Anthropic's New J-Space: Debunked?

[00:00:00]Hello com so great that you are back.

[00:00:03]Today we have a brand new scientific paper by entropic and they tell us something about a chase space. So we will have a look at the JS space just yet you get an information we will talk about transformer cause well mapping we will have a look at from activation to here an unmbedding matrix then we will look at the JS space itself of course it is here within the residual stream and we can see it here as an activation space and then I will show you why the result by entropic is not dil and let's start so this is the paper here entropic beautiful or test congratulation published July 16 2026 Six pantropic and look at the title. The title is strange.

[00:00:44]Do you notice it? Verbalize representations from a global workspace in language model. It is not a mathematical representation here from a global workspace within the transformer layers in our LLMs. Why do we have to have a verbalization of this mathematical representation inside a transformer?

[00:01:06]Now you know we have no problem with this. know we have a residual stream and in this video where we fine-tune supervised finetuner Laura here L&Ms we had the evolution of activation as a particular way to fine-tune something here this thing this was within the last week so we know how to deal with this no or here we know the temporal dynamics in the residual stream of a transformer and in this particular video we started here with the basic mathematics here of the core residual stream update functions so we know how to do this we know how computers so this is a familiar topic here. But of course, you know that also one of my last videos was here. Wait, I'm aware. So, selfawareness detected.

[00:01:49]And of course we were talking about here an introspective analysis and I showed you how you can train a model here or the authors published here a paper how to train your model that you can say hey I just detected somebody or something just in injected here an alien sword into me into my memory structure into my residual stream and I know exactly it is located after sentence 7. So we know this and now we go the next step and the next step is exactly here what entropic wants us to believe. So let's understand this now. Read the paper. I think it's a masterclass in uh presentation of a scientific fact but also in priming because it is about a functional analog in the communication patterns or is it not?

[00:02:41]Am I just read about this? No. In the abstract we start with a human brain process. Now out of everything here the human brain process is only a small fraction is yes yes yes and then we present now evidence entropic presents evidence that an analogous functional distinction has emerged in LLMs. So claude is now almost human. We have now evidence that a functional distinction that is analog to the human brain processes is happening in claude. So AI is just like us human. AI is just beautiful and man imagine what it means for the upcoming upcoming IPO for entropic. Unbelievable if they can prove that AI is similar to human. I mean believe this. So here we have a beautiful topic for marketing clawed in the human brain. How different are they? And you immediately understand that the topic of today's video shifted here from a pure scientific technical analysis what entropic presents in the way entropic presents us the paper. We see that there's a clear intention and it is timed for their financial offering.

[00:03:54]So let's have a first look at the introduction. Now again we start with the human swords. No our human sword. If the mind is an ocean, we spend our lives floating on the surface. Beneath us, an enormous amount of processing takes place without our knowledge. No. And yeah, human neuroscience beautiful. So we start with human. Everything is familiar. And then we say in this paper we present evidence that an analogous functional distinction has emerged in modern AI models, our claudi. Now what a coincidence. Now imagine this here as a headline absolutely gorgeous.

[00:04:35]So let's look a little bit closer and let's start here. Maybe now let's be clear they have access to their clo here in entropic that I have no access. So they are absolute specialist. They are the gods. They know everything here about the internal reasoning process because they cannot open up the black box of claude. No I can't. I just can read the paper and have a little bit of fun maybe together with you and we look at your argumentation and you decide at the end of this video is this a serious scientific paper or is this just kind of a marketing paper that is intended here to provide some marketing insights. So let's start all this humor talk. We have to have here a clear theory in neuroscience. No and we go here entropic goals with a global workspace theory.

[00:05:21]Absolutely. No, this is a prominent cognitive architecture and a theory of consciousness. What a coincidence. First proposed here in the 1980s by Bernard Bars. It models the brain as a collection of unconscious specialized modules operating in parallel. Now what a coincidence chose this particular global workspace theory. And if you want to see here some more up-to-date papers here, I found this interesting global workspace and prefrontal cortex the recent developments of course here. This is a paper here that is open here by NIH. You can go there and read it yourself. It's rather interesting to get give you a feeling. Yeah. So GWT suggested the human brain does lots of unconscious processing. I believe this but routes here the important data to a central internal workspace. No. So there we have a central specific space here where this is here the intelligent core of whatever the brain or the eye system.

