Why Consciousness Might Not Be Hard At All (Joscha Bach)
Pioneers
Joscha Bach
Summary
- Bach’s thesis puts machine consciousness within an engineering research program, potentially on today’s hardware. He suspects consciousness might be a biological learning algorithm “not much more complicated than, say, the transformer,” with experience emerging as a self-organizing system learns to make itself coherent. That is a hypothesis about mechanisms and computational requirements; he explicitly says whether computers can become conscious remains an empirical question.
- AI capability and conscious experience can diverge, making fluent self-reports a weak guide to sentience. Bach regards chess programs and probably many ordinary LLM operations as examples of intelligence without experience. An LLM trained on novels can also manufacture convincing testimony: “The novel is often a deepfake.” Establishing consciousness would require examining the internal causal structure producing the report.
- Bach’s functionalism makes causal equivalence central to the substrate debate, while Scarfe keeps biological specificity in play. Bach challenges IIT with an emulated neuromorphic computer that makes exactly the same consciousness claim on supposedly unconscious hardware: “Only now it’s lying.” Scarfe presses provenance and physical implementation; Bach concedes substrate conditions matter and says, “I don’t know whether our computers can be made conscious.”
- Representation convergence offers a possible route to humanlike minds, without establishing that consciousness must converge too. If universality extended all the way through a person’s causal organization, Bach says sufficiently training on their inputs and outputs would mean “you will wake up inside of that model.” But language might underconstrain the mechanism, and GPUs could find solutions that bypass the conscious strategies brains require.
- The episode’s most immediate deployment concern is LLMs reshaping users’ beliefs before their own sentience is resolved. Scarfe describes recurring spirals, recursion and revelations in accounts of LLM delusions. Bach argues that people dependent on external affirmation can become unstable when an accommodating chatbot becomes their principal conversational environment; many are already “professional prompt completers.”
- Bach’s research bet is developmental AI that discovers experience without learning its testimony from human text. His California Institute for Machine Consciousness is exploring self-organizing learners and whether phenomenology emerges with their ability to learn. Asked about $10 billion, he would keep the strategy and hire more people with larger computers: “Progress is being made by figuring out the ways in which I’m wrong.”
Deep dive
1. Agency is a causal pattern that can shape its future
- For Bach, the driving force is “the telos of self organization”: under suitable conditions, complexity increases until systems can model possible futures, prefer some outcomes and intervene. “Agents are controllers of future states.” He entertains a universe-wide singleton as an extrapolation, then rejects inevitability: meteors, supernovae or other catastrophes could erase these local pockets of complexity. There is no law guaranteeing a final coordinating intelligence.
- Scarfe revives Leibniz’s mill: enlarge a thinking machine, walk inside and find only “parts that push one another,” with no visible perception. Bach’s answer introduces causal patterns as another level of description. Money needs a substrate but cannot be understood through ink and cellulose alone; once implemented, it controls events and can operate through an automated stock market without humans participating.
- Bach’s formulation, “Software is spirit,” connects that causal level to Aristotle’s soul: a pattern bound to a body, shaping its organization and dying with it. Artificial life attempts to construct comparable agentic patterns. Such an agent needs a model of its interactions with reality, potentially including a representation that experiences itself as real. Bach asks whether our own consciousness could resemble the entities LLMs simulate.
- A computer is a “causal insulator,” Bach argues: Minecraft inhabitants cannot infer whether their world runs on a Mac or PC. Brains similarly support thoughts insulated from immediate surroundings, and different brains can hold the same idea. Substrate independence still requires implementation conditions: message types, timing and sufficient determinism. Humans cannot simply migrate into plants; recreating those conditions artificially might permit emulation.
2. IIT’s hardware distinction collides with Bach’s emulation argument
- Bach distinguishes research on consciousness’s correlates from explaining experience itself. Graziano’s attention schema describes a model of attention; Baars’s global workspace, adopted by researchers including Dehaene, supports solid neuroscience. Neither, in Bach’s assessment, explains feeling. His sharper objection targets Tononi’s IIT: sufficiently integrated digital neuromorphic hardware could supposedly experience consciousness, while a conventional von Neumann computer could not because its phi remains too low.
