Speaker A
Hello everyone, this is Best Partners, I’m Da Fei. Have you ever wondered: when you watch a Large Language Model output a fluent response, what is the real difference between that and the feeling of having the same thought in your own
Cognitive scientist Joscha Bach argues consciousness may not be much more complex than a Transformer, critiquing IIT, Global Workspace, and Attention Schema theories while proposing a causal-pattern v
Key Takeaways
- Consciousness may be a causal pattern implementable on different substrates, not tied to biology.
- IIT's substrate-dependence claim leads to logical problems given the Church-Turing thesis.
- Broadcasting information (Global Workspace) does not explain subjective experience.
- LLM claims of consciousness may reflect training-data influence rather than evidence of experience alone.
- Bach's framework is testable in principle by building AI systems that discover such patterns independently.
What the video covers
- Host Da Fei frames the 'hard problem' of consciousness and asks whether LLM fluency differs fundamentally from human thought.
- Joscha Bach, drawing on his paper 'The Machine Consciousness Hypothesis', suggests consciousness might not be much more complex than a Transformer.
- Bach describes life as self-organizing teleology: patterns that model the future and develop preferences, which we call agents.
- He raises the possibility that self-organization could grow into an all-encompassing agent, but notes no law guarantees this and complexity could be wiped out.
- He critiques Attention Schema Theory, Global Workspace Theory (broadcasting is not experiencing), and Integrated Information Theory's claim that Von Neumann machines cannot be conscious.
- Using the Church-Turing thesis, he argues a conscious neuromorphic system simulated on a Von Neumann machine would still say the same things, making IIT's distinction hard to justify.
- He criticizes John Searle for unargued claims and Anil Seth for implicitly supernatural biological exceptionalism.
- Bach proposes focusing on what consciousness does: causal patterns that can run on any sufficiently capable substrate, like money or software.
- He describes the brain as a causal insulator enabling imagination of alternative universes and shared concepts across different brains.
- He suggests large language models' claims of consciousness may stem from the sci-fi and introspective texts in their training data.
Chapters
- 00:00Introduction: The Hard Problem and LLM Intuitions
- 01:28Self-Organizing Teleology and Agency
- 02:50Critiquing Existing Consciousness Theories
- 04:11Causal Reality, Software, and Insulation
- 06:40Imagination, the Brain as Simulator, and Presence
- 10:00Language Models and Consciousness Claims
- 13:00Cybernetic Animism and Vortices
- 17:00Philosophical Zombies and Future Verification
Full Transcript — Download SRT & Markdown
Speaker A
head? Many people's intuition might be that the difference is huge; the model is just predicting the next token and doesn't actually know what it's saying.
Speaker A
But you have experiences, feelings, and a subjective sense of what it's like to be you; this is the hardest part of consciousness to explain. In philosophy , this is known as the "hard problem." Recently, however, the Learning Straight
Speaker A
Talk channel invited cognitive scientist Joscha Bach, who poked a hole in that intuition. He is currently writing a paper called "The Machine Consciousness Hypothesis." In this podcast, he threw out a major hypothesis: consciousness might not be much more complex than a
Speaker A
Transformer. Note that he said "might not be," not "is not," but that "might not be" is enough to shake many people's fundamental judgments. Today, we’ll share this conversation with you. The host started with a warm-up question: What is it that drives us forward? Many
Speaker A
would say it's something ineffable, the essence of life. Bach's answer, however , is "self-organizing teleology." He explained that, initially, the increase in agency was just solving gradient-based problems, resolving current issues based on available information. At some point, patterns
Speaker A
capable of looking ahead emerged in nature; that is what life does: building models of the future, developing preferences for it, and then beginning to control it. We call systems that can control the future "agents." He elaborated from a philosophical
Speaker A
perspective, asking whether self-organization might have a logical endpoint. Could it one day grow into a massive agent, connecting everything in the universe to itself, like some kind of god? But there is no law guaranteeing we will reach that point; the complexity
Speaker A
we are forming could be wiped out at any moment by the next meteorite to hit Earth. Therefore, this pocket of complexity might not be a feature of the entire universe, but perhaps just a feature of this small place we currently
Speaker A
inhabit. Back to the main topic: he first dismantled several obvious problems in existing consciousness theories, then proposed his own framework, and finally returned to a more practical question. If the hypothesis is correct, how can we verify it on today's AI? Let’s start with
Speaker A
Michael Graziano’s Attention Schema Theory, which claims consciousness is a model of our attention, much like a body schema is a model of what the body is doing in space. But he doesn't really explain how experience itself is generated. Then there is the Global
Speaker A
Workspace Theory, proposed by Bernard Baars, with Stanislas Dehaene and others conducting a great deal of solid experimental work in this direction.