[00:06:19]you know and from [snorts] this workspace from this core it's broadcasting into the rest of the brain.

[00:06:25]So this is here the central place of intelligence reasoning whatever you like to call it. Now entropic in its paper now maps here this cognitive science concept here of the global workspace theory onto its own clawed large language models. And I had a question why why are they doing this? Why they build up the paper? Why is the priming here of this storytelling in this paper in this particular way? And the answer is simple because entropic needs a special space. It needs the chase space in the transformer because we have to have something special if we ask for I don't know a trillion dollar validation.

[00:07:06]Yeah. So we have here some official wording here like hey out of everything the human brain processes here only a small fraction is continuous accessible and we present evidence here tells us entropic that an analogous functional distinction has emerged in large language. Well, and you say great. But we between you and me, we too, we know exactly we must immediately cut off the word conscious from feelings because this is something of a phenomenal consciousness and entropics preprint is solely investigating some computational routing. Now they are brilliant in their priming of the wording because they call the computational routing now access the consciousness. I mean beautiful wording here for all their marketing. No wall street headline. Yeah. Yeah. Access consciousness. No, it is just a computational routing between you and me. So what is the real mathematical reality? No, we know a transformer transform model massive series of matrix multiplication residual stream attention MLPS. Beautiful. Now in Tropic's basic assumption if you condense it to the mathematical core is simply this question. Is there a specific mathematical subspace? And I will show you it's not a subspace but never mind we go with entropic there in computer science. Is there a specific mathematical subspace inside the residual stream of our transformer where the intermediate data pulls together before becoming an output word a logic.

[00:08:31]this is it and they answer yes we found indications H and I ask myself hey wait a minute so is this JSPace so this JSPace is in the residual data stream of the transform okay and it is just a specific selection of orthogonal or oronormal subspaces and reading the paper I have now for me personally and I might be completely wrong the following answers and I share my answers with you I think it is defined inside the residual stream vector space. Yes, of course, it's a mathematical artificial vector space at each transformer layer.

[00:09:07]But it is not a subspace. It is a convex cone. If you go into mathematics here for some yeah non- negative coefficient, forget about it. So a sub the jsp space is better described not as a subspace but as a union of many sparse non- negative cones constructed from an overcomplete dictionary of non-ordonal token directories. So it is not one fixed linear subspace but yeah hey entropic is is computer science and not mathematics. So jspace is not a special or compartment inside the residual stream. It is a sparse vocabulary defined coordinate family imposed on and sometimes used by the residual stream representation itself. And you see this here I just asked here my chat GPT to bring up this image. And yeah I hope I made it clear. So we have here the activation space and we have some non- negative combination and you see the jsp space is a coordinate family not a physical module because I already read some interpretation and on Twitter and I got some even some questions here on email here is this now a new mathematical space that was discovered by tropic is jsp space now something new no not at all so union of sparse cones not one fixed by some low dimensional subspace Wait. So I built here this yesterday. So [laughter] I just want to show you, hey, we have a complete mathematical understanding of the inside of a transformer. No, and you too. We have here at the bottom our our input tokens here. Then we have here the embeddings. Then we have the positional encoding rope and all the 12 different rope things. And then we have here our transformer block. And we have the classical heads, the multi head attention layer norm. And then the MLP, the feed forward layers here. And then again the layer norm. And we go here from one layer here to multiple layers, multiple blocks. And then in the end we have our final layer norm. We have our logit our output token. And then for particular words like there or a or cat, we get a probability that tells us ah this is here the next token prediction.

[00:11:11]So we understand the machine and now I ask myself, hey, where is now this new JSpace? Show it to me. Point your finger to the exact space. Where is Jspace?

[00:11:24]And I thought about it and I said okay I mean clear you cannot read the neural network sorts which of course the internal data streams directly but you see I almost followed the entropics priming yeah the AI machine is dreaming and sinking and it is having feelings no it is an internal data stream and it's just a wall of numbers and the researchers here created your mathematical translator they called your jacobian lens.

[00:11:52]Now what is this? I mean here you have again this beautiful Bontropic I made a screenshot here. They make your practical window into a model's unspoken thinking.

[00:12:03]This this subconscious priming you know the AI machine is now sinking and it has an unspoken thinking space. I love it.

