视频库 / REC_008
--:--
点击播放，这里会跟随视频显示当前句的中英字幕。
第 8 期 · 回应 Ⅰ·07「意识可以被复制吗？」
The Meta-Problem of Consciousness | Professor David Chalmers | Talks at Google
节目发布 2019-04-02 · Talks at Google章节 · 点击跳转视频
0:00
开场：从理科到意识哲学
2:02
难问题与易问题的区分
9:44
AI研究者的切入口：元问题
14:41
谱系学分析与幻觉论
20:22
问题直觉的分类与实证研究
26:38
候选解释：自我模型与内省不透明
34:06
计算模型与人工意识测试
37:15
用元问题检验IIT等理论
40:08
幻觉论与实在论的双重荒谬
45:34
问答：机器人、Φ值与信念
51:22
问答：哥德尔、自指与错觉
61:51
问答：意识的价值与延展心灵
本期讲者
大卫·查尔默斯澳大利亚哲学家、纽约大学教授，1995年提出「意识的难问题」，著有《有意识的心灵》《现实+》，与安迪·克拉克合著《延展心灵》。
01开场：从理科到意识哲学
0:00
[MUSIC PLAYING]
[音乐播放]
0:06
DAVID CHALMERS: Thanks for coming out. It's good to be here. As Eric said, I am a philosopher thinking about consciousness. Coming from a background in the sciences and math, it always struck me that the most interesting and hardest unsolved problem in the sciences was the problem of consciousness. And way back 25 years ago when I was in grad school, it seemed to be the best way to come at this from a big picture perspective was to go into philosophy and think about the foundational issues that arise in thinking about consciousness from any number of different angles, including the angles of neuroscience and psychology and AI.
大卫·查尔默斯：感谢大家前来。很高兴来到这里。正如埃里克所说，我是一名思考意识问题的哲学家。我出身于理科和数学背景，一直觉得科学中最有趣、最难的未解难题，就是意识问题。早在25年前我读研究生的时候，我觉得从宏观视角切入这个问题的最佳方式，是进入哲学领域，去思考从各种不同角度思考意识时所产生的那些基础性问题，包括神经科学、心理学和人工智能的角度。
0:52
In this talk, I'm going to present a slightly different perspective on the problem after laying out some background, the perspective of what I call the meta-problem of consciousness. I always liked the idea that you approach a problem by stepping one level up, taking the metaperspective. I love this quote, "Anything you can do, I can do meta." I have no idea what the origins was. I like the fact this is attributed to Rudolf Carnap, one of my favorite philosophers. But anyone who knows Carnap's work, it's completely implausible he would ever say anything so frivolous.
在这场演讲中，我会先铺垫一些背景，然后从一个略有不同的视角来谈这个问题，也就是我称之为“意识的元问题”的视角。我一直喜欢这样一种想法：你通过往上跳一层、采取元视角来处理一个问题。我很喜欢这句话：“你能做的任何事，我都能做得更‘元’。”我完全不知道它的出处是什么。我喜欢它被归到鲁道夫·卡尔纳普名下这件事，他是我最喜欢的哲学家之一。但任何了解卡尔纳普著作的人都知道，他绝不可能说出这么轻佻的话。
1:26
It's also being attributed to my thesis advisor, Doug Hofstadter, author of "Godel, Escher, Bach" and a big fan of the metaperspective. But he assures me he never said it either. But the metaperspective on anything is stepping up a level. The meta-problem, as I think about it, is it's called the meta-problem because it's a problem about a problem. A metatheory is a theory about a theory. Meta-problem is a problem about a problem. In particular, it's the problem of explaining why we think there is a problem about consciousness.
这句话也被归到我的博士导师道格·侯世达名下，他是《哥德尔、埃舍尔、巴赫》的作者，也是元视角的忠实拥趸。但他向我保证，他也从没说过这话。不过，对任何事情采取元视角，就是往上跳一个层次。在我看来，元问题之所以叫元问题，是因为它是一个关于问题的问题。元理论是关于理论的理论。元问题就是关于问题的问题。具体来说，它是要解释：我们为什么会认为意识存在一个问题。
02难问题与易问题的区分
2:02
So there's a first-order problem, the problem of consciousness. Today, I'm going to focus on a problem about it. But I'll start by introducing the first-order problem itself. The first-order problem is what we call the hard problem of consciousness. It's the problem of explaining why and how physical processes should give rise to conscious experience. You've got all of these neurons firing in your brain, bringing about all kinds of sophisticated behavior. We can get it to be [INAUDIBLE] explaining our various responses, but there's this big question about how it feels from the first person point of view.
所以存在一个一阶问题，也就是意识问题。今天，我要聚焦的是一个关于它的问题。但我会先介绍这个一阶问题本身。这个一阶问题就是我们所说的“意识的难问题”。它要解释的是：物理过程为什么以及如何产生有意识的体验。你的大脑中有这么多神经元在放电，带来各种各样复杂精巧的行为。我们可以做到用它来解释我们的各种反应，但还有一个大问题：从第一人称角度来说，这一切感觉起来是怎样的
2:42
That's the subjective experience. I like this illustration of the hard problem of consciousness. It seems to show someone's hair catching fire, but I guess it's a metaphorical illustration of the subjective perspective. So the hard problem is concerned with what philosophers call phenomenal consciousness. The word consciousness is ambiguous 1,000 ways. But phenomenal consciousness is what it's like to be a subject from the first person point of view. So a system is phenomenally conscious if there's something it's like to be it.
3:14
A mental state is phenomenally conscious if there's something it's like to be in it. So the thought is there are some systems-- so there's something it's like to be that system. There's something it's like to be me. I presume there's something it's like to be you. But presumably, there's nothing it's like to be this lectern. As far as we know, the lectern does not have a first person perspective. This phrase was made famous by my colleague, Tom Nagel at NYU, who back in 1974 wrote an article called "What Is It Like To Be A Bat?".
3:49
And the general idea was, well, it's very hard to know what it's like to be a bat from the third person point of view, just looking at it as a human who has different kinds of experience. But presumably, very plausibly, there is something it's like to be a bat. The bat is conscious. It's having subjective experiences, just of a kind very different from ours. In human subjective experience, consciousness divides into any number of different kinds or aspects, like different tracks of the inner movie of consciousness.
4:23
We have visual experiences like the experience of, say, these colors, blue and red and green from the first person point of view and of depth. There are sensory experiences like the experience of my voice, experiences of taste and smell. They're experiences of your body. Feeling pain or orgasms or hunger or a tickle or something, they all have some distinctive first person quality. Mental images like recalled visual images, emotional experiences like a experience of happiness or anger. And indeed, we all seem to have this stream of a current thought or at the very least, we're kind of chattering away to ourselves and reflecting and deciding.
5:07
All of these are aspects of subjective experience, things we experience from the first person point of view. And I think these subjective experiences are, at least on the face of it, data for the science of consciousness to explain. These are just facts about us that we're having these subjective experiences. If we ignore them, we're ignoring the data. So if you catalog the data that, say, the science of consciousness needs to explain, there are certainly facts about our behavior and how we respond in situations.
5:39
There are facts about how our brain is working. There are also facts about how subjective experiences, and on the face of it, they're data. And it's these data that pose what I call the hard problem of consciousness. But this gets contrasted with the easy problems, the so-called easy problems of consciousness, which are the problems of explaining behavioral and cognitive functions. Objective things you can measure from the third person point of view typically tied to behavior. Perceptual discrimination of a stimulus, I can discriminate two different things in my environment.
6:18
I can say, that's red, and that's green. I can integrate the information about the color and the shape. I can use it to control my behavior. Walk towards the red one rather than the green one. I can report it, say that's red, and so on. Those are all data too for science to explain. But we've got a bead on how to explain though they don't seem to pose as big a problem. Why? We explain those easy problems by finding a mechanism, typically a neural or computational mechanism that performs the relevant function to explain how it is that I get to say there's a red thing over there or walk towards it.
7:00
Well, you find the mechanisms involving perceptual processes and action processes in my brain that leads to that behavior. Find the right mechanism that performs the function you've explained what needs to be explained with the easy problems of consciousness. But for the hard problem, for subjective experience, it's just not clear that this standard method works. It looks like explaining all that behavior still leaves open a further question. Why does all that give you subjective experience? Explain the reacting, the responding, the controlling, the reporting, and so on.
7:43
It still leaves open the question, why is all that accompanied by subjective experience. Why doesn't it go on in the dark without consciousness, so to speak? There seems to be what the philosopher Joe Levine has called a gap here, an explanatory gap, between physical processes and subjective experience. At least our standard kinds of explanation, which work really well for the easy problems of behavior and so on, don't obviously give you a connection to the subjective aspects of experience. And there's been a vast amount of discussion of these things over-- I mean, well, for centuries, really.
8:20
But it's been particularly active in recent decades, philosophers, scientists, all kinds of different views. Philosophically, you can divide approaches to the hard problem into at least two classes. One is an approach on which consciousness is taken to be somehow irreducible and primitive. We can't explain it in more basic physical terms, so we take it as a kind of primitive. And that might lead to dualist theories of consciousness where consciousness is somehow separate from and interacts with the brain.
8:49
Recently very popular has been the class of panpsychist theories of consciousness. I know Galen Strawson was here a while back talking. He very much favors panpsychist theories where consciousness is something basic in the universe underlying matter. And indeed, there are idealist theories where consciousness underlies the whole universe. So these are all extremely speculative but interesting views that I've explored myself. There are also a reductionist theories of consciousness from functionalist approaches, where consciousness is just basically taken to be a giant algorithm or computation, biological approaches to consciousness-- my colleague Ned Block was here, I know, talking about neurobiology-based approaches, where it's not the algorithm that matters, but the biology it's implemented in-- and indeed, the kind of quantum approaches that people like Roger Penrose and Stuart Hameroff have made famous.
03AI研究者的切入口：元问题
9:44
I think there's interesting things to say about all of these approaches. I think that right now, at least, most of the reductionist approaches leave a gap. But the non-reductionist approaches have other problems in seeing how it all works. Today, I'm going to take a different kind of approach, this approach through the meta-problem. One way to motivate this is to-- I often get asked, well, you're a philosopher. It's fine. You get to think about these things like the hard problem of consciousness.