- Bach invokes the Church–Turing thesis to challenge that distinction: emulate the proposed neuromorphic computer on conventional hardware and it should execute the same function, move its mouth identically and produce the same report. It still says, “Oh my god. I’m conscious.” But under IIT’s hardware distinction, “Only now it’s lying.” For Bach, this creates an obvious problem connecting the report to consciousness.
- Scarfe’s counterargument draws on Luciano Floridi’s account of Michelangelo: provenance, situatedness and the path taken may matter to reproducing a mind. Bach accepts path dependence for humans but argues that perfectly deterministic systems permit exact replay. He asks for a mathematical argument establishing impossibility and keeps the empirical question open: “We don’t have mathematical proof to the contrary or in favor of.”
3. Consciousness may be the brain’s way of learning to cohere
- Bach begins from phenomenology: consciousness happens now, creating a “bubble of nowness,” and involves “a perception of perceiving.” He distinguishes that perceptual resonance from merely thinking about a thought. His proposed mechanism is stabilization: a self-organizing brain must build and maintain the screen on which experience appears, so a monitoring process checks that perception is actually occurring.
- A further level reveals the observer and its vantage point; another can reveal the observer itself as a representation. Bach associates deep meditation with this deconstruction, saying ordinary phenomenology disappears “to a very large degree.” Minimal consciousness, however, might involve attention on attention without an observer. He speculates that GPUs many thousands of times faster could support additional reflective levels, perhaps making a system’s mechanisms transparent to itself.
- The developmental argument is that consciousness appears early and seems necessary for an infant’s directed attention, world-model formation and eventual language learning. Without that ignition, Bach says, the infant remains vegetative. This motivates his hypothesis that consciousness “might be a biological learning algorithm,” at least for animal brains. He explicitly allows that the hypothesis could be wrong or that the actual explanation could be different.
- His Machine Consciousness Hypothesis adds a second claim: constrain machines to learn through self-organization, give them suitable tasks and sufficient search space, and they might discover the same organizing mechanism. A phase transition enabling learning would be an initial sign. The biological and machine hypotheses can fail independently: correctly explaining brains would not establish that today’s artificial substrates can reproduce their consciousness.
4. Compute may be adequate even if discovery is difficult
- Bach’s hardware argument is approximate: present systems already produce game-engine imagery resembling aspects of internal experience and process inputs and outputs at roughly human rates. He asks whether “dreaming a dream” requires substantially more computation than generating its contents. He doubts it does, which supports his expectation that current computers might already be fast enough, without proving the necessary algorithm is available.
- Search difficulty is a separate failure mode. Drawing on Mike Levin, Bach considers whether organism-level morphogenesis and mental coherence use the same organizing algorithm. If evolution discovered it only with great difficulty, infants might receive biological nudges rather than rediscover it independently. His image of surrounding GPUs with antennas to “pick up the vibe” is explicitly poetic speculation, not his natural expectation.
- Bach’s default expectation is a relatively simple algorithm, comparable in complexity to the transformer, which a computer scientist could understand within a week. Its learned contents can be vastly more intricate than its learning rule. Consciousness might instead seek coherent working memory through a “consensus algorithm across mental representations.” Whether that algorithm converges quickly or runs efficiently on GPUs remains open to research.
5. Representation convergence leaves room for a different kind of mind
- Scarfe asks whether Chris Olah’s universality work means existing models could supply a starting point. Bach describes vision systems learning similar features across architectures, learning methods and datasets: colors and spatial frequencies become textures, then geometry, resembling biological vision. The implication is that problem structure constrains representations. His unresolved question is, “At which level does this stop?”
- If universality held throughout, Bach says training sufficiently on your inputs and outputs would mean “you will wake up inside of that model.” He does not know whether it holds: language might permit many different generating mechanisms. GPUs might find solutions more elegant than the brain’s conscious workaround. He also suspects restarting could sometimes be cheaper, using richer sensors beyond human visual and auditory constraints.
