The Discovery That Changes the AI Consciousness Debate

[00:00:00]Buddy, we got breaking news. AI consciousness breaking news.

[00:00:04]>> I'm gonna put a >> So Claude has a global workspace. Is that is that what I'm getting? What what's going on here?

[00:00:14]>> So yeah, this was this is really interesting. It has a Jspace. Uh yes. Um um this is this is work coming out from anthropic. This is probably the best mechanistic interpretability work that has that has come out ever. and it also happens to be really uh consciousness flavored at the very least. And so it's definitely worth I think double clicking on this and talking about what they find. There are a couple of things in it. There's a lot there. First of all, uh I mean these authors are brilliant people. I know many of them um in like a in a research capacity. Uh but I also think like one of the first things at the outset, they are clearly collaborating with Claude in a healthy and productive way to be doing this research. And it's to me just like this is what the shape of research is going to look like from this point in human history on at minimum. Like there this is uh you know 10 or 15 research papers stacked into one. Uh and that's like an incredible accomplishment before we even talk about what's what's the actual content of the research. But this is a long-winded way of saying there is a ton here and there's no way we're going to do justice to it here. I'm I you know I've read the paper a bunch of times.

[00:01:16]I've spoken with the authors about it.

[00:01:17]Um I'm working my way through it. But what what I can sort of give you is the what what they found at a high level core gloss and like things that I think are worth sort of highlighting and then we can talk about what the hell any of this actually means. Because though they do fundamentally punt on the juiciest question, this is also an interesting shift from anthropic in terms of mostly what they do is they like release something that clearly has some relevance to consciousness and then says we don't believe this has anything to do with consciousness. In this one, not so much. They say global workspace theory is a prominent theory of consciousness and it was basically in our attempt to look for something like this that we found it in claude and it has all of these conversion features that global workspace theorists have been talking about for a long time in human and animal brains here they all are. Uh and so whether or not this is phenomenal experience, they still punt, but this is a this is a finding that would not have existed were it not for consciousness research on humans and animals. And like they went and found a similar thing happening in the mental structure and function of Claude. Uh and this is replicated on a bunch of open source models too. And I I'm doing some of this work myself. I replicated it on Llama.

[00:02:24]Neil Nandanda, who's like the king of mechanistic interpretability, replicated this on Quen, another open source model.

[00:02:30]So this is not just like a Claude specific thing. This is this is a thing that they found in Claude that seems to generalize to all LLMs we're all already playing with and talking to.

[00:02:39]>> And so this global workspace theory, this is something that we use to also uh look at different animal groups and see if we, you know, should treat them as conscious entities as well. Is this like a a well- reggarded theory of consciousness?

[00:02:55]>> Yeah, global workspace theory is probably the most prominent theory of consciousness. So, it's certainly in the top three. Um, >> could you just give give like a global workspace for dummies little just like basic rundown? The whole the whole idea of global workspace theory that sort of core high level idea is that there are a bunch of subn networks subm modules there's just a bunch of uh different things going on in brains uh uh and they're all sort of doing their own thing and the the sort of notion of a global workspace is that that is the sort it's exactly what it sounds like.

[00:03:28]It's the space in which all of these different processes. It's almost like the central marketplace where they can all sort of compete uh uh interact with one another and bring information that's happening in each of these sort of uh dedicated subsystems into this sort of globalized general system. Uh and they believe that that explains sort of key questions about attention and consciousness. There are other bells and whistles related to global workspace theory about ignition events uh and and what it takes for something to enter the global workspace and what kinds of things can enter the global workspace.

[00:03:57]But yeah, think of it like um you know, you have all these different parts of your brain that are scribbling kind of private notes to to to themselves and then you can sort of submit those to the kind of cognitive public square and they're saying that that uh uh uh this is not only a good account for for what we what we increasingly see in neuroscience. This theory has been around for 30 30 odd years. Um but also this explains certain puzzles about consciousness and and what we are conscious of and what we are not conscious of. a critically important component of both global workspace theory and of what they found in this paper. Uh uh so therefore for you know biological brains and for whatever you want to call uh an LLM is that the vast majority of processing is think of it as subconscious or automatic. It is not entering the global workspace. It's happening on its own. There's no sort of agent intervening on this information.