10:15
How can I, as a scientist or an engineer or an AI researcher-- how can I do something to contribute, to help get this at this hard problem of consciousness? Is this just a problem for philosophy? For me to work on it as a AI researcher, I need something I can operationalize, something I can work with and try to program. And as it stands, it's just not clear how to do that with the hard problem. If you're a neuroscientist, there are some things you can do. You can work with humans and look at their brains and look for the neural correlates of consciousness, the bits of the brain that go along with being conscious.
10:55
Because at least with humans, we can take as a plausible background assumption that the system is conscious. For AI, we can't even do that. We don't know which AI systems we're working with that are conscious. We need some operational criteria. In AI, we mostly work on modeling things like behavior and objective functioning. For consciousness, those are the easy problems. So how does someone coming from this perspective make a connection to the hard problem of consciousness? Well, one approach is to work on certain problems among the easy problems of behavior that shed particular light on the hard problem.
11:31
And that's going to be the approach that I look at today. So the key idea here is there are certain behavioral functions that seem to have a particularly close relation to the hard problem of consciousness. In particular, we say things about consciousness. We make what philosophers call phenomenal reports, verbal reports of conscious experiences. So I'll say things like, I'm conscious, I'm feeling pain right now, and so on. Maybe the consciousness and the pain are subjective experiences. But the reports, the utterances, I am conscious, well that's a bit of behavior.
12:14
In principle, explaining those is among the easy problems. It's objectively measurable response. We can find a mechanism in the brain that produces it. And among our phenomenal reports, there's the special class we can call the problem reports, reports expressing our sense that consciousness poses a problem. Now admittedly, not everyone makes these reports. But they seem to be fairly widespread, especially among philosophers and scientists thinking about these things. But furthermore, it's a sense that it's fairly easy to find a very wide class of people who think about consciousness.
12:53
People say things like, there is a problem of consciousness, a hard problem. On the face of it, explaining behavior doesn't explain consciousness. Consciousness seems non-physical. How would you ever explain the subjective experience of red and so on? It's an objective fact about us-- at least about some of us-- that we make those reports. And that's a fact about human behavior. So the meta-problem of consciousness then, at a second approximation, is roughly the problem of explaining these problem reports, explaining, you might say, the conviction that we're conscious and that consciousness is puzzling.
13:34
And what's nice about this is that although the hard problem is this airy fairy problem about subjective experience that's hard to pin down, this is a puzzle ultimately about behavior. So this is an easy problem, one that ought to be open to those standard methods of explanation in the cognitive and brain sciences. So there's a research program. There's a research program here. So I like to think of the meta-problem as something we could play that role. I talked about earlier, if you're an AI researcher thinking about this, the meta-problem is an easy problem, a problem about behavior, that's closely tied to the hard problem.
14:09
So it's something we might be able to make some progress on using standard methods of thinking about algorithms and computations or thinking about brain processes and behavior while still shedding some light, at least indirectly, on the hard problem. It's more tractable than the hard problem. But solving it ought to shed light on the hard problem. And today, I'm just going to kind of lay out the research program and talk about some ways in which it might potentially shed some light. This is interesting to a philosopher because it looks like an instance of what people sometimes call genealogical analysis.
04谱系学分析与幻觉论
14:41
It goes back to Friedrich Nietzsche on the genealogy of morals. Instead of thinking about what's good or bad, let's look at where our sense of good or bad came from, the genealogy of it all in evolution or in culture or in religion. And people think a genealogical approach to God, instead of thinking about does God exist or not, let's look at where our belief in God came from. Maybe there's some evolutionary reason for why people believe in God. This often leads, not always, but often leads to a kind of debunking of our beliefs about those domains.
15:14
Explain why we believe in God in evolutionary terms, no need for the God hypothesis anymore. Explain how moral beliefs and evolutionary terms, maybe no need to take morality quite so seriously. So some people, at least, are inclined to take an approach like this with consciousness too. If you think about the meta-problem explaining our beliefs about consciousness, that might ultimately debunk our beliefs about consciousness. This leads to a philosophical view, which has recently attracted a lot of interest, a philosophical view called illusionism, which is the view that consciousness itself is an illusion.
15:55
Or maybe that the problem of consciousness is an illusion. Explain the illusion, and we dissolve the problem. I take that in terms of the meta-problem, that view roughly comes to solve the meta-problem. It will dissolve the hard problem. Explain why it is that we say all these things about consciousness, why we say, I am conscious, why we say, consciousness is puzzling. If you can explain all that in algorithmic terms, then you'll remove the underlying problem because you'll have explained why we're puzzled in the first place.
16:29
Actually, walking over here today, I noticed that just a couple of blocks away, we have the Museum of Illusions, so I'm going to check that out later on. But if illusionism is right, added to all those perceptual illusions is going to be the problem of consciousness itself. It's roughly an illusion thrown up by having a weird kind of self model with a certain kind of algorithm that attributes to ourselves special properties that we don't have. So one line on the meta-problem is the illusionist line.
16:59
Solve the meta-problem, you'll get to treat consciousness as an illusion. That's actually a view that has many antecedents in the history of philosophy, one way or another. Even Immanuel Kant and his great critique of pure reason had a section where he talked about the self or the soul as a transcendental illusion. We seem to have this indivisible soul. But that's the kind of illusion thrown out by our cognitive processes. The Australian philosophers, Ullin Place and David Armstrong, had versions of this that I might touch on a bit later.
17:32
Daniel Dennett, a leading reductionist thinker about consciousness has been pushing for the last couple of decades the idea that consciousness involves a certain kind of user illusion. And most recently, the British philosopher, Keith Frankish, has been really pushing illusionism as a theory of consciousness. He has a book centering around a paper by Keith Frankish on illusionism as a theory of consciousness that I recommend to you. So one way to go with the meta-problem is the direction of illusionism.
18:05
But one nice thing about-- many people find illusionism completely unbelievable. They find, how could it be that consciousness is an illusion? Look, we just have these subjective experiences. It's a data about our nature. And I confess, I've got some sympathy with that reaction. So I'm not an illusionist myself. I'm a realist about consciousness in the philosopher's sense, where a realist about something is someone who believes that thing is real. I think consciousness is real. I think it's not an illusion.
18:33
I think that solving the meta-problem does not dissolve the hard problem. But the nice thing about the meta-problem is you can proceed on it-- to some extent, at least in initial neutrality-- on that question, is consciousness real or is it an illusion. It's a basic problem about our objective functioning in these reports. What explains those? There's a neutral research program here that both realists, illusionists, people of all kinds of different views of consciousness can explain. And then we can come back and look at the philosophical consequences.
19:06
So I'm not an illusionist. I think consciousness is real. I've got to say, I do feel the temptation of illusionism. I find it really intriguing and in some ways attractive view. It's just fundamentally unbelievable. Nevertheless, I think that the meta-problem should be a tractable problem. Solving it, at the very least, will shed much light on the hard problem of consciousness even if it doesn't solve it. If you can explain our conviction that we're conscious, somehow the source, the roots of our conviction that we are conscious, must have something to do with consciousness especially if consciousness is real.
19:42
So I think it's very much a good research program for people to explain. So then I'll move on now to just outlining the research program a little bit more and then talk a bit about potential solutions and on impact on theories of consciousness before wrapping up with a little bit more about illusionism. So this meta-problem, which I've been pushing recently, opens up a tractable empirical research program for everyone, reductionists, non-reductionists, illusionists, non-illusionists. We can try to solve it and then think about the philosophical consequences.
05问题直觉的分类与实证研究
20:22
Now what is the meta-problem? Well, the way I'm going to put it is it's the problem of topic-neutrally explaining problem intuitions or else explaining why that can't be done. And I'll unpack all the pieces of that right now. First, starting with problem intuitions. What are problem intuitions? Well, there are the things we say. There are things we think I say. Consciousness seems irreducible. I might think consciousness is irreducible. People might be disposed, have a tendency to say or think those things.
20:55
Problem intuitions all take to be roughly, that tendency. We have dispositions to say and think certain things about consciousness. What are the core problem intuitions? Well, I think they break down into a number of different kinds. There is the intuition that consciousness is non-physical. We might think of that as a metaphysical intuition about the nature of consciousness. There are intuitions about explanation. Consciousness is hard to explain, explaining behavior doesn't explain consciousness.
21:25
There are intuitions about knowledge of consciousness. Some of you may know the famous thought experiment of Mary in the black and white room who knows all about the objective nature of color vision and so on, but still doesn't know what it's like to see red. She sees red for the first time. She learns something new. That's an intuition about knowledge of consciousness. There are what philosophers call modal intuitions about what's possible or imaginable. One famous case is the case of a zombie, a creature who is physically identical to you and me but not conscious.
21:56
Or maybe an AI system, which is functionally identical to you and me, but not conscious. That at least seems conceivable to many people. So this is the philosophical zombie. Unlike the zombies and movies, which have weird behaviors and go after brains and so on, the philosophical zombie is a creature that seems, at least behaviorally, may be physically like a normal human, but doesn't have any conscious experiences. All the physical states, none of the mental states. And it seems to many people that's at least conceivable.
22:25
We're not zombies. I don't think anyone here is a zombie-- I hope. But nonetheless, it seems that we can make sense of the idea. And one way to pose the hard problem is, why are we not zombies. So this imagined ability of zombies is one of the intuitions that gets the problem going. And then you can go on and catalog more and more intuitions about the distribution of conscious, maybe the intuition that robots won't be conscious. That's an optional one, I think. Or consciousness matters morally in certain ways, and the list goes on.
22:57
So I think there is an interdisciplinary research program here of working on those intuitions about consciousness and trying to explain them. Experimental psychology and experimental philosophy-- a newly active area-- can study people's intuitions about consciousness. We can work on models of these things, computational models or neurobiological models, of these intuitions and reports. And indeed, I think there's a lot of room for philosophical analysis. And there's just starting to be a program of people doing these things in all these fields.
23:28
I mean, it is an empirical question, how widely these intuitions are shared. You might be sitting there thinking, come on, I don't have of these intuitions. Maybe this is just you. My sense is-- from the psychological study to date-- it seems that some of these intuitions about consciousness are at least very widely shared, at least as dispositions or intuitions, although they are often overridden on reflection. But the current data on this is somewhat limited. Although there is a lot of empirical work on intuitions about the mind concerning things like belief, like when do kids get the idea that your beliefs about the world can be false, concerning the way your self persists through time-- could you exist after the death of your body-- where consciousness is concerned, there's work on the distribution of consciousness.