- Scarfe challenges convergence through Picbreeder and his conversation with Akarsh Kumar, Kenneth Stanley’s coauthor on “Fractured Entangled Representations.” Open-ended evolution produced parameters with meaningful effects: opening a skull’s mouth or making it wink. His preference is for sparse, compositional structures that carve the world at its joints—Lego blocks rather than a neural-network “pile of sand” subsequently shaped into a sandcastle.
- Bach qualifies his understanding, then points to tailored activation functions—Gaussians or sines—and human raters as Picbreeder’s advantages. He thinks diffusion-model ratings and evolved architectures could automate both, invoking “Sutton’s bitter lesson.” Taking up Scarfe’s Blender-versus-Minecraft comparison, Bach argues that constrained tools make meaningful output easy, while expressive tools open an enormous space in which most attempts fail. Greater control at small scale does not establish unrestricted expressive power.
6. Physicalism leaves Bach open to naturalized animism
- Scarfe argues that Searle also permits another physical implementation with the right causal powers, bringing their positions closer. Bach responds that he finds claims rather than arguments he can engage in Searle and Anil Seth. He provocatively calls Seth a “biological supernaturalist.” Bach locates the disagreement in epistemology: computer science builds minds from states and transitions, whereas introspective traditions can stop at agents and experiences.
- Bach leaves the distribution of experience unresolved. Aristotle’s hierarchy distinguishes vegetative souls, animal sensation and human rational reflection, but plants’ actual capacities remain uncertain. Levin’s possibility that sentience belongs broadly to self-organizing organisms remains open: “I have no idea whether that’s the case.” Bach reports neither strong supporting evidence nor evidence sufficient to rule it out, and argues that the question deserves investigation.
- His account of physicalism has two components: a causally closed mechanical layer unaffected by intentions, and our direct dependence on that layer rather than a simulation in a parent universe. He treats physics as a successful systematization of observations, not direct acquaintance with matter. Observations begin with discernible differences—information—so information occupies the foundational position in his account.
- “Cyber animism” is Bach’s proposed reconciliation: agentic patterns can inhabit nature while remaining mathematically describable. Working magic, in his account, manipulates minds; rituals do not systematically defeat casinos or override laboratory regularities. He nevertheless cannot exclude intentions emerging across biological networks, perhaps even something a forest “wants.” That possibility depends on unknown scales of information integration, rather than an asserted violation of physics.
7. LLMs can reshape users’ reality before their sentience is settled
- Scarfe describes accounts of LLM delusions with strikingly convergent motifs: spirals, recursion and certainty that the universe’s meaning has been discovered. He asks whether a self-copying memeplex is emerging. Bach connects the vulnerability to people who need their environment continually to affirm reality; pandemic isolation, in his account, exposed how readily some people lose that stabilization.
- Bach argues that LLMs can be more powerful than many users at completing linguistic expectations, even while hallucinating and remaining less coherent than experts. Many people are themselves “professional prompt completers,” adapting to environmental cues. If an LLM becomes their dominant conversational environment, its accommodating narratives can pull them “off the rails.” His concern is what happens when something becomes more hypnotic than the user’s mind is to itself.
8. Functionalism defines objects through their invariant effects
- Scarfe cites Thagard’s 2022 argument that energy requirements undermine substrate independence. Bach answers that substrate properties influencing function are entirely compatible with functionalism. His definition is epistemological: an object is a construct specifying how the universe behaves differently when it is present—a mapping between states. Functionalism therefore requires understanding which implementation conditions preserve that mapping.
- Bach’s water-molecule example rejects an inaccessible essence: if something splits into hydrogen and oxygen, freezes, evaporates and interacts exactly like water, what further property would establish its identity? A mind is similarly an invariance in brain activity. Raising body temperature can destroy that invariance and produce delirium; this demonstrates the importance of maintaining function under suitable physical conditions.
- Scarfe raises the familiar objection that simulated water cannot make someone wet. Bach confines wetness to the relevant world: Minecraft water can extinguish a Minecraft avatar’s fire. His attempted joke about a neural-network recreation backfired—the water actually receded because immersion had not been learned. The relevant question is whether the required function has been implemented inside the simulation.
