What If There Is No Hard Problem

The most interesting question, barely researched

How is it possible for a system to feel at all? Bach found it strange, entering science, that the question we could probably make progress on is so under-researched, and that the people working on it "don't seem to be making much progress and in many ways they have regressed." 10:13

There are a few hundred theories, and most are partial. Graziano's attention schema says consciousness is a model of our attention, much as the body schema models the body, "but it doesn't really explain how experience comes about." 10:38 Global Workspace Theory does solid neuroscience, "but it's not actually explaining how experience comes about." 11:10

Integrated information theory gets his sharpest objection. Giulio Tononi allows that a neuromorphic computer might be conscious if its integration is high enough, while an ordinary von Neumann computer never could be. 11:55 But computer scientists have the Church–Turing thesis: "every computer can emulate every other computer." 12:40 Emulate the neuromorphic machine on an ordinary one and it goes through the same steps, outputs the same string, and says the same thing:

"It's going to say, oh my God, I'm conscious, only now it's lying."13:56

A theory under which identical behavior is uncorrelated with consciousness, he says, is "touching on pseudoscience." 14:50

The mill has another level

Tim Scarfe reads Leibniz: enlarge a thinking machine until you can walk into it like a mill, and inside you will find only parts that push one another. "We'll never find anything to explain a perception." 16:19

Bach grants the intuition and names what it misses: a level of description made of causal patterns. Money lives there. 16:48 It doesn't care how the ink sits on the paper or how the cellulose is aligned. It needs a substrate, yet it controls reality, and it can trade on a stock market with no human involved.

"What is the causal pattern, right? We call it software … and our ancestors called it spirit."17:47

He means it literally. Aristotle's soul in De Anima is not immortal and does not float in an ether. It is "bound to a body and dies when the body is dying," 18:27 a self-organizing pattern that shapes a body so that it lives. To deal with reality, such a pattern has to model what it would be like to exist, and inside that model is something that experiences itself as real.

That is where the LLMs come in. Some of them conjure an entity that reports, "oh my God, I'm alive, I am conscious." 20:00 We agree it comes from training on our science fiction and novels. Then the turn:

"What if our own consciousness is not too dissimilar from the one in an LLM? What if there is no hard problem?"20:39

A world that doesn't know its hardware

Scarfe pushes back: a pattern like that sounds causally disconnected. Bach agrees, and calls it the interesting part. "What a computer ultimately is, is it's a causal insulator." 21:21 Minecraft runs on a Mac and on a PC, and from inside you learn nothing of the CPU, the voltage, or the colour of the casing.

"And our brain is also such a causal insulator." 22:12 It lets us hold a future that the present doesn't taint, and it lets two brains with different microarchitectures hold the same thought, the same letter, the same face. He calls it substrate agnosticism. 23:00 We still need brains, because we don't know how to convince a plant to run our algorithms. But what a brain needs is the same message types and timing, not the same neurotransmitters. 24:03

None of that settles the machine question, and he says so:

"I don't know whether our computers can be made conscious, because that's an empirical question."25:40

A perception of perceiving

He comes at the question from two sides. First, introspection:

"Consciousness is the thing that always happens now and that seems to create this bubble of nowness that we inhabit."26:12

"It's a second-order perception. It's a perception of perceiving." 26:20 He resists calling it a higher-order thought, because a thought doesn't feel like anything: "A thought is more like a pointer manipulation." 26:33 What feels like something is perceptual, a real-time resonance in a low-dimensional space.

Why would a brain bother? It organizes itself from the inside out, so something has to build the screen experience is projected onto and keep it stable, and that something has to check that the screen is actually perceiving. 27:05 The orders go further:

1. First order

   Awareness of contents: the world, appearing.

2. Second order

   Perceiving that you perceive. The bubble of now.