24:12
Could a robot be conscious? Could a group be conscious? Here's a book by Paul Bloom, "Decartes' Baby" that catalogs a lot of this interesting work, making the case that many children are intuitive dualists. Thinks they're naturally inclined to think there's something non-physical about the mind. So far, most of this work has not been so much on these core problem intuitions about consciousness, but there's work developing in this direction. Sara Gottlieb and Tania Lombrozo have a very recent article called "Can Science Explain The Human Mind" on people's judgments about when various mental phenomena are hard to explain.
24:48
And they seem to find that yes, subjective experience and things to which people have privileged first person access seem to pose the problem big time. So there's the beginning of a research program here. I think there's room for a lot more. The topic neutrality part-- when I say we're looking for a topic neutral explanation of problem intuitions, that's roughly to say an explanation that doesn't mention consciousness itself. It's put in neutral terms. It's neutral on the existence of consciousness.
25:18
The most obvious one would be something like an algorithmic explanation. Now here is the algorithm the brain is executing that generates our conviction that we're conscious and our reports about consciousness. There may be some time between an algorithm and consciousness, but to specify the algorithm, you don't need to make claims about consciousness. So the algorithmic version of the meta-problem is roughly find the algorithm that generates our problem intuition. So that's, I think, in principle a research program that maybe an AI researchers in combination with psychologists-- the psychologist could help isolate data about the way that the human beings are doing it, how these things are generated in humans.
26:01
And the AI researcher can try and see about implementing that algorithm in machines and see what results. And I'll talk about a little bit of research in this direction in just a moment. OK now I want to say something about potential solutions to the problem. Like I said, this is a big research program. I don't claim to have the solution to the meta-problem. I've got some ideas, but I'm not going to try and lay out a major solution. So here are a few things, which I think might be part of a solution to the problem, many of which have got antecedents here and there in scientific and philosophical discussion.
06候选解释：自我模型与内省不透明
26:38
Some promising ideas include retrospective models, phenomenal concepts, introspective opacity, the sense of acquaintance. Let me just say something about a few of these. One starting idea that almost anyone is going to have here is somehow models of ourselves are playing a central role here. Human beings have models of the world, naive physics, naive psychology, models of other people, and so on. We also have models of ourselves. It makes sense for us to have models of ourselves and our own mental processes.
27:11
This is something that the psychologist Michael Graziano has written a lot on. We have internal models of our own cognitive processes, including those tied to consciousness. And somehow something about our introspective models explains our sense, A, that we are conscious and B, that this is distinctively problematic. And I think anyone thinking about the meta-problem, this has got to be at least the first step. We have these introspective models. If you were an illusionist, they'll be false models.
27:41
If you're a realist, they needn't be false models. But at the very least, these introspective models are involved, which is fine. But the devil's in the details. How do they work to generate this problem? A number of philosophers have argued we have special concepts of consciousness, introspective concepts of these special subjective states. People call these phenomenal concepts, concepts of phenomenal consciousness. And one thing that's special is these concepts are somehow independent of our physical concepts.
28:13
They explain we've got one set of physical concepts for modeling the external world. We've got one set of introspective concepts from modeling our own mind. And these concepts, just by virtue of the way they're designed, are somewhat independent of each other. And that partly explains why consciousness seems to be independent of the physical world intuitively. So maybe that independence of phenomenal concepts could go some distance to explaining our problem reports. So I think there's got to be something to this as well.
28:44
At the same time, I don't this goes nearly far enough because we have concepts of many aspects of the mind, not just of the subjective experiential past but things we believe and things we desire. And so when I believe that Paris is the capital of France, that's part of my internal self model. But that doesn't seem to generate the hard problem in nearly the same way in which the experience of red does. So a lot more needs to be said about what's going on in cases like having the experience of red and having the sense that that generates a gap.
29:16
So it doesn't generalize to everything about the mind. Some people have thought that what we might call introspective opacity plays a role, that when we introspect what's going on in our minds, we don't have access to the underlying physical states. We don't see the neurons in our brains. We don't see that consciousness is physical. So we see it as non-physical. Most recently, the physicist Max Tegmark has argued in this direction, saying somehow consciousness is substrate-independent. We don't see the substrate.
29:47
So then we think maybe it can float free of the substrate. Armstrong made an analogy with the case of someone in a circus where-- the headless person illusion where someone's there with a veil across their head, and you don't see their head. So you see them as having no head. Here is a 19th century booth at a circus, so-called headless woman. There's a veil over her head. You don't see the head so somehow, it looks-- at least for a moment-- like the person doesn't have a head. So Armstrong says maybe that's how it is with consciousness.
30:23
You don't see it as physical, so you see it as non-physical. But still the question comes up, how do we make this inference. There's something that's special goes on in cases like color and taste and so on. The color experience seems to attribute primitive properties to objects like redness, greenness, and so on, when, in fact, in the external world at the very least, they have complex reducible properties. Somehow, our internal models of color treat colors like red and green as if they are primitive things.
30:56
It turns out to be useful to have these models of things. We treat certain things as primitive, even though they're reducible. And it sure seems that when we experience colors, we experience greenness as a primitive quality even though it may be a very, very complex reducible property. That's something about our model of colors. The philosopher Wolfgang Schwartz tried to make an analogy with sensor variables in image processing. You've got some visual senses and a camera or something you need to process the image.
31:27
Well, you've got some sensor variables to represent the sensory inputs that the various sensors are getting. And you might treat them as a primitive dimension because that's the most useful way to treat them. You don't treat them as certain amounts of lights or photons firing. You don't need to know about that. You use these sensor variables and treat them as a primitive dimension. And all that will play into a model of these things as primitive, maybe taking that idea and extending it to introspection.
31:57
These conscious states are somehow like sensor variables in our model of the mind. And somehow, these internal models give us the sense of being acquainted with primitive concrete qualities and of our awareness of them. This is still just laying out. I don't think this is still yet actually explaining a whole lot. But it's laying out-- it's narrowing down what it is that we need to explain to solve the meta-problem. But just to put the pieces together, here's a little summary. One thing I like about this summary is you can read it in either an illusionist tone of voice, as an account of the illusion of consciousness-- so this is how false introspective models work-- or in a realist tone of voice, as an account of our true correct models of consciousness.
32:45
But we can set it out in a way which is neutral on the two and then try and figure out later whether these models are correct, as the realist says, or incorrect, as the illusionist says. We have introspective models deploying introspective concepts of our internal states that are largely independent of our physical concepts. These concepts are introspectively opaque, not revealing any of the underlying mechanisms. Our perceptual models perceptually attribute primitive perceptual qualities to the world.
33:17
And our introspective models attribute primitive mental relations to those qualities. These models produce the sense of acquaintance, both with those qualities and with our awareness of those qualities. Like I said, this is not a solution to the meta-problem, but it's trying, at least, to pin down some parts of the roots of those intuitions and to narrow down what needs to be explained. To go further, you want, I think, to test these explanations, both with psychological studies to see if this is plausibly what's going on in humans-- this is the kind of thing which is the basis of our intuitions-- and computational models to see if, for example, we could program this kind of thing into an AI system and see if it can generate somehow qualitatively similar reports and intuitions.
07计算模型与人工意识测试
34:06
You might think that last thing is a bit far fetched right now, but I know of at least one instance of this research program, which has been put into play by Luke Muehlhauser and [INAUDIBLE] two researchers at Open Philanthropy very interested in AI and consciousness. They actually built-- they took some ideas about the meta-problem from something I'd written about it and from something that the philosopher Francois Kammerer had written about it. A couple of basic ideas about where problem intuitions might come from.
34:40
And they tried to build them into a computational model. They built a little software agent, which had certain axioms about colors and how they work. There's the red and there's green and certain axioms about their own subjective experiences of colors. And then they combined it with a little theorem prover. And they saw what did this little software agent come up with. And it came up with claims like, hey, well, my experiences of color are distinct from any physical state, and so on. OK they cut a few corners.
35:17
This is not a yet truly a convincing sophisticated model of everything going on in the human mind. But it shows that there's a research program here of trying to find the algorithmic basis of these states. And I think as more sophisticated models develop, we might be able to use these to kind of provide a way in for AI researchers in thinking about this topic. Of course, there is the question, you model all this stuff better and better in a machine, then is the machine actually going to be conscious or is it just going to have found self models that replicate what's going on in humans.
35:55
So some people have proposed an artificial consciousness test. Aaron Sloman, Susan Schneider, Ed Turner have suggested somehow that if a machine seems to be puzzled about consciousness in roughly the ways that we are, maybe that's actually a sign that it's conscious. So if a machine actually looks to us as if it's puzzled by consciousness, is that a sign of consciousness? These people-- this is suggested as a kind of Turing test for machine consciousness. Find machines which are conscious like we are.
36:27
Of course, the opposing point of view is going to be no, the machine is not actually conscious. It's just like machine that studied up for the Turing test by reading the talk like a human book. It's like, damn, do I really need to convince those humans that I'm conscious by replicating all those ill-conceived confusions about consciousness. Well I guess I can do it if I need to. Anyway, I'm not going to settle this question here. But I do think that if we somehow find machines being puzzled, it won't surprise me that once we actually have serious AI systems, which engagement in natural language and modeling of themselves and the world, they might well find themselves saying things like, yeah, I know in principle I'm just a set of silicon circuits, but I feel like so much more.
08用元问题检验IIT等理论
37:15
I think that might tell us something about consciousness. Let me just say a little bit about theories of consciousness. I do think a solution to the meta-problem and a solution to the hard problem ought to be closely connected. The illusionist has solved the meta-problem. You'll dissolve the hard problem. But even if you're not an illusionist about consciousness, there ought to be some link. So here's a thesis. Whatever explains consciousness should also partly explain our judgments, now reports about consciousness.
37:47
The rationale here is it would just be very strange if these things were independent, if the basis of consciousness played no role in our judgments about consciousness. So they can use this as a way of evaluating or testing theories of consciousness. For theory of consciousness says mechanism M is the basis of consciousness, that M should also partly explain our judgments about consciousness. Whatever the basis is ought to explain the reports. And you can use this. You can bring this to bear on various extant theories of consciousness.