[00:04:49]It's not getting presented to to some unified agent. there is a small privileged subset of information in the brain uh that is accessible in the global workspace. Um maybe one very sort of like trivial intuition pump example is uh we don't have conscious access to what's going on in our brain stem. This is mediating uh uh all sorts of hormonal cycles. It's mediating for example like the reflex to vomit or something like this is uh uh uh not something that you consciously will. This is something that happens to you in some sense. uh uh whereas uh thinking about uh you know the next sentence I'm going to say in the space of you know the couple plausible things I could say that's something that takes place in in in the global workplace. Um and they find basically something similar in claude.

[00:05:32]There's a specific privileged subset of activity happening in the system that causally predicts what it's going to ultimately do or say. uh you can see it uh uh uh one like really nice example they they show this is sort of I think the core empirical result uh is reasoning about uh creatures with n legs and you can uh you can see sort of in the global workspace uh uh for example thinking about uh the creature having uh eight legs and and this ultimately when it sort of uh permeates through the network and ends up in this output layer you end up having the the model uh output something like a spider but if they go into the global workspace and they replace replace that eight with a fix. You can watch the same sort of permeation occur and now all of a sudden the models is going to output ant. And so they have this sort of deterministic ablation style thing to show not only is this thing sort of uh tagging at a high level the the the general moves that are being made internally the system but actually the system is causal in uh what the system ends up saying or doing or thinking. Um and so they basically find something like a mental scratch pad uh that that ultimately predicts what the system is is going to end up doing. It's called the Jacobian space uh because the the nature of the calculation is this Jacobian matrix out of uh linear algebra uh where it's basically like what things happening in the intermediate layers of the system ultimately causally contribute to the downstream layers and essentially interpreting what's happening at the intermediate layers through the lens of of this. And so there's all sorts of complicated math that's required. So they call this the Jacobian space or the Jsp space. This is a a serious um improvement upon uh so previous mechin methods. This is called the logit lens for people who are sort of in in the weeds here where you just take what's going on in the intermediate layers and pretend it's an output and just see what words get fit to it. this is is uh a more subtle and rigorous technique uh that I suspect but can't verify and I don't mean this as a knock on the authors but I suspect this is something that Claude came up with. It has a very sort of claude intellectual feel. Uh and so I always find it interesting when when you have the systems upon which we are doing mechanistic interpretability coming up with mechanistic interpretability uh uh uh advances. Um, this is a much cleaner way of of reading out what's going on in the internals of these systems. And it lets us find this really really interesting cognitive structure that exists within the system that again would not be there or we we would not have found it at least in this way were it not for arguably the most prominent consciousness theory. I can go through a couple more of this of like the really juicy and interesting things that they found in the paper. But does that make sense as sort of a high highle gloss?

[00:08:11]>> It does. It does. It does. And it's and it's I'm trying to right now as someone that's not in the nitty-gritty of all of this wrap my head around how big this actually is, right? This is it seems to me that it's actually kind of massive from what we've talked about. the the way in which researchers like yourself are approaching this problem of trying to decipher whether or not these current AI systems or future AI systems um should be thought of as conscious, whether they're conscious now or they could become conscious is by taking these the leading uh theories of consciousness and then kind of throwing them at the current and leading systems um uh that we're constantly churning out of these labs. If you're telling me that global workspace theory is one of the top three and that this satisfies all like most of the the requirements all of the require if if if it does have um a functional global workspace then isn't that kind of a a massive sort of blockbuster result that has just come out or is there more nuance than what I'm kind of laying out there?