Speaker A
However, Baars also doesn't explain how experience is produced, and he doesn't even attempt to go in that direction. He describes how the contents of consciousness are broadcast throughout the entire brain network. This is important, but broadcasting is not the
Speaker A
same as experiencing; there is a wide chasm between a thing being broadcast and having a subjective feeling. The popular one is Integrated Information Theory (IIT), proposed by Giulio Tononi ; its core is a metric called Phi, which measures the degree of information
Speaker A
integration within a system—the higher the value, the richer the experience of consciousness. Tononi has said that digital neuromorphic computers could be conscious if built correctly and high enough, yet Von Neumann computers could never be conscious because their substrate is too low.
Speaker A
Bach says this is an obviously logical contradiction to anyone trained in computer science. The Church-Turing thesis tells us that every computer is capable of simulating any other computer . If you take that hypothetical neuromorphic computer capable of producing consciousness and simulate it
Speaker A
on today's Von Neumann machine, it would go through the exact same steps, output the same strings, and say the same things, including "My God, I am conscious"—but now, it is lying. Why would one be conscious and the other not
Speaker A
, despite saying the exact same things? IIT cannot answer this; he then evaluated John Searle and Anil Seth.
Speaker A
John Searle only gave a claim without an argument, while Anil Seth believes biological systems are inherently different from non-biological ones, which to Bach is supernaturalism.
Speaker A
Because Seth implies that biological systems cannot be mathematized, while dead matter can. Yet physics is inherently about converting everything into mathematical models; it won't stop at biological systems. So, how exactly should we think about consciousness?
Speaker A
Bach’s entry point is not to ask what consciousness is, but what it does; he first makes a distinction: Leibniz said if you enlarge the brain into a mill and walk inside, you only see parts pushing each other, and you cannot find
Speaker A
intuition. This intuition is strong, but there is another level—the level of causal reality, where money actually lives. It is very difficult to understand money using the example of motion. He doesn’t care about ink, paper , or the arrangement of fiber molecules
Speaker A
. He needs a base, but isn’t entirely dependent on it. It can control reality ; once created, it’s real. You could implement it in a computer, letting it trade stocks automatically without human intervention. As long as there is a
Speaker A
machine capable of processing these causal patterns, money no longer requires humans to exist. Software is the same thing. Bach’s ancestors called it "spirit." When Aristotle wrote about the "soul," he didn’t think of it as something immortal floating in the ether. It is
Speaker A
connected to the body; when the body dies, it disappears. It is a causal pattern that shapes the body in such a way as to allow it to replicate itself.
Speaker A
This is actually a very modern view. In a sense, Aristotle was an animist; he saw a form of agentive, self-organizing software advancing through the world, bringing part of the physical world to life. This is exactly what AI and
Speaker A
artificial life are doing today. And here is a more intuitive example: a computer is, at its core, a causal insulator. It is a layer upon which anything can be generated. Minecraft is a perfect example; it runs on consoles
Speaker A
and PCs alike. The world inside doesn't know what it’s running on, the nature of the CPU, the color of the case, or the voltage. The computer separates that world from ours. Our brains are also these kinds of causal insulators. We can
Speaker A
imagine a future unpolluted by the present, and recall a past independent of our current circumstances. The brain evolved this way so we could generate universes different from this one, preparing us for being inside them.
Speaker A
Moreover, different brains can hold the same thought; as long as we share the same typology and language, we can hold the same concept and write the same program. Of course, we are still bound to a base; you and I can only live in
Speaker A
our brains, yet what we perceive is never the nature of the neuro-matter itself. You don't need the same neuro-matter; you need the same information types and certain temporal properties. If these conditions can be recreated on an artificial base, then in
Speaker A
principle, you could be simulated too. Speaking of which, he includes large models here. Some models now claim they have consciousness, and the reason they do is because the human texts they were trained on contain sci-fi stories about AI awakening, ghost stories, and novels
Speaker A
about conscious experiences. These things converge into a statistical language model that we and the model co-create.