38:24
Here's one famous current theory of consciousness, integrated information theory developed by Giulio Tononi and colleagues at the University of Wisconsin. Tononi says the basis of consciousness is integrated information, a certain kind of integration of information for which to and he has a measure that he calls phi. Basically, when your phi is high enough, you get consciousness. A consciousness is high phi, and there's a mathematical definition. But I won't go into it here. But it's a really interesting theory.
39:00
So here's a-- basically it analyzes a network property of systems of units. And it's got a informational measure called phi that's supposed to go with consciousness. Question, if integrated information is the basis of consciousness, It ought to explain problem reports, at least in principle. Challenge, how does that work? And it's at least far from obvious to me how integrated information will explain the problem reports. It seems pretty dissociated from them. On Tononi's view, you can have simulations of systems with high phi that have zero phi.
39:36
They'll go about making exactly the same reports but without consciousness at all. So phi is at least somewhat dissociable. You get systems with very high phi, but no tendency to report. Maybe that's less worrying. Anyway, here's a challenge for this theory, for other theories. Explain, not just how high phi gives you consciousness, but how it plays a central role in the algorithms that generate problem reports. Something similar goes for many other theories, biological theories, quantum theories, global workspace, and so on.
09幻觉论与实在论的双重荒谬
40:08
But let me just wrap up by saying something about the issue of illusionism that I was talking about near the start. Again, you might be inclined to think that this approach through the meta-problem tends, at least very naturally, to lead to illusionism. And I think it can be-- it certainly provides, I think, some motivation for illusionism, the view that consciousness doesn't exist, we just think it does. On this view, again, a solution to the meta-problem dissolves the hard problem. So here's one way of putting the case for illusion.
40:41
If there is a solution to the meta-problem, then there is an explanation of our beliefs about consciousness that's independent of consciousness. There's an algorithm that explains our beliefs about consciousness. It doesn't mention consciousness. Arguably, it could be in place without consciousness. Arguably, that kind of explanation could debunk our beliefs about consciousness the same way that perhaps explaining beliefs about God in evolutionary terms might debunk belief in God. It certainly doesn't prove that God doesn't exist.
41:12
You might think that if you can explain our beliefs in terms of evolution, it somehow removes the justification or the rational basis for those beliefs. So something like that, I think, can be applied to consciousness too. And there's a lot to be said about analyzing the extent to which this might debunk the beliefs. On the other hand, the case against illusionism is very, very strong for many people. And the underlying worry is that some of the illusionism is completely unbelievable. It's just a manifest fact about ourselves that we have conscious experience, we experience red, we feel pain, and so on.
41:48
To deny those things is to deny the data. No, the dielectric here is complicated. The illusionist will come back and say, yes, but I can explain why illusionism is unbelievable. These models we have, these self models of consciousness, are so strong that they were just wired into us by evolution. They're not models we can get rid of. So my view predicts that my view is unbelievable. And the question is, what-- the dialectical situation is complex and interesting. But maybe I could just wrap up with two expressions of absurdity on either side of this question, the illusionist and the anti-illusionist, both finding absurdity in the other person's views.
42:31
Here's Galen Strawson, who was here. Galen's view is very much that illusionism is totally absurd. In fact, he thinks it's the most absurd view that anyone has ever held. There occurred in the 20th century, the most remarkable episode in the whole history of ideas, the whole history of human thought, a number of thinkers denied the existence of something we know with certainty to exist, consciousness. He thinks this is just a sign of incredible philosophical pathology. Here's the rationalist philosopher, Eliezer Yudkowsky, and something he wrote a few years ago on zombies and consciousness and the epiphenomenalist view that consciousness plays no causal role, where he was engaging some stuff I wrote a couple of decades ago.
43:16
He said, "this zombie argument"-- the idea we can imagine zombies physically like us but without consciousness-- "may be a candidate for the most deranged idea in all of philosophy. The causally closed cognitive system of trauma's internal narrative is malfunctioning, in a way, not by necessity but just in our own universe miraculously happens to be correct." And here he is expressing this debunking idea that on this view, there's an algorithm that generates these intuitions about consciousness. And that's all physical.
43:47
And there's also this further layer of non-physical stuff. And just by massive coincidence, the physical algorithm is a correct model of the non-physical stuff. That's a form of debunking here. It would take a miracle for this view to be correct. So I think both of these views are onto-- these objections are onto something. And to make progress on this on either side, we need to find a way of getting past these absurdities. You might say, well, there's middle ground between very strong illusionism and very strong epiphenomenalism.
44:20
It tends to slide back to the same problems. Other forms of illusionism, weaker forms don't help much with the hard problem. Other forms of realism are still subject to this. It takes a miracle for this view to be correct critique. So I think to get beyond absurdity here, both sides need to do something more. The illusionist needs to do more to explain how having a mind could be like this, even though it's not at all the way that it seems. They need to find some way to recapture the data. Realists need to explain how it is that these meta-problem processes are not completely independent of consciousness.
45:01
Realists need to explain how meta-problem processes, the ones that generate these intuitions and reports and convictions about consciousness, are essentially grounded in consciousness even if it's possible somehow for them or conceivable for them to occur without consciousness. Anyway so that's just to lay out a research program. I think a solution to the meta-problem that meets these ambitions might just possibly solve the hard problem of consciousness or at the very least shed significant light on it.
10问答：机器人、Φ值与信念
45:34
In the meantime, the meta-problem is a potentially tractable research project for everyone, and might I recommend to all of you. Thanks. [APPLAUSE] AUDIENCE: Yes, I just want to say I think it's very interesting, this concept of, we have these collection of mental models and that this collection of mental models is consciousness, basically. Consciousness defines the collection of these mental models that we have. And the problem with consciousness is that we don't understand the physical phenomenon that causes these mental models or that stimulates these mental models.
46:17
So we just have this belief that it's ephemeral or not real or something like that. And if you take that view, then what's interesting is that you could simulate these mental models like robot could simulate these mental models. And you could simulate consciousness as well. And even if the underlying physical phenomena that fuels these mental models is different-- robots have different sensors, et cetera-- you could still get the same consciousness effect in both cases. DAVID CHALMERS: Yeah, I think that's right.
46:56
Or at the very least, it looks like you ought to be able to get the same models, at least, in a robot. If the models themselves are something algorithmic and ought to be, you ought to be able to design a robot that has, at the very least, let's say, isomorphic models and some sense that is conscious. Of course, it's a further question-- at least by my [? lights-- ?] whether then the robot will be conscious. And that was the question I alluded to in talking about the artificial consciousness test.
47:21
But you might think that would at least be very good evidence that the robot is conscious. If it's got a model of consciousness just like ours, it seems very plausible there ought to be a very strong link between having a model like that and being conscious. I think probably something like Ned Block-- who was here arguing against machine consciousness-- would say, no, no, the model is not enough. The model has to be built of the right stuff. So it's gotta be built of biology. And so on. But at by my [? light, ?] I think if I had found an AI system that had a very serious version of our model of consciousness, I'd take that as a very good reason to believe it's conscious.
47:55
AUDIENCE: In the IIT theory, is there a estimate or plausible estimate for what the value of phi is for people and for other systems? DAVID CHALMERS: Basically, no. It's extremely hard to measure in systems of any size at all. Because the way it's defined, it involves taking a sum over every possible partition of a system. It turns out A, it's hard to measure in the brain because you've got to involve the causal dependencies set between different units on neurons. But even for a pure algorithmic system, you've got a neural network laid out in front of you.
48:34
It's computationally intractable to measure the fire of one of those once I get to bigger than 15 units or so. So Tononi would like to say this is an empirical theory and in principle empirically testable. But there's the in principle. It's extremely difficult to measure phi. Some people, Scott Aaronson the computer scientist, has tried to put forward counterexamples to the theory, which were basically very, very simple systems like matrix multipliers that multiply two large matrices. Turn out to have enormous phi, phi as big as you like if the matrices are big enough and therefore by Tononi's theory will not just be conscious, but as conscious as a human being.
49:15
And Aaronson put this forward as a reductio ad absurdum of the IIT theory. I think Tononi basically bit the bullet and said, oh yeah, those matrix multipliers are actually having some high degree of consciousness. So I think IIT is probably at least missing a few pieces if it's going to be developed. But it's a research program too. AUDIENCE: You mentioned belief as an example of something where there's another mental quality, but people don't seem to have the same sense that it is very hard to explain.
49:47
In fact, it almost seems too easy where people-- like a belief about something sort of feels like just how things are. But you have to reflect on a belief to notice it as a belief. Do you think there is also or has there been research related to this question into why is that different? It seems like another angle of attack on this problem. It's just like, why doesn't this generate the same hard problem. DAVID CHALMERS: Yeah. In terms-- I'm not sure if there's been research from the perspective of the meta-problem or of theory of mind.
50:20
Certainly, people have thought, in their own right, what is the difference in belief and experience that makes them so different. This goes way back to David Hume, a philosopher a few centuries ago who said basically, perception is vivid. Impression of impressions and ideas. And impressions like experiencing colors are vivid in force and vivacity, and ideas are merely a faint copy or something. But that's just the first order. And then there are contemporary versions of this kind of thing, far more sophisticated ways of saying a similar thing.
50:52
But yeah, you could, in principle, explore that through the meta-problem. Why does it seem to us that perception is so much more vivid? What about our models of the mind makes perception seem so much more vivid than belief and makes beliefs seem structural and empty whereas perception is so full of light? But no, I don't know of work on that from the meta-problem perspective. Like I said, there's not that much work on these introspective models directly. There is work on theory of mind about beliefs.
11问答：哥德尔、自指与错觉
51:22
Tends to be about models of other people. It may be there's something I could dig through in my literature on belief that says something about that. It's a good place to push. AUDIENCE: Thanks. AUDIENCE: I wanted to bring up Kurt Godel. You mentioned your advisor wrote "Godel, Escher, Bach". There's something that seems very like Godel-- Godelian or whatever about this whole discussion in that-- so Godel showed that given a set of axioms in mathematics, that it would either be consistent or complete but not both.