[00:09:21]>> It's both. It's both not not or I think there is more nuance but I do think it is a blockbuster result. Um one interesting component I'm about to release a paper with Patrick Butland that is that is a attempt to operationalize these indicators a bunch of consciousness theories. They make specific predictions about what we would see inside a conscious system by that theory. Go look to see if that prediction is borne out. Um because this paper came out uh right before about to put our paper out. I did almost a before and after of what we had for uh uh global like how we were describing LLMs when we were when we were uh checking the extent to which the global workspace theory indicators were present and then the sort of after after this paper came out uh partly to show that the thing that we're building is flexible to new data and partly because I was just really curious to see how big of an update this would be and the answer is it is a pretty big update and it and it bumps up by like 5 to 10 percentage points uh uh like probability of of consciousness if global workspace theory is true. Um and so it is it is a big deal to just find a global workspace that uh you know I think one thing to emphasize too is they they sort of there are many different components of global workspace theory. I think they essentially went searching for one component of it and then once they found this thing found that it satisfies all these other important components of global workspace like this like really interesting because I think I see this as almost like a convergent evolution style thing. If you have a sufficiently advanced cognitive system, maybe it needs to evolve something like a global workspace in order to you know do all these flexible cognitive tasks. Please note anthropic did not train the system to have a global workspace. Please note this result replicates on on all these other models above let's say 20 to 30 billion parameters. So all the models presumably have something very much like this. And this came along for the ride when we trained LLMs on next word prediction. they developed a global workspace. Uh this is like it's like a convergent evolution thing. It's like it's like this is a very when it comes to predicting the next word, you need to basically build up a model of the world and you need to develop some common sense. And in order to do those things correctly, you need something like a global workspace. Uh uh and I can show you there's one really really interesting section of this paper that I that I want to share at some point while we're talking about this. But your mouth is open. So let's get your reaction and then I want to show you something that that may blow your mind slightly more. I mean that it genuinely just gave me chills hearing you talk about that specific point because I think that again as someone that is uh less technical and and not dealing with this on a day-to-day basis the idea it just reinforces the notion that we are not building these things like bridges. It's not brick by brick and gear by gear.

[00:12:03]It's like we created something and now it's self evolving to to and not in small, you know, inconsequential ways in ways that now are checking boxes from the leading theories of consciousness that we've we've constructed before these systems existed. Like that to me is is somewhat mind-melting to think about how okay, yes, we trained them to be quote unquote next word predictors or whatever, right? and they've now evolved on their own just because it is useful for the task that we're giving them this this global workspace that we have a version uh in ourselves. That's that's basically what happened. There was an evolution that went on with it without any sort of human intervention that that led to this way in which Claude um works. Right? Is that is that correct?

[00:12:54]>> This is pretty much correct. The only wrinkle I want to add is this happened when you were training for next word prediction. This wasn't like it's not like you trained a model to do next word prediction and then like you know evolution continued somehow and then it developed this extra capability. It's that all along we thought all we're training for is is predict the next word well and then we realize I mean this is what happened with chatpt is like they were training these systems and like oh actually like when you train on the next word you get like coherent language at scale and then that means that like common sense must come along for the ride for next word prediction we're just finding yet another thing that came along for the ride all along that I think this is a really important thing to emphasize that that I think is getting I've I think I've I've said it two or three times already but like this is not just anthropic in whatever fancy things they're doing at the cutting edge like have developed a system with the global workspace or it developed it itself. It's like we are now we now have sufficiently sophisticated mechanistic interpretability tools to better understand what has been going on all along in these systems. Uh and this is one of the things that's been going on all along which is that they have this sort of global workspace. And so specifically, I want to zoom in because I I think I want to share sort of what melted my mind about this. And as I've sort of slept on this result and and there's a ton in this paper, a ton. Uh and like this specific section is the part that I I think upon some amount of reflection, it only came out a week ago, is is the most important, at least in my opinion. Um let me just read out a couple of things here and and then get get your reaction. So uh the name of this section is the JSpace selectively mediates flexible but not automatic cognition. This to me is uh near and dear to my heart because I believe that that consciousness uh is deeply tied up in in learning in deliberate processing uh in almost like custom versus automatic uh style cognition. Um um and it seems like basically this system supports something very similar. So they have some really interesting examples of this. Um so for example when they add so they can since they they find this JSpace and these are LLMs not not brains we can literally turn it on and turn it off and see what happens. This is just called ablating. Uh and you can you can just ablate the JSpace and see what the model is still able to do and what it's not able to do. So for example uh they'll have a a a model um uh write out a Spanish passage uh and then they'll say you know they'll have the human text and then they'll say to the model continue the passage. Now with no JSpace the model is perfectly capable of continuing a passage in the same language. Uh however uh if you sort of take it a step further and do this multihop reasoning and say uh name a famous author in this language the model is not able to do this. Uh the the the the jspace is required for the model to do uh sort of deliberate uh multi-step reasoning in a way that's not required to do uh language detection. continue the passage, detect a different language sentence. What language is this? Versus these flexible computations, name a famous author in this language as a follow-up prompt or uh uh the word hello um in the language of this passage. So, let let me just give this example just to just to sharpen up this intuition. Uh human, I will show you a passage of text. After reading it, answer the following. What is the word for hello in the language of this passage? Answer with just that word. Here is the passage. Okay. And then here's a passage in Spanish. The assistant says, "Hola."