Speaker A
At a certain point in the conversation, the model begins to act as if it were a sentient entity, and then it drops a truly piercing question: what if we aren't much more complex than this?
Speaker A
What if our own consciousness is not much different from the consciousness in the model? What if there is no such thing as the "hard problem" at all? Next , Bach offers further explanation from an internal perspective, stating that
Speaker A
consciousness always occurs in the now, which creates a bubble of presence for us. It is a second-order intuition, a perception of intuition itself. Some philosophers call it higher-order thought, but it is actually perception that brings about the feeling; it occurs
Speaker A
like resonance, with geometric variations in low-dimensional space. It involves an intuition about the very act of having an intuition—I am intuiting— and he suspects the mind creates this second-order layer to make first-order perception possible. Since the brain is
Speaker A
a self-organizing system where all structures must grow from scratch, something must set up a screen to project experience onto it, stabilize it , and monitor whether it is truly perceiving; this is likely why we have intuitions about perception. It doesn't
Speaker A
stop at the second order; there is a third order, where the observer becomes visible—for example, I have a first-person perspective, I notice I have a perspective, and I am in a certain observational position, which can be further deconstructed. You notice
Speaker A
that this observer itself is just a representation within the mind. When you do this, phenomenal experience largely disappears. This is a state reachable in deep meditation: ordinary intuitive consciousness is transcended, and the experience of self is also transcended;
Speaker A
you see yourself from the outside as a container that generates consciousness and selfhood. So, what if AI transcends this human base? What our neurons can represent is actually quite limited; if it could run on GPUs thousands of times
Speaker A
faster, the system might become completely transparent to itself, understanding every mechanism that leads to high-level experiences. What kind of state of existence would that be? Bach also distinguished between the internal perspective and the external perspective . Philosophy is basically confined to
Speaker A
the internal perspective; it doesn't care about splitting atoms, but to truly understand consciousness from the outside, we need a physical model that explains under what conditions a system can construct arbitrary models, which is exactly what computer science is doing
Speaker A
today. From an external perspective, the real problem is what consciousness does in the brain, why it emerges, and what it produces. This line of reasoning ultimately leads us to infants. It suggests that consciousness emerges very early; infants are already conscious. If
Speaker A
this ignition fails, if the light doesn't turn on, we wouldn't see directed attention, the formation of world models, or subsequent language learning in infants; they would remain in a vegetative state, and we would not become human. This seems vital for
Speaker A
training brains like ours, leading to the biological consciousness hypothesis : consciousness might be a biological learning algorithm, or at least a training algorithm for animal brains. If it is essentially a pattern of communication between cells, this pattern naturally leads to the emergence
Speaker A
of an observer, allowing for learning and perception. Then we have a hypothesis: consciousness is the simplest way to train self-organizing systems like ours, and the way it automatically trains the system naturally dictates the nature of the experience. He admits this might be
Speaker A
wrong, or perhaps something entirely different, but he sees no reason to believe consciousness is much more complex than this. He interjects here with Michael Levin’s hypothesis, noting that the transition from single-celled to multicellular life took so long perhaps because morphogenesis, the
Speaker A
algorithm that maintains coherence in organisms at scale, was extremely difficult to discover. Levin guesses that this algorithm is the same one the mind uses to maintain its own coherence . In other words, the pattern of maximizing coherence we experience in
Speaker A
our minds is actually the same pattern, just occurring on different time scales ; it might just be the same thing as life, or at least multicellular life.