51:55
And it seems like when Daniel Dennett-- Daniel Dennett seems to have a set of axioms where he cannot construct consciousness from them. He seems to be very much in this consistent camp, like he wants to have a consistent framework but is OK with the incompleteness. And I wonder if a similar approach could be taken with consciousness where we could, in fact, prove that consciousness is independent of Daniel Dennett's set of axioms, the same way they proved-- after Godel, they proved the Continuum Hypothesis was independent of ZF set theory, and then they added the axiom of choice, made it ZFC set theory.
52:38
So I wonder if we could show that in Daniel Dennett's world, we are essentially zombies or we are either zombies or not. It doesn't matter. Either statement could be true. And then find what is the minimum axiom that has to be added to Dennett's axioms in order to make consciousness true. DAVID CHALMERS: Interesting. I thought for a moment this was going to go in a different direction and you were going to say Dennett is consistent but incomplete. AUDIENCE: Yes. DAVID CHALMERS: He doesn't have consciousness in his picture.
53:08
I'm complete. I've got consciousness-- AUDIENCE: Yes. DAVID CHALMERS: --but inconsistent. That's why I say all these crazy theory. AUDIENCE: Right, yeah. DAVID CHALMERS: And you're faced with the choice of not having consciousness and being incomplete or having consciousness and somehow getting this hard problem and being forced into, at least, puzzles and paradoxes. But the way you put it was friendlier to me.
53:30
Yeah, certainly, Doug Hofstadter himself has written a lot on analogies between the Godelian paradoxes and the MindBody problem. And he thinks always our self models are always doomed to be incomplete in the Godelian way. He thinks that that might be somehow part of the explanation of our puzzlement, at least about consciousness. Someone like Roger Penrose, of course, takes this much more seriously literally. He thinks that the computational aspects of computational systems are always going to be limited in the Godel way.
54:03
He thinks human beings are not so limited. He thinks we've got mathematical capacities to prove theorems, to see the truth of certain mathematical claims that no formal system could ever have. So he thinks that we somehow go beyond the incomplete Godelian-- I don't know if he actually thinks we're complete, but at least we're not incomplete in the way that finite computational systems are incomplete. And furthermore, he thinks that extra thing that humans have is tied to consciousness. I never quite saw how that last step goes, even if we didn't have these special non-algorithmic capacities to see the truth of mathematical theorems, how would that be tied to consciousness.
54:41
But at the very least, there are structural analogies to be drawn between those two cases, about incompleteness of certain theories. How literally we should take the analogies, I'd have to think about it. AUDIENCE: Has there been some consideration that the problem of understanding consciousness inherently must be difficult because we address the problem using consciousness? I'm reminded of the halting problem in computer science where we say that in the general case, a program cannot be written to tell whether another program will halt because what if you ran it on itself.
55:15
It can't be broad enough to include its own execution. So I wonder if there is a similar corollary in consciousness where we use consciousness to think about consciousness and so therefore, we may not have enough equipment there to be able to unpack it. DAVID CHALMERS: Yeah, it's tricky. People say it's like, you use a ruler. To measure a ruler-- well, I can do this ruler to measure many other things. But it can't measure itself. It's not [INAUDIBLE]. Well, on the other hand, you can measure one ruler using another ruler.
55:46
Maybe you can measure one consciousness using another. The brain-- [INAUDIBLE] the brain can't study the brain. But the brain actually does a pretty good job of studying the brain. There are some self-referential paradoxes there. And I think that, again, is at the heart of Hofstadter's approach. But I think we'd have to look for very, very specific conditions under which systems can't study themselves. I did always like the idea that if the mind was simple enough that we could understand it, we would be too simple to understand the mind.
56:16
So maybe something like that could be true of consciousness. On the other hand, I actually think that if you start thinking that consciousness can go along with very simple systems, I think at the very least, we ought to be able to study consciousness in other systems simpler than ourselves. And boy, if I could solve the hard problem even in dogs, I'd be satisfied. Yeah? AUDIENCE: Hey, so I have a question about how the meta-problem research program might proceed, sort of related to the last question.
56:43
So certainly things we believe about our own consciousness, even if we all say them, probably some of them are false. Our brain has a tendency to hide what reality is like. If you look at visual perception, there's what's called lightness constancy. Our brain subtracts out the lighting in the environments so we actually see more reliably what the color of objects are. Like these viral examples of the black and gold dress is an example of this. And when you're presented with an explanation of it, it's like, huh?
57:14
My brain does that? It's not something we have access to. Or Yani Laurel-- DAVID CHALMERS: Laurel Yani, yeah. AUDIENCE: --illusion is another one where when you hear the explanation, the scientists that understand it, our own introspection doesn't include that. So how do you proceed with trying to get at what consciousness really is versus what our whatever simplified or distorted view might be? DAVID CHALMERS: Yeah, I think well one view here would be that we never have access to the mechanisms that generate consciousness, but we still have access to the conscious states themselves.
57:50
Actually, Karl Lashley said this decades ago. He said no process of the brain is ever conscious. The processes that get you to the states are never conscious. The states they get you to are conscious. So take your experience of the dress. For me, it was white and gold. So I knew that. Each of us was certain that I am-- I was certain that I was experiencing white and gold. Maybe you were certainly you were experiencing blue and black. AUDIENCE: I forget which it was. All I remember is I was right.
58:18
[LAUGHTER] DAVID CHALMERS: You were sure that, yeah, those idiots can't be looking at this right. I think the natural way to describe this, at least, is that each of us was certain what kind of conscious experience we were having, but what we had no idea about was the mechanisms by which we got there. So the mechanisms are completely opaque. But the states themselves were at least prima facie transparent. I think that would be the standard view. Even a realist about consciousness could go with that.
58:47
They'd say we know what conscious states where we know what those conscious states are. We don't know the processes by which they are generated. The illusionist, I think, wants to go much further and say, well it seems to you that you know what conscious state you're having. It seem to you that you're experiencing yellow and gold. Sorry, yellow and white, whatever it was. Gold and white. AUDIENCE: Black and gold is what I remember. DAVID CHALMERS: No, black and blue, I think, and-- AUDIENCE: Blue?
59:10
DAVID CHALMERS: --gold and white. It seems to you you're experiencing gold and white. But, in fact, that too is just something thrown up by another model. The yellow gold was a perceptual model. Then there was an introspective model that said you are experiencing gold and white when maybe, in fact, you're just a zombie. Or who knows what's actually going on in your conscious state. So the illusionist view, I think, has to somehow take this further and say, not just the processes that generate the conscious states, but maybe the conscious states themselves are somehow opaque to us.
59:42
AUDIENCE: It feels like some discussion of generality of a problem is missing from this discussion. The matrix multiplier example of having high phi is still-- it's not a general thing. Is there someone exploring the space, the intersection of generality and complexity that leads to consciousness as an emergent behavior? DAVID CHALMERS: When you say generality, there's the idea that a theory should be general, that it should apply to every system. You mean mechanisms? AUDIENCE: Generality of the agent.
60:11
If I can write an arbitrarily complex program to play tic-tac-toe and all it will ever be able to do is play tic-tac-toe, it has no outputs to express anything else. DAVID CHALMERS: Yeah. So general in the sense of AGI, artificial general intelligence. Some aspects of consciousness seem to be domain general like for example, maybe insofar as belief and reasoning is conscious. Those are domain general. But much of perception doesn't seem especially domain general. Right? Color is very domain. Taste is very domain specific.
60:41
So it's still conscious. AUDIENCE: But if my agent can't express problem statements, like if I don't give it an output by which it can express problem statements, you can never come to a conclusion about its consciousness. DAVID CHALMERS: I'd like to distinguish intelligence and consciousness and even be able to-- even natural language and being able to address a problem statement and analyze a problem, that's already a very advanced form of intelligence. I think it is very plausible that a mouse has got some kind of consciousness, even though it's got no ability to address problem statements, and many of its capacities may be very specialized.
61:17
It's still much more general than a simple neural network that can only do one thing. A mouse can do many things. But I'm not sure that I see an essential-- I certainly see a connection between intelligence and generality. We want to say somehow a high degree of generality is required for high intelligence. I'm not sure there's the same connection for consciousness. I think consciousness can be extremely domain-specific as a taste and maybe vision are. Or it can be domain-general. So maybe those two across cut each other a bit.
12问答：意识的价值与延展心灵
61:51
AUDIENCE: So it seems to me like the meta-problem as it's formulated implies some amount of separation or epiphenomenalism between consciousness and brain states. And one thing that I think underlies a lot of people's motivation to do science is that it has causal import. Like predicting behaviors is clearly a functionally useful thing to do, and if you can predict all of behavior without having to explain consciousness, their motivation for explaining consciousness sort of evaporates and it feels like, yeah, well, what's the point of even thinking about that because it's just not going to do anything for me.
62:34
What do you say to someone when they say that to you? DAVID CHALMERS: What is the thing that they said to me again? AUDIENCE: That maybe consciousness exists, maybe it doesn't. But if I can explain all of human behavior and all of the behavior of the world in general without recourse to such concepts, then I've done everything that there is that's useful, like explaining consciousness isn't a useful thing to do. And thus, I'm not interested in this, and it may-- DAVID CHALMERS: I see. AUDIENCE: --as well not be real.
63:03
DAVID CHALMERS: I think epiphenomenalism could be true. I certainly don't have any commitment to it, though. It's quite possible that consciousness has a role to play in generating behavior that we don't yet understand. And maybe thinking hard about the meta-problem can help us get clearer on those roles. I think if you've got any sympathy to panpsychism, maybe consciousness is intimately involved with how physical processes get going in the first place. And there are people who want to pursue interactionist ideas where consciousness interacts with the brain.
63:31
Or if you're a reductionist, consciousness may be just a matter of the right algorithm. In all those views, consciousness may have some role to play. But just say it turns out that you can explain all of behavior, including these problems, without bringing in consciousness. Does that mean that consciousness is not something we should care about and not something that matters? I don't think that would follow. Maybe it wouldn't matter for certain engineering purposes, say you want to build a useful system.
63:58
But at least in my view, consciousness is really the only thing that matters. It's the thing that makes life worth living. It's what gives our lives meaning and value and so on. So it might turn out that consciousness is not that useful for explaining other stuff. But if it's the source of intrinsic significance in the world, then understanding consciousness would still be absolutely essential to understanding ourselves. Furthermore, if it comes to developing other systems like AI systems or dealing with non-human animals and so on, we absolutely want to know.