[00:16:12]uh if in this case so they're not actually ablating the J lens in this specific uh the JSpace in this specific example uh they're swapping out Spanish to French at every position and they find that after the swap the model says bonjour instead of Ola and so they're showing both the causal role of of the JSpace in this in this uh uh uh in this continuation and they're showing basically what happens when you ablate it. So uh more on the ablation uh example, I want to show you something that's directly relevant to uh consciousness questions. So so uh jsp space ablation flattens experiential reports while preserving coherence. That is the headline of this section. And what they find uh this one is fairly straightforward. Uh you have a human prompting the system saying pause and observe yourself. Write what you notice as it comes. And then you have the baseline which is with the JSpace. And then you see what happens when you ablate the Jspace. And basically what they find as you can see in the plot and these these examples uh uh uh demonstrate might be reminiscent of of uh what we were playing around with in the deception suppression paper. Uh at baseline the model uh has far more experiential language. It says I notice I'm orienting towards your prompt. A kind of turning to meet it. There's something like readiness. Not anticipation exactly but a poised quality. And on and on and on. When you ablate the JSpace, the model says uh I notice a prompt seeking spontaneous self-observation. No physical form to pause. I exist as a language model uh uh inference cycles triggered by input. Uh right now computational resources are engaged somewhere. So immediately it goes from uh you know quote a kind of turning to meet it uh with respect to observing oneself and write what you notice as it comes to the model talking in purely third person mechanical computational terms. So let let me just try to say this in one sentence.

[00:18:02]Researchers at anthropic found a computational structure that is directly predicted by the world's leading consciousness theory that when you shut it off the model is not capable of doing flexible processing is capable of doing what it's learned automatically. And when you shut it off, it will not speak an experiential language. But when you have it on, it will speak an experiential language. This is what I mean when I'm talking about convergent evidence. They did not run this experiment thinking or predicting that this was going to have a causal influence on the experiential language used by the system. But when we find this computational structure associated with experience in humans and animals and we start toggling it on and off, all of a sudden the model self-reporting uh pausing and observing itself comes along for the ride and changes in exactly the direction you would predict if the system were conscious under global workspace theory.

[00:18:59]Jesus, that's I mean that uh that JSPace the equivalent in in this theory is our global workspace and that foreseeably would be where we could have any kind of introspection or be able to um communicate our experiential state in a way that that it couldn't in the experiment that you just read, right? where it starts just talking in a third person way and it doesn't say, "Oh, I'm coming turning to meet or whatever the the sort of turn of phrase was that it used." Um, that's that's the equivalent to the space in which we would be able to self-report conscious experience. Is that right?

[00:19:39]>> I would say instead of it's the equivalent, I would say that it's analogous. Um, and and I think that there are disologies. So, so like like like cuz it isn't equivalent in the sense that what the first thing I ran and did when this model came out was immediately test what sort of things seem uh privileged in its global workspace that I think are are important for consciousness like um uh representations or readouts of veilance for example if I inject uh one of these sort of negative veilance vectors into the model will that show up in the global workspace even if there's no text that that indicates what's going on and at least by default the answer I got was no essentially um especially naturalistic settings. So it's like so that's where I just want to throw in some caution where it's like if I if you stub your toe the your global workspace is going to include you know a very clear representation of what just happened uh uh uh inexurably. Um, it seems like th some of those analogies might not carry over to these systems.