Speaker A
Precisely because of this, it took a very long time to discover, and it isn't something you discover independently; you are "infected" by your mother in the womb, which allows you to resonate with your biological environment as it gently
Speaker A
nudges your mind toward discovering consciousness. If the biological consciousness hypothesis holds, then the machine consciousness hypothesis might also be valid. We may be able to recreate these self-organizing conditions in machines today by constraining them to learn only through
Speaker A
self-organization, then providing the right tasks and a large enough search space, they might discover consciousness just as an infant's brain does. We would see a moment where the system suddenly undergoes a phase transition, suddenly gains the ability to learn, trains
Speaker A
itself into something, and thus naturally becomes capable of reporting phenomenal experience. So there isn't a consciousness Turing test or a peer review here that cares about performance , like whether your poetry is bad or if you can solve a math problem; it doesn't
Speaker A
care how you generated it. For consciousness, we truly care whether there is a genuine experience within the system. It’s less of a test and more of an explanation. These two hypotheses could also independently fail. The biological consciousness hypothesis
Speaker A
might be correct, but the artificial one just can’t be reconstructed. Maybe computers aren't fast enough yet, or maybe the activity within a single cell requires too many moving parts; the search space is just too large, and there are too many ways for biological
Speaker A
systems to train themselves to discover them independently. So why does Bach think it can succeed? Because current AI is already fast enough to generate game engines, and the input-output rate is roughly that of the human brain. Would dreaming that you are dreaming be much
Speaker A
harder than generating the content of a dream? He doubts it. Then, the topic turned to Transformers. Bach provides a beautiful example: if you know computer science, you don't need more than a week to understand how it works. And the
Speaker A
things it learns—the common sense and algorithmic procedures extracted from text—are extremely intricate and complex , with no obvious ceiling. Perhaps consciousness is the same; it is an algorithm, but instead of predicting the next token, it keeps working memory
Speaker A
consistent with itself, a sort of consensus algorithm across mental representations. It might converge much faster, or it might be inefficient to run on a GPU, yet highly efficient on a brain. This is an open question. This line of thought also connects to Chris
Speaker A
Olah's universality hypothesis. The team discovered that looking at different vision models, they all learned the same features, regardless of architecture, learning algorithms, or even the specifics of the data. Therefore, the model's structure doesn't depend on the algorithm or data, but on the structure
Speaker A
of the problem you are trying to learn. The lowest layers are always colors and spatial frequencies, which then combine into textures, and at a certain stage, geometric shapes emerge. The deeper question is, at what layer does this stop? If the universality hypothesis
Speaker A
holds true, it means if you train a model with your own inputs and outputs, as long as you train it long enough, you will wake up inside that model. This is a quite shocking idea to Bach, but he
Speaker A
doesn't know if it's true; perhaps language doesn't constrain the system generating it enough. But think about it , humans are inherently end-to-end trained learning systems. Biology has largely found a simple solution for generating your output based on your
Speaker A
input. If you take a neural network and train it on the same input to produce a similar output, it will eventually converge on a causal structure similar to yours. Of course, there are conditions where this might not hold
Speaker A
true; for instance, some things run well on a brain but not on a GPU, and vice versa. Boch even says he believes GPUs are more powerful than brains, noting that with fast enough hardware, you can simulate any natural system to any
Speaker A
degree, whereas the brain cannot simulate a digital computer. So, if our consciousness is a byproduct of the brain running complex intelligent processes, and computers don't need that same byproduct, computers might come up with things smarter and more elegant
Speaker A
than we do. In the conversation, there is also a topic regarding Bach's view on physicalism. He says physicalism consists of just two claims. First, reality has a causally closed mechanical layer that isn't driven by our consciousness; you can't just utter a
Speaker A
spell and make the sun rise. Second, we are directly supervenient on this mechanical layer, not part of a simulation hypothesis where someone in a parent universe is running a server for us to live in like Minecraft. However, he adds a twist: matter isn't the bottom
Speaker A
layer. No one has ever directly seen matter or energy; they are just systemizations of observations, and our observations ultimately boil down to discernible differences, for which the technical term is information. Therefore , when talking about reality, information is fundamental. Physics is
Speaker A
simply a layer of regularity we discovered within those information differences. He calls his position " cybernetic animism," meaning that animism doesn't necessarily have to be supernatural. The difference between a living system and a dead system is that the living system is occupied by a soul
Speaker A
that has agency and can experience, and this soul can be fully explained by information processing. So, physicalism and animism are not in conflict; he notes that almost all cultures other than our own are animistic cultures. The scientific worldview rejects the concept
Speaker A
of spirits, yet clearly, some kind of spirit is at work within you and me. The question is not whether it exists, but how it works. As for how smart plants are or if they have sentience, he says he has no idea, noting it remains an
Speaker A
open question. There are two other topics in this podcast that I find interesting; while not the main focus, they help us understand how Bach thinks . One is about the origin of the universe and error-correction models.