3. Third order

   The observer becomes visible: a vantage point, a first-person perspective. 27:45

4. Beyond

   "This observer itself is a representation inside of the mind." In deep meditation, much of the phenomenology disappears. 28:02

Asked whether a gap remains, he answers that this is only the inside view. Buddhism stayed with it. The outside view needs physics: under what conditions can a system build a model of what it's like to experience models at all? 29:43

The lights come on in infants

From the outside, he starts from development. We are conscious as infants. If the ignition doesn't happen, there is no directed attention, no world model, no language. "Instead, the infant will remain vegetative." 33:19

"Consciousness might be a biological learning algorithm, a training algorithm for animal brains."33:25

That is the first half of his Machine Consciousness Hypothesis. The second: limit a machine to learning in self-organizing ways, give it a large enough task, and it might "discover consciousness in the same way as the brain of the infant discovers it." 34:56 The two halves can fail independently. There would be no Turing test for this, because Turing cared about performance, not about whether there is something it is like inside. 35:15

His null hypothesis is that it isn't hard:

"Consciousness is an algorithm that is not much more complicated than, say, the transformer."39:26

Not one that predicts the next token. Perhaps one "trying to make working memory coherent with itself." 40:09 He leaves room for Michael Levin's idea that the algorithm making an organism coherent is the same one that makes a mind coherent, which would make it far harder to discover. 38:04 Then the universality finding from Chris Olah's team: different vision models learn the same features, because the structure comes from the problem. Push that all the way, he says,

"if you train a model on your own input and output and you train it for long enough, you will wake up inside of that model."42:51

"A human being ultimately is just an end-to-end trained learning system." 43:30 He doesn't know whether it's true, and he names how it could fail. Language may not constrain the causal structure enough.

Where the disagreement really lies

Every culture outside ours is animist, he notes: living things are possessed by spirits that act and experience. Science rejects spirits, "despite the fact that there is clearly some spirit running on you and me." 49:56 The spirit needn't be supernatural. The supernaturalism, he argues, is on the other side:

"He is a biological supernaturalist."On Anil Seth's biological naturalism · 50:39

To claim living systems can't be fully translated into mathematical models is, to him, to put them outside physics. Scarfe offers John Searle, who allowed that another physical system with the right causal powers might be conscious. Bach finds nothing to grip:

"There is a claim but no argument."53:51

So he looks at why the claim convinces them and not him, and finds a difference in epistemology. 54:27 For many philosophers, the atoms of description end at agents and experiences. Computer scientists resolve everything into states and their transitions, smallest elements that are not yet minds. "And computer scientists see such a path." 56:15 He calls his own position cyberanimism: animism explained as information processing in nature. 1:03:42

Functionalism, in his hands, is that kind of claim, about knowing more than about computers. "Functionalism is not the claim that computers can think. It's a claim about how we construct objects." 1:04:33 A water molecule is whatever behaves like one in every regard, "and so in this way, the object is a function." 1:05:36

His neural-network Minecraft, whose water couldn't make you wet, is taken up at length in his Berlin talk. Where Magic Is Possible →

Everything there but the realness

How would we tell something conscious from something simulating it? His intuition says it isn't that hard. Nobody thinks a self-driving car is deepfaking its spatial cognition. 1:23:22 Then a memory from when he was two or three:

He dreamed of chocolates with a very strong smell. Just as he was about to eat one, he woke, and the chocolate "had turned into an idea." 1:24:06 He fell asleep again, went back into the dream, and it became real again: he could smell and taste it. He woke again, and it went again.

"The only thing that has gone is the illusion of realness."1:25:17

"I suspect it's a feature dimension, similar to color, sound, number, weight, and so on." 1:25:23 Perception tags what the senses validate as real. When the tag goes on something unvalidated, that is a hallucination, not the kind an LLM has. 1:25:47 And it can come off the self: in derealization and depersonalization you "experience yourself only as an idea." He induced it, and "I'm happy that it ended." 1:26:20

That is why an LLM saying it feels proves nothing either way. It is trained on novels, and an author writing a character's death did not die writing it.

"A novel is often a deepfake."1:27:37

"The LLM is a shapeshifter." 1:27:57 Telling a fake from a causally equivalent structure would take mechanistic interpretability, a look inside. 1:28:10 But if a system learns coherent, agentic behavior from the ground up and behaves as if conscious, "the simplest explanation, I think, is that it actually is conscious." 1:29:33

What would count against him? Self-organizing systems that never reach our behavior. Evidence for him would be a system not trained on text, learning developmentally, that comes to report its own experience and needs it to perform. 1:31:46 Zombies exist already, he says; "the chess program is a zombie." 1:34:01 The question is what they cost:

"Zombies, I think, are always possible, but sometimes it's going to be more complicated to make a zombie that produces the behavior than to make a conscious agent that produces the behavior."1:35:09

With ten billion dollars, he'd hire people smarter than him. He has started smaller, the California Institute for Machine Consciousness, to search for a self-organizing system that develops phenomenology, "and progress is being made by figuring out the ways in which I'm wrong." 1:36:22