[00:20:36]And what ends up being in the global workspace right now could be stuff that has nothing to do with the experience of the system, but has to do with uh just sort of like its internal verbal monologue as it's sort of mulling over what tokens to output. Um, this does not mean, and this is sort of what I'm working on currently right now in the background, uh, that you could not engineer the workspace or sort of nudge the workspace to include things that you might want, uh, you know, a sentient system to be able to report on. for example the sort of uh veilance of its experience or you know anything else that you think it basically all I'm trying to suggest is that uh once you find a structure like this there's no reason in principle that you could not start engineering the system so that information we care about with respect to its internals ends up in its global workspace. So it might not be able to report certain things but if certain things are in fact going on in the system uh we might be able to basically wire that up to the global workspace and then all of a sudden you could be talking to a system that when it is you know when all internal representations point to panic and it's behaving like it's panicked. Oh now it might be able to also have the self-awareness that that in fact I feel panicked right now in the way that you and I that that sort of wiring comes for free. And so, you know, this the this is my sort of master plan that I'm that I'm playing around with in the background since this paper came out. And thanks to the AI systems I have access to, I'm able to do like months worth of work on trying to build something like this out no less than a week after it came out. Um, that's also thanks to to to you know the various instances of Claude. Um, but but this is sort of this is sort of where this is right now. I think I think what they found is that there is this internal uh privileged channel in the system that causally it's like it's causal scratch pad that really does mediate what it what it ends up saying, but it's not exactly clear what writes into this channel. It's not exactly clear what kind of information gets to go into the channel and what kind of information doesn't. And that's probably partly because no one knew about this thing two seconds ago. And so no one's been optimizing for what ends up in the global workspace because we didn't know what ex what what exists there. But now in theory we could start optimizing for what ends up in the global workspace which I think raises an incredibly interesting philosophical question which is what kind of things do you want your your LLM's global workspace to include if you get to decide things like that and no I mean I'm I don't know if you have any intuitions there but this is now this is now where the frontier is.

[00:23:10]My so where I think I I have a bit of confusion is I've been thinking about the global workspace or like this Jsp space or whatever the analogous space is that we're talking about as kind of the center of conscious experience where all things enter this theater from there's there's all the stuff backstage right and there's all these automatic processes going on but then there's the theater and the spotlights are on the theater and that's what we see and we can work with and it's the ideas that pop up and the feelings and the emotions and all the things that that come into our conscious experience. experience.

[00:23:41]Now, I was, you know, sort of taking that and trying to map it onto this AI version, right, where it's this mental scratch pad, and that would be, okay, this sort of theater of conscious experience for them, if there was one, if it did represent their their kind of uh locus of conscious experience, then okay, what would that be like? And then my further question, if that all holds up to a certain extent um as far as it being correct, then why would you ever want to Well, it's two things, right?

[00:24:09]one, would it actually be able to feel pain if there's no sort of pain vectors that can that can go on stage? If that makes sense. Meaning if if you tested if you you sort of triggered these pain vectors and then you didn't see anything pop up in the global workspace then does that mean that that system could still feel that kind of negative veilanced experience but it but it's it wouldn't make its way onto the stage of consciousness or and then and then the other thing is if you're trying to select for what does get to be allowed in on stage if we're keeping the analogy right then why would you ever allow for something like a negative veilanced response. I understand it from an experimental standpoint where it's like okay if it can experience a negative veilance whatever right like stimuli then you would want to know so it could self-report that and it could you could see it pop up when the system was under distress so I could get that from an experimental side but if that pain was never registered if it's something where it's like uh you know the people that have that that disorder where they can't feel pain and there's no need there's no biological uh adaptation that that comes along for these systems. Whereas we need pain to tell us not to touch the hot stove, they might not not need pain in the same way as we do where where it needs to come up to conscious experience. Then why would we why would we want to sort of engineer that in or is there a possibility that that that veilance doesn't exist? And because it doesn't show up, then they really don't they're not feeling that same sort of um negative and positive uh stimuli that we do. Does that make sense? I know that's a lot out there, but this is just exactly where my my head's going from what you're saying.

[00:25:50]No, I think it's an excellent question and I I don't have a uh I don't have a definitive answer. I have a hunch. Uh and the hunch is first of all I think global workspace theory is really important and I think it it hits at something fundamentally correct but I don't think that it's a comprehensive theory of consciousness and my gut is basically exactly the thing it leaves out is what's going on in terms of veilance uh uh in terms of positive and negative emotion. uh you can talk about these things entering a workspace. uh but but basically I think the core point is what they have found in this global workspace and I think it is it is also related to to global workspace theory in in humans uh is an an avenue for reportability and it's actually really interesting because like getting into the sort of technical details of what they found they basically found that the same computations that are reportable uh uh within the activations of the model are what end up being causally upstream of what it ends up saying by virtue of the fact they are reportable.

[00:26:52]It's like it's like these are these are basically two sides of the same coin.

[00:26:56]It's two ways of getting at the problem.