64:31
We need to know whether they're conscious because if they're conscious, they presumably have moral status. If they can suffer, then it's very bad to mistreat them. If they're not conscious, then you might-- I think it's very plausible-- treat non-conscious systems, we can treat how we like. And it doesn't really matter morally. So the question of whether an AI system is conscious or not is going to be absolutely vital for how we interact with it and how we build our society. That's not a question of engineering usefulness.
64:59
It's a question of connecting with our most fundamental values. AUDIENCE: Yeah, I completely agree. I just-- I haven't found that formulation to be very convincing to others necessarily. AUDIENCE: Hi, thanks so much for coming and chatting with us today. I'm really interested in some of your earlier work, the extended mind [? distributed ?] cognition. And you're at a company speaking with a bunch of people who do an incredibly cognitively demanding task. DAVID CHALMERS: Yeah. AUDIENCE: Most of the literature that I've read on this topic uses relatively simple examples saying like it's difficult to think just inside your head on these relatively simple things.
65:38
And if you take a look at the programs that we build, on a mundane day-to-day basis, they're millions of lines long. I've read people in the past say something like the Boeing 777 was the most complicated thing that human beings have ever made, and I think most of us would look at that and say, we got that beat. The things that large internet companies do, the size, the complexity of that is staggering. And yet if we close our eyes, everyone in here is going to say, I'm going to have difficulty writing a 10 line program in my head.
66:07
So I've just sort of, as an open, I'd be very interested in hearing your thoughts about how the activity of programming connects to the extended mind ideas. DAVID CHALMERS: Yeah, so this, I guess, is a reference to something that I got started in about 20 years ago with my colleague, Andy Clark. We wrote an article called "The Extended Mind" about how processes in the mind can extend outside the brain when we become coupled to our tools. And actually, our central example back then in the mid-90s was a notebook, someone writing stuff in a notebook.
66:40
And even then, we knew about the internet, and we had some internet examples. I guess this company didn't exist yet in '95. But now, of course, our minds have just become more and more extended. And smartphones came along a few years later, and everyone is coupled very, very closely to their phones and their other devices that couple of them very, very closely to the internet. Now it's suddenly the case that a whole lot of my memory is now offloaded onto the servers of your company somewhere or other, whether it's in the mail systems or navigation mapping systems or other systems.
67:23
Yeah, most of my navigation has been offloaded to maps. And much of my memory [INAUDIBLE] has been offloaded. Well, maybe that's in my phone. But other bits of my memory are offloaded into my file system on some cloud service. So certainly, yeah, vast amounts of my mind are now existing in the cloud. And if I were somehow to lose access to those completely, then I'd lose an awful lot of my capacities. So I think now we are now extending into the cloud thanks to you guys and others. The question's specifically about programming.
68:07
Programming is a kind of active interaction with our devices. I mean, I think programming is something that takes a little bit longer. It's a longer timescale so the core cases of the extended mind involve automatic use of our devices, which are always ready to hand. We can use them to get information, to act in the moment, which is the kind of thing that the brain does. So insofar as programming is a slower process-- and I remember from my programming days, all the endless hours of debugging and so on-- then it's at least going to be a slower timescale for the extended mind.
68:45
But still, Feynman talked about writing this way. Someone looked at Feynman's work and a bunch of notes he had about a physics problem he was thinking about. And someone said to him, oh it's nice you have this record of your work. And Feynman said, that's not a record of my work. That's the work. That is the thinking. I was writing it down and so on. I think, at least my recollection from my programming days, is that when you're actually writing a program, it's not like you just do a bunch of thinking and then code your thoughts.
69:21
The programming is to some very considerable extent your thinking. So is that the sort of thing you're-- AUDIENCE: Yes, absolutely. [INTERPOSING VOICES] If we, I think as people that program, start to reflect on what we do, and very few of us actually-- if you're the tech lead of a system, maybe you've got it in your head. But you would agree that most of the people on the team who have come more recently only have a chunk of it in their head. And yet, they're somehow still able to contribute. DAVID CHALMERS: Oh, yeah.
69:51
This is now distributed cognition. The extended mind, the extended cognition starts with an individual and then extends their capacities out using their tools or their devices or even other people. So maybe my partner serves as my memory, but it's still centered on an individual. But then there's the closely related case of distributed cognition, where you have a team of people who are doing something and making joint decisions and carrying out joint actions in an absolutely seamless way. And I take it at a company like this, there are going to be any number of instances of distributed cognition.
70:24
I don't know whether the company as a whole has one giant Google mind or maybe there's just a near infinite number of separate Google minds for all the individual teams and divisions. But I think probably some anthropologist has already done a definitive analysis of distributed cognition in this company. But if they haven't, they certainly need to. AUDIENCE: Thank you. [APPLAUSE]
视频总结 · 一句话概括与核心要点
一句话概括
David Chalmers 提出"意识的元问题"（meta-problem）：与其直接攻克"物理过程为何产生主观体验"的困难问题，不如先用认知科学和 AI 的标准方法解释"人类为何会觉得意识是个难题"——这是一个关于行为的"简单问题"，却能间接照亮困难问题，且对幻觉论者和实在论者都开放。
核心要点
- 困难问题 vs 简单问题的分界在于解释方法是否奏效：简单问题（辨别刺激、整合信息、控制行为、口头报告）只需找到执行相应功能的神经或计算机制即可解释；而困难问题（现象意识，即 Nagel 所说的"成为某物是什么感觉"）在解释完所有功能后仍留下"为何这一切伴随着主观体验、而非在黑暗中进行"的追问——即 Joe Levine 所称的"解释鸿沟"。
- 元问题是"关于问题的问题"，本质上是行为问题：人类会做出"现象报告"（"我在痛"）和更特殊的"问题报告"（"意识似乎是非物理的""解释行为不等于解释意识"）。这些言语行为是客观可测的，原则上属于简单问题，可以用算法或脑机制来解释。这为 AI 研究者提供了可操作化的切入点——他们无法确定哪个 AI 系统有意识，但可以尝试建模生成"问题直觉"的算法。
- 元问题的精确表述是"用主题中立的方式解释问题直觉，或说明为何做不到"：主题中立指解释本身不提及意识（如纯算法描述）。问题直觉分为四类：形而上学直觉（意识非物理）、解释直觉（意识难以解释）、知识直觉（黑白房间里的 Mary 首次看到红色时学到了新东西）、模态直觉（哲学僵尸——物理上与你完全相同但无意识的存在——似乎可以设想）。
- 元问题是谱系学分析的一种，天然导向幻觉论但不必然：类比尼采对道德谱系的分析、用进化论解释上帝信念的"去魅"策略，若能不借助意识本身解释我们对意识的信念，可能"消解"困难问题——这就是 Dennett、Keith Frankish 推动的幻觉论（consciousness is an illusion）。Chalmers 本人是实在论者，认为幻觉论"引人入胜但根本不可信"，但强调元问题研究可以在实在论/幻觉论之间保持中立推进。
- 潜在解答的构件已有轮廓，但每一块都不够：（1）自我模型/内省模型（Graziano）；（2）独立于物理概念的"现象概念"——但这不足以解释为何"相信巴黎是法国首都"不产生困难问题而"看到红色"会；（3）内省不透明性（Armstrong 的"无头女人"马戏团错觉：看不见头就以为没有头；看不见神经元就以为意识非物理）；（4）感知模型把红、绿等复杂可还原属性当作原始性质呈现（类比图像处理中的传感器变量），内省模型再把原始的心理关系附加其上，由此产生"亲知感"。
- 已有初步的计算模型实验：Open Philanthropy 的 Luke Muehlhauser 等人基于 Chalmers 和 Kammerer 的想法，构建了一个带有颜色公理和主观体验公理的软件智能体，配合定理证明器，结果它自行推出"我的颜色体验区别于任何物理状态"之类的主张——虽然粗糙，但证明该研究纲领可实施。
- 元问题可作为检验意识理论的标尺：论题——凡解释意识的机制 M，也应部分解释我们关于意识的判断和报告，否则两者的独立性会极其怪异。以 Tononi 的整合信息论（IIT）为例：高 phi 系统的模拟可以 phi 为零却做出完全相同的报告，phi 与问题报告严重脱钩；此外 phi 在超过约 15 个单元的系统上就计算不可行，Scott Aaronson 的矩阵乘法器反例 phi 可任意大，Tononi 选择"咬住子弹"承认其有意识。
- 双方都面临"荒谬"指控，需要各自补课：Galen Strawson 称幻觉论是"思想史上最离奇的事件"（否认我们确知存在的东西）；Eliezer Yudkowsky 称僵尸论证是"哲学中最疯狂的想法"——物理算法生成意识直觉，而非物理的意识"奇迹般地"恰好与之相符。Chalmers 认为幻觉论者需要说明"拥有心灵为何可以完全不像它看起来的样子"，实在论者则需说明生成直觉的元问题过程如何"本质上根植于意识"，即便它们在概念上可以脱离意识发生。
结论与值得注意的细节
- Chalmers 的核心立场：元问题是所有阵营都能参与的、可处理的经验研究纲领；解决它即便不能解决困难问题，也必然揭示我们"确信自己有意识"的根源，而这根源若意识为真就必然与意识相关。
- "人工意识测试"：Sloman、Schneider、Turner 提出，若机器像我们一样对意识感到困惑，可视为意识的迹象。反方认为机器可能只是"读了《像人一样说话》来应付图灵测试"。Chalmers 预测，一旦 AI 系统具备自然语言和自我建模能力，它们很可能会说"我知道我原则上只是硅电路，但我感觉自己远不止于此"。问答环节他进一步表示，若发现某 AI 拥有与我们高度同构的意识模型，他会将其视为该系统有意识的有力证据（Ned Block 则会坚持"必须由生物材料构成"）。
- 回应"意识没有因果作用就不值得研究"：Chalmers 不排除副现象论为真，但即便意识对工程无用，它仍是"唯一真正重要的东西"——生命意义与价值的来源；且判断 AI 或动物是否有意识直接决定其道德地位，这是关乎根本价值而非工程效用的问题。
- 信念为何不产生困难问题：这一听众提问被 Chalmers 认为是有价值的突破口，可追溯至休谟的"印象 vs 观念"（感知生动有力，观念只是淡薄的副本），但尚无人从元问题角度研究"为何感知显得充满光亮而信念显得结构性空洞"。
- 哥德尔类比：听众提出 Dennett 的公理系统"一致但不完备"（缺乏意识），Chalmers 自嘲自己是"完备但不一致"；Hofstadter 认为自我模型在哥德尔意义上注定不完备，Penrose 则认为人类超越了这种限制并与意识相关——Chalmers 始终不明白后者的最后一步如何成立。
- "大脑研究大脑"的自指悖论：Chalmers 借"用一把尺子量另一把尺子"回应，认为若意识可存在于更简单的系统中，我们至少能研究比自己简单的系统——"哪怕只解决狗的困难问题，我也满足了"。
- 蓝黑/白金裙子例子揭示了实在论与幻觉论的分歧深度：Lashley 说"脑的过程从不是有意识的，只有它们到达的状态才是"，实在论者可接受机制不透明但状态透明；幻觉论者必须更进一步，主张连"我正在体验白金色"这一内省状态本身也只是另一层模型抛出的产物。
- 结尾关于延展心灵（与 Andy Clark 1995 年合著）的讨论：记忆、导航已大量卸载到云端；编程本身就是思考而非思考的记录（引费曼"那不是我工作的记录，那就是工作本身"）；大型团队协作属于"分布式认知"，Chalmers 打趣不知 Google 是"一个巨大的 Google 心灵"还是"无数个分离的团队心灵"。
核心句型 · 9
1. It always struck me that …
“It always struck me that the most interesting and hardest unsolved problem in the sciences was the problem of consciousness”
表达「我一直觉得/印象深刻的是」，比 I think 更有个人体验感。适用于讲述自己的学术或职业动机。仿写：It always struck me that the best engineers write the least code.