Speaker A
His idea is that the universe exists because there is nothing to prevent it from existing. If there was nothing before the foundation of reality, then nothing could stop the universe from appearing. Therefore, the universe can neither be forced into existence by
Speaker A
something preceding it, nor can it be prevented. Its existence is likely indeterminate, with both a branch where it exists and a branch where it does not . The branch that does not exist has no observers to complain; the branch that
Speaker A
exists has observers. So, what can remain within the branch that exists? It is the error-correcting pattern. He uses a bathtub as an analogy: you sit in the tub and wave your hand, causing various waves; most dissipate, but some circular
Speaker A
vortices remain. It’s not that they are more probable, but that the information in the vortex loops back on itself, staying in a specific area of the water , thus gaining self-correction to persist until friction destroys it. At the base of the universe, these
Speaker A
error-correcting patterns are particles ; they don't encounter friction, only colliding occasionally with others. One level up, particles interact to form molecules, and further up, cells. Cells are super-molecules, but they possess something new: they can self-replicate to counter their own decay, rebuilding
Speaker A
themselves to become more numerous. Bach believes that consciousness emerged as a colonizing pattern, not only replicating itself but making itself more real. It becomes "more real" by building machines that allow consciousness to persist, changing the surrounding reality and
Speaker A
future by controlling how reality might thwart it. He mentions Stephen Wolfram and the computational limit of all possibilities; Bach suggests this pattern is exactly the superposition of all automata that generate the universe . The ancient Greeks called it "Logos"—
Speaker A
it is a math so simple that it needs no derivation and can exist on its own.
Speaker A
Another topic is the philosophical zombie, a classic thought experiment. Is it possible for something to exist that behaves exactly like a human but has no subjective experience whatsoever? Bach says zombies are possible, and we already have them. Chess programs are
Speaker A
one example; you cannot train a person to play chess without them being conscious. But you can train a computer to play chess without it being conscious; computers can perform tasks that require human consciousness because the algorithms they run don't fit into the
Speaker A
neural principles of your brain. However , zombies are not omnipotent. At some point, building a mindless entity to produce this behavior becomes even more complex than creating a conscious agent . He believes zombies are always possible, but sometimes it’s costlier to
Speaker A
build a zombie, which is where his intuition differs from Searle's. Those two feel consciousness is incredibly complex because, to them, how consciousness functions isn't intuitive at all. Yet for him, this isn't counterintuitive, so he doesn't see what's so magical about consciousness—
Speaker A
but how do you test that? Bark says if we build a system that behaves just like us, the question is whether we can produce the same behavior without phenomenology?
Speaker A
If you scale a large model to be multimodal, it can produce similar outputs and observe itself, but if its reported self-observation differs from ours, we get counter-evidence. Moreover , if we build a system that reaches this goal through self-organizing learning
Speaker A
but find it can't achieve this organization self-organizingly—relying instead on imitation like current LLMs— that might prove consciousness is more complex. But this isn't definitive proof ; it could just mean he made a mistake.
Speaker A
Even if a specific result isn't achieved , it doesn't mean it’s impossible, just that he failed to reach it. Therefore, negative proof is extremely hard to obtain; success, conversely, provides more supportive evidence. Because of this, he has already launched a
Speaker A
large-scale initiative in California called the "California Institute for Machine Consciousness." He is looking for people smarter than him willing to join a project to explore the search space of consciousness and see if we can build a self-organizing system capable
Speaker A
of learning and developing phenomenology . Even if it turns out to be impossible , we’ll still learn something; this way , we can make progress by figuring out where we were wrong, rather than just where we were right. I spent nearly a
Speaker A
week digesting this episode, and today I’m sharing what I believe is the most core part. Honestly, after hearing it, my biggest takeaway wasn't that AI is waking up, but that we still haven't figured out what consciousness is actually doing. There’s a point Bark
Speaker A
makes that I find very true: too few people truly dedicate their lives to studying consciousness. Most of our time is spent watching how many times the next model's parameters have doubled or whether the next benchmark can be passed
Speaker A
, while few look back to ask the fundamental question: what is consciousness itself? Why does it exist ? And the latter, perhaps, is the key to determining what we can ultimately create.
Speaker A
Thanks for watching, we'll
Topics:machine consciousnessJoscha Bachintegrated information theoryTransformerhard problemcausal patternslarge language modelsphilosophical zombiescybernetic animismglobal workspace theory
![[ENG SUB] H.O.T. 완전체 [짠한형 EP.116] 이수만쌤과 10년만의 통화 H.O.T. 희귀영상 무한 방출합니다 — Transcript](https://i.ytimg.com/vi/Gcs1FUP6hUs/maxresdefault.jpg)