[00:12:12]But the reality is different. Huh? J lens identifies a representation of a model is poised to verbalize at any point in its processing revealing unspoken thinking. And you might say, "What the hell is an unspoken sinking?"

[00:12:25]Yeah. Now, just to be clear, JSPace is not a discovered physical module. It is not a component and dedicated set of neurons or sharply isolated fraction of the network. It is a representational construction based on token specific direction. Remember we are operating in a vector space in I don't know 4,000 dimensional vector space. So we do have specific direction we have token specific directions that are derived now from hold on to your socks averaged Jacobian matrix elements.

[00:12:57]Now I know if you're new to AI welcome.

[00:13:00]What is a Jacobian matrix? It is here a matrix of first order derivatives. So it describes here how and this is important as small a infinitismal change in a multivariable input changes here the multivariable output parameter so this is something in physics we operate with here when we have our differential beautiful and we use this here now in AI no problem at all the jacobian here of classical form your row I is here the ingredient of the output f of i with respect to all input coordinates and a column j describes how changing the input coord that it's XJ effects now all the outputs. So you can take now the Jacobian like in theoretical physics as a linear approximation for underline sufficiently small perturbation delta x we get this formula and therefore our delta f is exactly now the jacobian here that operates on the function. So our J lens score if you have a high JL score does not prove at all the following not that the model contains a discrete concept identical to the token label because some of the secondary publication claim this this is not correct this is not true the representation is used in the in the current computation. No, the representation has a single meaning here. The model will eventually output a token probably but maybe not at all. The direction is causally sufficient in every context. No. And the model is conscious of the information.

[00:14:30]This is absolute nonsense. This is BS.

[00:14:34]Remember the Jacobian also provides here or only provides the first order information. This is the term that we get now here if we add this. So for finite intervention and let's say our EI and our real world experience when we have here our English language we have finite intervention the ignored higher order terms may matter but we just ignore them because the methodology that entropic uses here is just look at the first order okay this is a first paper this is here the first approach the first step but just notice a larger steering intervention may leave the region in which the linear approximation is accurate So we just leave out maybe the most important part. Now they design now something like a Jacobian lens or a J lens. So what is it? It is simple. At a model layer L, it simply estimates here this mathematical expression. What is it? So at first we have our delta h final t dash and this is the later activation from which the output token probabilities are calculated in relation to our delta hlt. This is the mall's residual stream activation at a particular layer L at a specific token position T. And then what we get the lens is simply what what it does what is the the function of the lens. The lens estimates how a small change a small minuscule change in the earlier activations of our transform architecture would on average because yeah we have to average over the Jacobian change later output activation.

[00:16:10]This is it. So we go in, we change a little bit the activation and then we just want to have an estimation over an average Jacobian set on average how this would change the later output activations.

[00:16:24]And you might say what? This is it. Yes, I read it twice. This is it. There's nothing else.

[00:16:31]And then the mapped activation are passed through the mouse output vocabulary matrix beautifully. So they assign vocabulary labels. This is what we call words in English like spider or Spanish or panic or failure to direction in this new mathematical space with corresponding effects on later token logits.

[00:16:51]Now they present this as something special but I personally fail to see this as something special because yes of course we build it in that way. We use the Jacobian exactly for this particular purpose.

[00:17:05]Because the lens, the J lens is designed to identify the activation direction that can affect the verbal output. Now, if we modify this activation direction, imagine we have a little knob somewhere here in the transformer layer and we turn the knob either a little bit shift to the right or to the left side. Yes, of course, we want this effect to happen. So, to be surprised that this is happening, I mean, we build it particular in this way. So, hm question mark. So another one and there are multiple beautiful examples and I just put up I just show you some of them. Now please read the paper yourself get yourself an impression I just want to give you some hints but it's a very nice paper by entropic but understand for what audience they've wrote this paper so a direction labeled let's say panic no we have somewhere in our four dimensional vector space here a vector that is clearly having a direction here to the panic subspace cone but for example what it does mean it does not mean that like I've said your I've read on Twitter. This EI system feels now panicking. The eye system is panicking and it the eye system has emotion. This is all because what it means is changing now the activation direction which we do manual. We have an we insert here our new values to this vector. Of course, it affects the computation and the output probability that are now associated with a panic related word or a sequence or a sentence or a language part. Of course, we adjusted it for panic. So guess what the result will be that it indicates to panic.