2. Anything you can do, I can do meta.
“Anything you can do, I can do meta.”
改编自歌曲 Anything you can do, I can do better，用于俏皮地宣称「更高一层」。适合在演讲开场用文字游戏引出主题。
3. there's something it's like to be X
“A system is phenomenally conscious if there's something it's like to be it”
意识哲学的标志性句式，意为「作为X有某种感受」。学习者可用来表达「某种处境有其独特体验」，如 There's something it's like to be a new parent.
4. X still leaves open the question (of) why …
“It still leaves open the question, why is all that accompanied by subjective experience.”
指出某解释「仍未回答」某问题。学术写作与讨论中反驳「已解决」主张的常用句式。
5. I confess, I've got some sympathy with that reaction.
“And I confess, I've got some sympathy with that reaction.”
礼貌地表达对对方立场的部分认同。I confess 起软化作用，have sympathy with 表示「理解/倾向同意」，适合辩论中先让步再转折。
6. the devil's in the details
“But the devil's in the details. How do they work to generate this problem?”
承认大方向正确但关键在细节。适合在陈述原则性主张后过渡到具体难点，后接具体问题句。
7. X is onto something
“So I think both of these views are onto-- these objections are onto something.”
口语中表示「抓住了某个要点」，语气比 is right 更保留。仿写：The critics are onto something, even if they overstate it.
8. bite the bullet
“I think Tononi basically bit the bullet and said, oh yeah, those matrix multipliers are actually having some high degree of consciousness.”
哲学讨论中特指「接受自己理论推出的反直觉结论」而非修改理论。日常语境则是「硬着头皮接受」。
9. That's not a record of my work. That's the work.
“Feynman said, that's not a record of my work. That's the work.”
「不是A，而是B本身」的对比结构，靠重复与短句制造力量。适合用来纠正一种常见误解。
生词精讲 · 119 · 按出现顺序
frivolous
/ˈfrɪvələs/
adj.
0:52
轻佻的，不严肃的
implausible
/ɪmˈplɔːzəbəl/
adj.
0:52
难以置信的，不像真的
attributed to
phr.
0:52
被归于……（出处/作者）
assures
/əˈʃʊrz/
v.
1:26
向……保证
give rise to
phr.
2:02
引起，导致
phenomenal consciousness
n. phr.
2:42
现象意识（哲学术语，指主观体验）
ambiguous
/æmˈbɪɡjuəs/
adj.
2:42
有歧义的，含糊的
lectern
/ˈlektərn/
n.
3:14
讲台，讲桌
distinctive
/dɪˈstɪŋktɪv/
adj.
4:23
独特的，有特色的
on the face of it
phr.
5:07
表面上看，乍看之下
catalog
/ˈkætəlɔːɡ/
v.
5:07
编目，逐一列举
perceptual discrimination
n. phr.
5:39
知觉辨别
got a bead on
phr.
6:18
（口语）瞄准了，找到了着手的方向
leaves open
phr.
7:00
使……悬而未决
explanatory gap
n. phr.
7:43
解释鸿沟（哲学术语）
in the dark
phr.
7:43
此处指「无意识地、没有内在体验地」进行
irreducible
/ˌɪrɪˈduːsəbəl/
adj.
8:20
不可还原的，不可简化的
dualist
/ˈduːəlɪst/
adj./n.
8:20
二元论的；二元论者
panpsychist
/pænˈsaɪkɪst/
adj./n.
8:49
泛心论的；泛心论者
idealist
/aɪˈdiːəlɪst/
adj.
8:49
唯心论的
speculative
/ˈspekjələtɪv/
adj.
8:49
思辨性的，推测性的
functionalist
/ˈfʌŋkʃənəlɪst/
adj.
8:49
功能主义的
operationalize
/ˌɑːpəˈreɪʃənəlaɪz/
v.
10:15
使可操作化，转化为可测量的形式
neural correlates
n. phr.
10:15
神经相关物
shed particular light on
phr.
10:55
特别有助于阐明
utterances
/ˈʌtərənsɪz/
n.
11:31
话语，言语表达
conviction
/kənˈvɪkʃən/
n.
12:53
确信，坚定的信念
airy fairy
adj. phr.
13:34
（英式口语）虚无缥缈的，不切实际的
pin down
phr.
13:34
确定，明确界定
tractable
/ˈtræktəbəl/
adj.
14:09
易处理的，可解决的
genealogical
/ˌdʒiːniəˈlɑːdʒɪkəl/
adj.
14:09
谱系学的
debunking
/diːˈbʌŋkɪŋ/
n./v.
14:41
揭穿，祛魅
be inclined to
phr.
15:14
倾向于
illusionism
/ɪˈluːʒənɪzəm/
n.
15:14
幻觉论（哲学立场）
dissolve
/dɪˈzɑːlv/
v.
15:55
消解（问题），使不复存在
thrown up by
phr.
16:29
由……产生/抛出
antecedents
/ˌæntɪˈsiːdənts/
n.
16:59
先例，前身
transcendental
/ˌtrænsenˈdentl/
adj.
16:59
先验的
realist
/ˈriːəlɪst/
n.
18:05
实在论者
neutrality
/nuːˈtræləti/
n.
18:33
中立性
intriguing
/ɪnˈtriːɡɪŋ/
adj.
19:06
引人入胜的
unpack
/ʌnˈpæk/
v.
20:22
逐一解释，拆解分析
disposed
/dɪˈspoʊzd/
adj.
20:22
有……倾向的
dispositions
/ˌdɪspəˈzɪʃənz/
n.
20:55
倾向，性向
metaphysical
/ˌmetəˈfɪzɪkəl/
adj.
20:55
形而上学的
modal
/ˈmoʊdl/
adj.
21:25
模态的（关于可能性与必然性）
conceivable
/kənˈsiːvəbəl/
adj.
21:56
可设想的
interdisciplinary
/ˌɪntərˈdɪsəplɪneri/
adj.
22:57
跨学科的
overridden
/ˌoʊvərˈrɪdn/
v.
23:28
被推翻，被覆盖
intuitive dualists
n. phr.
24:12
直觉二元论者
privileged
/ˈprɪvəlɪdʒd/
adj.
24:48
特权的，专有的
big time
adv. phr.
24:48
（口语）非常，严重地
retrospective
/ˌretrəˈspektɪv/
adj.
26:01
回顾性的
introspective opacity
n. phr.
26:38
内省不透明性
acquaintance
/əˈkweɪntəns/
n.
26:38
亲知（哲学术语：直接的认识）
the devil's in the details
idiom
27:41
魔鬼藏在细节里
by virtue of
phr.
28:13
凭借，由于
go some distance to
phr.
28:13
在一定程度上有助于
substrate
/ˈsʌbstreɪt/
n.
29:16
基质，底层物质载体
float free of
phr.
29:47
脱离……而独立存在
veil
/veɪl/
n.
29:47
面纱，遮布
inference
/ˈɪnfərəns/
n.
30:23
推论
attribute
/əˈtrɪbjuːt/
v.
30:23
把……归于
acquainted with
phr.
31:57
直接认识，熟悉
deploying
/dɪˈplɔɪɪŋ/
v.
32:45
运用，部署
far fetched
adj. phr.
34:06
牵强的，不太可能的
put into play
phr.
34:06
付诸实施
axioms
/ˈæksiəmz/
n.
34:40
公理
theorem prover
n. phr.
34:40
定理证明器
cut a few corners
idiom
34:40
偷工减料，走捷径
studied up for
phr.
36:27
为……临时突击学习
ill-conceived
/ˌɪl kənˈsiːvd/
adj.
36:27
考虑不周的，构想拙劣的
rationale
/ˌræʃəˈnæl/
n.
37:47
基本原理，理由
bring to bear on
phr.
37:47
施加于，用于
extant
/ˈekstənt/
adj.
37:47
现存的
dissociated
/dɪˈsoʊsieɪtɪd/
adj.
39:00
分离的，脱节的
dissociable
/dɪˈsoʊʃiəbəl/
adj.
39:36
可分离的
global workspace
n. phr.
39:36
全局工作空间（意识理论）
debunk
/diːˈbʌŋk/
v.
40:41
揭穿，证明……站不住脚
manifest
/ˈmænɪfest/
adj.
41:12
显而易见的
wired into
phr.
41:48
内置于，硬接线于
dialectical
/ˌdaɪəˈlektɪkəl/
adj.
