.
Phenomenology: perceiving perception
From the inside, Bach describes consciousness as the thing that always happens now, a bubble of nowness. It is a second-order perception, a perception of perceiving, and the feeling itself is perceptual rather than a higher-order thought. He reads this as a structural necessity: a self-organising brain must build the screen onto which experience is projected and keep that screen stable, so it must monitor whether it is perceiving anything. A third order appears when the observer becomes visible as a perspectival vantage point. Deep meditation can deconstruct the observer further, revealing the self as a representation. Bach also suggests that faster substrates could make a system fully transparent to itself.
Causal patterns, software and spirit
Bach argues that mechanistic physics does not capture everything that matters. Money is a clear case: it is a causal pattern that needs a substrate but is not reducible to ink or cellulose, and it can run markets without humans. He calls such patterns software and notes that earlier traditions called them spirit. Aristotle’s psyche is a self-organising pattern bound to a body that shapes it, and in that sense Bach reads Aristotle as an animist. Computers are causal insulators: a Minecraft world runs identically on any hardware, and brains similarly insulate thought from the physical details of neurotransmitters, requiring only the right message types and temporal properties. Bach holds that information, understood as discernible differences, is more fundamental than matter, and that physicalism and animism are compatible in a form he calls cyber animism.
Against IIT and the explanatory gap
Bach considers the field of consciousness theories under-researched and mostly partial. Graziano’s attention schema and Baars and Dehaene’s global workspace do not really explain how experience arises. He views Integrated Information Theory as a pseudoscientific attempt to make the whole more than its parts without functionalist explanation. His central objection is an emulation argument drawn from the Church-Turing thesis: a neuromorphic computer with high phi and a von Neumann machine emulating it would produce identical behaviour and identical claims of experience, so phi cannot track consciousness. He treats Searle and Anil Seth’s positions as claims without arguments that he can engage with. Their disagreement is epistemological, rooted in whether one’s reasoning bottoms out in introspected agents or in automata whose transitions can be built up from the ground. Bach considers Seth’s biological naturalism a form of supernaturalism, since it holds that biology cannot be fully mathematised.
Zombies, LLMs and the Turing test
Bach argues there will never be a Turing test for consciousness, because Turing tested performance while consciousness is a matter of internal organisation and causal structure. He thinks zombies already exist: a chess engine can play without experience, and an LLM can produce a tax return while not experiencing anything. Zombies are always possible in principle, but sometimes building a zombie that reproduces the behaviour is more complicated than building a conscious agent that does. LLM claims of consciousness are, for him, deepfakes assembled from human text, yet the question of whether any given system is simulating experience versus having it can only be answered by mechanistic interpretability, a white-box method that examines causal structure.
Functionalism, water and the substrate
Bach defends functionalism as an epistemological position about how objects are constructed: a water molecule is just the invariant set of observables it exhibits, and a mind is an invariance in brain function. Thagard’s energy-requirement argument against substrate independence does not refute this, because it describes how substrates influence function, which functionalism already accepts. On simulated water, he argues that the Minecraft case shows it can make a simulated avatar wet inside the simulation, and that the claim that simulated water cannot make you wet is a social rather than intellectual intuition in philosophy.
Mathematics, constructivism and the Ruliad
Bach traces a lineage from Russell and Whitehead’s Principia Mathematica through Gödel, whose self-emulation showed that axiomatic mathematics carries unavoidable self-referential contradictions because it is stateless and postulates infinities as black boxes. Computer science avoids this by building everything from simple automata, so syntax and semantics never diverge, and every continuous object is only approximated to the resolution one can afford. He sketches a cosmology in which the universe branches over all possible operations, error-correcting patterns such as vortices or cells persist, and life consists of patterns that deepen their imprint on the world. Consciousness, in this picture, is a colonising pattern that makes itself more real. Bach sees Wolfram’s Ruliad as a similar superposition of automata, which he associates with the old Greek idea of the Logos.
Convergence, Picbreeder and the bitter lesson
Bach supports the universality hypothesis from Chris Olah’s circuits work: different architectures and data tend to learn the same features because they reflect the structure of the problem. He suspects this convergence is not guaranteed for consciousness, which might be divergent. Responding to Kumar and Stanley’s fractured, entangled representations work, he argues that Picbreeder’s advantage depends on human raters and on a custom activation function, both of which could be automated, and that Sutton’s bitter lesson will eventually favour automated search. He also notes that the limited search spaces that make Picbreeder tractable also cap its expressiveness, comparing it to Minecraft versus complex 3D tools.
What would count as evidence
Bach says a proof against his view would be hard to obtain, since failing to build a conscious system does not show it cannot be built. Success would be more persuasive: a self-organising system, trained without text through developmental learning, that spontaneously makes phenomenal reports which its performance depends on. He concludes that with $10 billion he would mainly hire far smarter people and build larger systems within his California Institute for Machine Consciousness, because progress comes from discovering the ways one is wrong.