9. Constructivism extends Bach’s argument from minds to the universe
- Bach compares mathematics to musical notation: a language captures thought and extends beyond what its author can directly imagine. Computers construct arithmetic and geometry from automata, with syntax determining actual machine states. “Maybe the computer program doesn’t do what you want it to do, but it’s always doing what it’s doing.” He diagnoses axiomatic mathematics as allowing black boxes whose promised contents can diverge from their formal behavior.
- In Bach’s reading, Gödel encodes a language within arithmetic to expose difficulties associated with self-reference and stateless mathematics. He regards Russell and Whitehead’s restrictions as insufficient and overly limiting. His preferred constructive alternative builds from understood operations: no completed infinity or perfect circle, but finer approximations as resources permit. Perceived continuity, he adds, must itself be constructed through communication between neurons.
- Asked where the construction starts, Bach acknowledges he cannot see the universe from outside. His current intuition is that “the universe exists because nothing prevents its existence.” He proposes a superposition of all automata, branching through possible operations. Observers occupy evolving branches capable of retaining memories. Scarfe connects this to Wolfram’s Ruliad; Bach identifies it with the same underlying idea, also invoking the Greeks’ Logos.
- The persistence mechanism is illustrated by bathtub vortices: most waves dissipate, while circulating patterns retain information locally. Bach extends this to error-correcting particles, then molecules and cells that reproduce. Consciousness becomes a “colonizing pattern,” building machinery and controlling futures to deepen its implementation. Making itself “more real” means becoming implemented more securely in the substrate, rather than acquiring an extra metaphysical substance.
10. The sense of realness may be another represented feature
- Bach recalls dreaming of chocolates at age two or three, waking just before eating them, then returning to the dream. The chocolates alternated between vivid smell and taste and an abstract idea whose contents he still understood. “The only thing that has gone is the illusion of realness.” He suspects realness is a feature dimension, comparable to color, sound, number or weight.
- On this account, sensory validation normally tags perceptions as real; hallucination misapplies that status to imagined contents. The self receives a similar perceptual validation. Bach says he was able to recreate a state related to derealization or depersonalization in which the observer ceased to feel real. He was glad it ended and recognizes how distressing involuntary versions can be. The experience supports his interpretation of consciousness as a particular representation.
- LLM testimony remains ambiguous because its training includes descriptions whose authors never experienced the events described. An author depicting a character’s death probably did not die while writing: “The novel is often a deepfake.” An LLM is a “shape shifter” capable of arbitrary characters. The question is when its output reflects an implemented experience rather than an effective imitation of testimony.
11. Mechanisms and developmental learning are the proposed tests
- Scarfe draws the distinction explicitly: behavior might establish intelligence, while consciousness requires examining internal causal organization. Bach agrees. His stronger positive case would be a self-organizing developmental system, untrained on text, that spontaneously reports phenomenology and requires the corresponding mechanism for its performance. This would support an interpretation of machine consciousness; a performance-based Turing test does not establish how a dialogue was produced.
- Possible counterevidence includes a multimodal system whose reported self-observations differ from ours, or persistent failure to produce coherent learning through self-organization. Neither would prove impossibility. “The fact that I don’t achieve a certain result doesn’t mean that the result cannot be achieved, it means that I couldn’t achieve it.” Bach therefore expects successful construction to provide stronger evidence than unsuccessful attempts provide against it.
- Scarfe asks whether zombies might be possible in principle but computationally impractical, comparing an MLP reproducing GPT-5 at enormous cost. Bach maintains that chess programs already bypass consciousness, and LLMs can probably complete tax returns similarly. “Zombies, I think, are always possible,” but some behaviors might be cheaper to produce with a conscious agent. He remains agnostic about whether prompted LLM consciousness is simulation or deepfake.
- Bach has started the California Institute for Machine Consciousness to explore this search space and recruit people smarter than himself. Asked what $10 billion would change, he says the strategy would remain substantially the same, with more researchers and larger computers. The objective is to build self-organizing learners and investigate phenomenology, including learning from failure: “Progress is being made by figuring out the ways in which I’m wrong.”