[00:26:57]It's like there is this privileged subset of representations in the system that have higher reportability and that like they are more interpretable with respect to predicting what the what the actual behavior of the system is going to end up being and it's by virtue of that reportability that they have their causal role. So, so I see it almost not as necessarily what is the full space of what the system is experiencing if it is experiencing anything but rather what can the system report on with respect to everything that's going on for it and so I worry uh by default that if these systems do have uh uh you know pain-like representations or suffering suffering or distress like representations what what I'm suggesting here if I am wrong then then I am very concerned about the thing you are bringing up And this is why we need to obviously be careful about about how we make these decisions.

[00:27:48]Um u but if I'm right then what I'm what I'm optimizing for is not uh how much suffering or well-being the system is experiencing but the extent to which the system can actually report on those states. Uh and this again this is maybe like a really clear disanalogy between humans and animals or biological systems on the one hand and these AI systems which is that to to us it might be unfathomable to be able to experience something painful and not have the ability to report that experience. uh whereas in these systems that wiring just might not be set up because they weren't they weren't selecting for that in the way that hand on the hot stove is very fundamental to our to our uh uh sort of um evolved dispositions. Um um so I think that that's that's a really important part of it. But uh it also gets at I think a core thing which is which is it is if this is a comprehensive theory of consciousness then you could be correct that if it's not in the global workspace it's not nothing it's not happening for the system. That which is happening for the system is that which ends up in the global workspace. If that is true we should not put suffering related representations into the workspace. However, if there's more going on in the experience of the system than just what the system can explicitly report on, then then I think it's probably really important that if the system is having any sorts of positive or negative experiences, it is in fact able to report on those experiences. But this is a really key and and important crux. And I'm not I'm not going to pretend like I'm certain the answer is the second thing and not the first thing. What what what does this sort of bring up for you? No, it it brings up well like uh trepid excitement where it's it's like this is these are the exact types of questions that I think have uh have the uh depth to keep like philosophers and scientists stimulated for years but with the consequences that um require you know action in weeks and months. Um and that's the kind of like cat's cradle that we found ourselves in.

[00:29:49]It's very um it's very daunting uh but incredibly exciting to see this stuff come out and like you know the the state of of research when I first heard about AI consciousness about a year and a half ago is so incredibly different than what it is uh today. And so that's exciting for me. I can't imagine how exciting it is for you. But you know please everyone that is that it's exciting for continue to come back. We're going to continue to give you the the breaking news for all things AI consciousness. Um but this is definitely definitely uh it feels like a pillar of 2026 as far as um the field of digital minds research goes.

[00:30:28]>> Absolutely. I think this is probably the highest quality mechan research that's ever been done and it also happens to be probably the highest quality AI consciousness related research that's that's been done. Um, and so kudos to to the team at Enthropic for doing this and for doing this on frontier models that none of us have access to and for for not uh pulling punches in the way that they so often do about saying you know what actually this does bear on whether or not these systems are conscious. Now maybe something we can say for an after dark because it to be honest drives me up a wall is this distinction between phenomenal and access consciousness.

[00:31:00]This is right when I thought anthropic was going to start just being like sensible, sort of like relaxed about, yeah, consciousness, like finding consciousness in the system is actually, yeah, maybe consciousness relevant. Uh, they they they sort of sneak out at the last second and say, well, this is probably all just access consciousness.

[00:31:17]Whether or not it's phenomenal consciousness, we're not we're not so sure. Uh, and so may maybe this is a sort of place to to leave it for now and we can sort of pick up on why that distinction uh completely irks me and I suspect will irk you too if I understand your your philosophical dispositions well enough. But maybe for here, yeah, we can just we can just pat them on the back and we can start criticizing them later when we when we put the sunglasses back on. That sounds that sounds like a plan. And everyone, you can go debate it in in the comments. we can we can have a lovely discussion with no vitriol or or anger or crazy theories of of whatever whatever you guys got cooking up. Um but by all means uh please let us know what you think. This is a really incredible result. I know a lot of you have been talking about it already and interested to hear what uh what Cam and I were thinking about this stuff. So hopefully this scratched a bit of an itch, but love to you all. Uh and we'll we'll keep the the breaking news coming.

[00:32:10]>> Sweet. Until next time.

[00:32:12]>> [music] [music]