[00:18:48]Yeah, I don't think that this is an indication that Yeah. So what is the scientific reflection? It uses calculus.

[00:18:55]Come on, there's no magic. There's no super space or some some members here have written here on on on Twitter, forget about it. So, it's a Jacobian matrix to measure the internal residion stream at the layer L and simply ask, hey, if we tweak the numbers right here, just a little bit, which final vocabulary words are most likely to pop out at the end. This is it. This is the main mathematical insight.

[00:19:23]You just modify the internals of this black box in a particular way and you look at the end the final vocabulary of words are suddenly stress or spider or aunt or grandmother.

[00:19:34]Great. So it maps abstract mathematical vectors in a particular convex cone structure of the residual space to the eyes actual dictionary. This is all there is. There is no selfawareness.

[00:19:48]There's no self awakening of the eye like I've read on Twitter. This is all [clears throat] not what you should um assume.

[00:19:57]What is now my working definition and this is a working definition. This is not a final one. I think a Jacobian lens can be understood as an output oriented linear interpretability method that maps an intermediate transformer activation through a corpus average Jacobian of downstream hidden state dynamics and then through the vocabulary unmbedding matrix. So the classical way that we calculate our residual data stream in the transformer.

[00:20:28]So again we do have a complete mathematical understanding of a transformer and we would really be amazed if suddenly there is now a new JS space because yeah but who knows maybe there is some new configuration but now we understand it is here in the activation space just a family of coordinates. So please whatever you read bring it down to mathematics. Great.

[00:20:54]Yeah, as I told you because the J lens is explicitly designed to find vector direction that predict verbal output.

[00:21:00]Finding now that these vectors predict the verbal output. Oh wow. Uh yeah, but some would expect it. No, because we built the tool exactly for this. So we cannot use this as an argument. Yeah, you get it. Yeah, I mean it's a nice tool. Don't get me wrong. It's a nice tool, but it it creates a semantic scratch pad. That's all that it is. Not a literal stream of consciousness.

[00:21:21]Careful. Entropic is here. It's a masterpiece. They're on the edge. No.

[00:21:27]And I can understand that some other people who read the article and then published their knowledge on Twitter say, "Oh yeah, this is the proof of consciousness of AI." No, it's not at all. It is a semantic scratch bed. Just get it right.

[00:21:43]This way back the immediate internal reasoning intervening on them is sufficient to redirect the conclusion.

[00:21:49]Guess what? Yeah. Now in multihop logical prompts here asking for the number of legs on a web spinning animal and this is their main proof. This is their main example. So let's talk about it. The J lens reads the word spider maybe somewhere in the middle layers of our transformer before the output generation at all. And now guess what?

[00:22:11]When the researcher now manually open the black box and overwrite hard code overwrite the mathematical vector for spider with the mathematical vector in this 4,000 dimensional space but the object ai outputs now six instead of eight. Why?

[00:22:30]Because spider has eight legs and an aunt has maybe six legs. And this is for them the indication or the evidence the proof that this JSPace is something above a scratch pad. I don't think so.

[00:22:49]Why?

[00:22:51]Assume we have this magic space this Jspace. So one question I ask myself what is the complexity of this JS space?

[00:23:00]Now how big is it? How complex is it?

[00:23:02]How much can this workspace hold on complex information? Does it hold full sword? So what is it able to do? What is it? Now entropic claims and just to make it clear that this workspace holds on the order of tens of concepts at a time and displays directed modulation. Now directed modulation we understand because the J lens was built for this.

[00:23:24]But tense of concept what is a concept now in the definition? What is it mathematically? What is it exactly? Now it turns out if you read the paper and hear my interpretation and I might be wrong please correct me in the comments when the preprint says tense of concept the data shows that it actually means token directions but sorry a token direction is something different than a concept no so maybe a little bit [clears throat] so if the J lens reads spider legs and eight it isn't holding a grammatical sentence the spider has eight legs It is just a statistical bag of words.

[00:24:06]Why? It is not analyzing sentences.

[00:24:10]It is focusing on words. It is focusing on tokens that are maybe subwords.

[00:24:17]It has not the semantic complexity of a grammatical correct English sentence.