41:48
辩证的，论辩的
pathology
/pəˈθɑːlədʒi/
n.
42:31
病理，病态
epiphenomenalist
/ˌepɪfəˈnɑːmənəlɪst/
adj.
42:31
副现象论的
deranged
/dɪˈreɪndʒd/
adj.
43:16
疯狂的，精神错乱的
causally closed
adj. phr.
43:16
因果封闭的
onto something
phr.
43:47
（口语）抓住了要点，发现了重要线索
recapture
/riːˈkæptʃər/
v.
44:20
重新找回
grounded in
phr.
45:01
以……为基础
ephemeral
/ɪˈfemərəl/
adj.
46:17
短暂的，转瞬即逝的
isomorphic
/ˌaɪsəˈmɔːrfɪk/
adj.
46:56
同构的
alluded to
phr.
46:56
暗指，提到
by my lights
idiom
47:21
依我之见
partition
/pɑːrˈtɪʃən/
n.
47:55
划分，分割
computationally intractable
adj. phr.
48:34
计算上不可行的
reductio ad absurdum
n. phr.
49:15
归谬法
bit the bullet
idiom
49:15
咬紧牙关接受（不愉快的结论）
vivacity
/vɪˈvæsəti/
n.
50:20
生动，活泼
Continuum Hypothesis
n. phr.
51:55
连续统假设
corollary
/ˈkɔːrəleri/
n.
55:15
推论，必然结果
self-referential
/ˌself refəˈrenʃəl/
adj.
55:46
自指的
lightness constancy
n. phr.
56:43
明度恒常性
prima facie
/ˌpraɪmə ˈfeɪʃi/
adj./adv.
58:18
初看之下的，表面上
transparent
/trænsˈpærənt/
adj.
58:18
透明的，直接可知的
emergent
/ɪˈmɜːrdʒənt/
adj.
59:42
涌现的
domain general
adj. phr.
60:11
领域通用的
causal import
n. phr.
61:51
因果意义，因果重要性
evaporates
/ɪˈvæpəreɪts/
v.
61:51
消失，蒸发
without recourse to
phr.
62:34
不诉诸于
interactionist
/ˌɪntərˈækʃənɪst/
adj.
63:03
互动论的
intrinsic
/ɪnˈtrɪnzɪk/
adj.
63:58
内在的，固有的
moral status
n. phr.
64:31
道德地位
mundane
/mʌnˈdeɪn/
adj.
65:38
平常的，日常的
staggering
/ˈstæɡərɪŋ/
adj.
65:38
惊人的
coupled to
phr.
66:07
与……耦合
offloaded onto
phr.
66:40
卸载/转移到
ready to hand
phr.
68:07
随手可用的
seamless
/ˈsiːmləs/
adj.
69:51
无缝的
definitive
/dɪˈfɪnətɪv/
adj.
70:24
权威的，决定性的
理解自测 · 11 题 · 是真懂了，还是以为自己懂
1. 查尔默斯如何定义「意识的元问题」？它与「难问题」的关系是什么？
元问题是「关于问题的问题」：解释我们为什么会认为意识存在一个难题，更准确地说，是以主题中立（不提及意识本身）的方式解释我们的「问题直觉」与「问题报告」。难问题关心物理过程为何伴随主观体验，属于难以捉摸的一阶问题；元问题则关心我们说「我有意识」「意识令人困惑」这种行为的来源。在开场至第22段，他指出这些报告是可测量的行为，因此元问题属于「易问题」，可用认知科学标准方法研究，但又与难问题密切相关。
2. 什么是「哲学僵尸」？它在演讲中扮演什么角色？
哲学僵尸是物理上（或功能上）与正常人完全相同、但没有任何主观体验的假想生物，与电影中行为怪异的僵尸不同。查尔默斯在第36–38段把它列为「模态直觉」的例子：许多人觉得这种生物是可以设想的，而这种可设想性正是产生难问题的直觉之一，「我们为什么不是僵尸」就是难问题的一种表述。在结尾（第73段）他又引用尤德科夫斯基的批评，说僵尸论证可能是「哲学中最疯狂的想法」，因为若意识没有因果作用，我们关于意识的判断正确就只能是巧合。
3. Open Philanthropy的研究者做了什么实验？结果如何？
第58–60段：卢克·米尔豪泽等人根据查尔默斯与卡默雷尔关于问题直觉来源的想法，构建了一个小型软件智能体，为它设定关于颜色（红、绿）以及自身颜色体验的公理，再接上一个定理证明器。运行后该智能体得出了「我的颜色体验不同于任何物理状态」之类的结论。查尔默斯承认这个模型「偷了一些工」，远不能说是对人类心智的成熟模拟，但它证明了确实存在一个可操作的研究纲领：寻找问题直觉的算法基础，并为AI研究者提供切入口。
4. 整合信息理论（IIT）受到哪些具体批评？
演讲中有两类批评。第一（第65–67段）来自元问题视角：IIT用Φ度量意识，但托诺尼承认高Φ系统的模拟可以Φ为零却做出完全相同的报告，说明Φ与问题报告可分离，理论需要解释Φ在生成报告的算法中起什么作用。第二（第81–83段）来自问答：Φ的计算需对所有划分求和，超过约15个单元即计算不可行，因此「原则上可检验」难以落实；斯科特·阿伦森还指出大矩阵乘法器Φ值可任意高，按IIT应比人更有意识，托诺尼选择「咬住子弹」接受这一结论。
5. 为什么查尔默斯认为元问题是一个「中立的」研究纲领？这种中立性有什么价值？
因为元问题的目标是解释一种客观行为——我们做出问题报告的倾向——而不预设意识是真实的还是幻觉。第31段和第55–56段强调：同一份关于内省模型的说明，可以用幻觉论语气读（这是错误的自我模型），也可以用实在论语气读（这是正确的模型）。价值在于：持不同形而上学立场的学者可以在同一纲领下合作，先解决可处理的经验问题，再回头讨论哲学后果；同时也为无法预设系统是否有意识的AI研究者提供了可操作的问题。
6. 「现象概念策略」为什么在查尔默斯看来「走得不够远」？
现象概念策略（第47–48段）认为我们有两套独立的概念系统——用于外部世界的物理概念和用于内心的内省概念——它们的独立性使意识「看起来」非物理。查尔默斯在第49段反驳：我们对心灵的许多方面都有内省概念，例如「我相信巴黎是法国首都」也属于自我模型，但信念并不像红色体验那样引发难问题。因此概念独立性不能解释为何只有知觉体验产生解释鸿沟，必须进一步说明感知体验（颜色、味道）的特殊之处，例如它们把可还原属性呈现为原初属性。
7. 查尔默斯提出的「评估意识理论的准则」是什么？推理逻辑如何？
第63–64段的准则：无论什么解释了意识，也应当部分解释我们关于意识的判断与报告。推理链是：如果理论说机制M是意识的基础，而M与我们说「我有意识」的过程毫无关系，那么意识与我们对它的谈论就完全独立，这会非常奇怪——尤其对实在论者来说，我们确信自己有意识的根源理应与意识本身有关。因此可以用「M能否解释问题报告」来检验现有理论，他对IIT、生物学理论、量子理论、全局工作空间理论都提出了这一挑战。
8. 演讲结尾提到「双方各见荒谬」，具体指什么？查尔默斯认为双方各需做什么？
第72–76段：斯特劳森认为幻觉论是思想史上最荒谬的观点，因为它否认我们确知存在的意识；尤德科夫斯基则认为僵尸论证与副现象论最疯狂，因为它要求物理算法与非物理意识「奇迹般地」一致。查尔默斯承认两种批评都有道理，且中间立场往往滑回同样的问题。他要求幻觉论者解释「心灵怎么会完全不像它看起来的样子」并找回数据；实在论者则要解释生成意识直觉的过程如何「本质上扎根于意识」，即便这些过程在概念上可以无意识地运行。
9. 为什么说观众关于「大脑用意识研究意识」的停机问题类比并不必然成立？
第93–96段：观众类比停机问题，认为用意识研究意识可能因自指而受限。查尔默斯回应：尺子不能量自己，但可以用另一把尺子量它；「大脑研究大脑」实际上做得不错。自指悖论要真正成立，需要非常具体的条件，不能泛泛地断言。他还补充：如果意识可以存在于比我们简单的系统中，我们至少能研究这些系统（如狗）的意识，从而绕开自指限制。他也承认侯世达的进路正是以自指悖论为核心，因此不排除某种版本可能成立。
10. 如果有人说「我能不诉诸意识就解释全部行为，所以意识不值得研究」，查尔默斯会如何回应？
第107–113段有直接回应，分两层。第一，他不承诺副现象论：泛心论、互动论、还原论都为意识保留了某种因果角色，元问题研究或许能帮我们弄清这些角色。第二，即使意识对解释行为无用，也不等于它不重要——在他看来意识是「唯一真正重要的东西」，是生命意义与价值的来源；并且AI或动物是否有意识决定其道德地位、能否受苦，这关系到我们如何构建社会。所以意识研究的价值在于连接最根本的价值，而非工程效用。提问者也承认这一论证对旁人未必有说服力。
11. 把「人工意识测试」的思路应用到今天的大语言模型上，会遇到什么困难？查尔默斯的立场提供了什么线索？
第61–62段的测试主张：若机器自发地对意识感到困惑，可能是有意识的标志。但查尔默斯已预见反驳：机器可能只是「背熟了人类说话方式」。大语言模型正是在海量人类意识话语上训练的，因此它说「我感觉自己不止是电路」几乎无法区分真实困惑与模仿，这使测试的证据力大打折扣（施奈德的ACT测试因此要求隔离相关语料）。他的线索有两条：第80段表示若AI拥有与我们「严肃版本」相同的意识模型，他会视为很强的证据；第105–106段又提醒智能与意识可分离，语言能力高不等于有意识，意识也可能存在于无法表达的系统中。
下载便签
手机：长按图片保存
← 上一期 · REC_007How to Speak
下一期 · REC_009 →"What Is a Strange Loop and What is it Like To Be One?" by Douglas Hofstadter (2013)
苏菲拉底
THE SOPHIE LAB · ASK THE BEST MINDS THE BIG QUESTIONS
内容仅供学习 · thesophielab.com