[00:24:22]This is above its pay grade.

[00:24:26]So careful what you interpret into the functionality and into the possibilities of a J lens. It is a statistical bag of words. And if you have ever coded your bird or your sentence transformer, you know exactly here, you know the first part of the T5 transformer, the encoder part, the bird part, you know exactly what a statistical bag of work means and what it caused us for problems five years ago when we started with sentence transformers here.

[00:24:57]Now the preprint kind of claims that if you tell the eye don't think of X the J lens shows that the eyes is struggling to suppress no X. But I'm sorry in my understanding a transformer the mere present of the word let's say X in the prompt mathematically activates the vector for X. No I mean come on. And this is a standard semantic priming. And this is not an indication that the eye has some psychological struggle going on inside the black box. In my interpretation, what is this paper entropics most important empirical claim for me and maybe for you? I don't think that it is that the JS base contains here literal words if you go for token equal words identities. I think it is that some of these vocabulary defined directions align with naturally occurring intermediate variables inside the blackbox transformer and causally participate in later computation. I personally think that this is the sweet spot the beauty of this paper by entropic.

[00:26:05]Now let's talk about this global workspace idea because remember we are mapping the idea of a global workspace from neuroscience to computer science.

[00:26:16]Is a global workspace here really defined in neuroscience?

[00:26:21]Hm. I would say the human theory used as context here is itself unsettled.

[00:26:28]And yeah, wherever you read here with this ignition and the prefontal representation of this particular theory here of human consciousness, I refer you to an article by nature 2025. And just to tell you here my interpretation of reading this article, it did not validate here this new theory as established neuroscience.

[00:26:50]So let's put it in friendly words. Maybe it is just a theory that maybe is correct or maybe not.

[00:26:59]But I understand why Entropic chose this particular model. Yeah. And here we have the article here by nature. So here you have all the document uh identifier go there uh published here April 30th 2025.

[00:27:13]It is open access which is beautiful here. One of the few articles by nature that really have here open access. You don't have to pay for this. You can go there and you see here the global neural workspace theory the GNWT and they have another one. Yeah, they go here with an integrated information theory. But just focusing on the part of the global neuro neural workspace theory. My goodness. And yeah and if you have a look at this you will come to the conclusion it is not validated as the correct theory.

[00:27:47]So this means entropic is chosen here and this is 2025 a model that they know that is not empirically validated but it fits beautifully. Yeah, because this model here, this neuroscience model predicts kind of a special space, this workspace and this is what entropic can use in their presentation and now they call it a G-space and they can say hey look our machines are almost like human.

[00:28:17]So therefore we take now here as a little bit of a scientific approach the step we have to separate the term consciousness the term access and the term self-consciousness. Now for LLM models, no and I would argue we have here a ladder. We start with the representation. This is the AI system contains information about a particular fact X. Beautiful. It has a representation either as a vector or whatever functional access. That information can influence reports in multiple computation. And then we come now to an interesting part of meter cognition. This is here where the eye system represents aspects of its own processing or of its own reliability.

[00:28:58]This is where my last two three videos were about midcognition. Can we train an IML to listen inside its own let's call mathematical processes and train it to trigger certain events and understand when I see this trigger here in my activation I have a spike in my activation manifold. Hey this is here an adversarial prompt that was injected here in my sort processes.

[00:29:25]Then we have the level four. So this is really high a self-consciousness level.

[00:29:29]No. And then we have here the phenomenal consciousness itself. No. So I think it does not at all establish anything close to level four or five. You might say level one. Yes. Level two. Yeah. Partly and level three. We are starting to talk about it. Now we are here in the first weeks of having here the first study coming out here on meta cognition. But this is it. This is this is the limit.

[00:29:53]This is the sky.

[00:29:56]Now I have to say I like this study by Entropic because they provide of course you have to find it on the entropic server here. This is the file with this beautiful name of and they show you hm they have an external commentary on this particular study and they have it here from two expert in the field here of neuroscience and those are the people that they ask here to have a look at the study. This is beautiful. This is gorgeous. This is exactly what you should do. But please uh yeah why leave it out there because you hardly find it. Yeah. And I just show you here this is here from these two people here and they answer the question does claude possess or this our system entropic possess a conscious global workspace. So is it true what here entropic claims and they end here with caution in drawing parallels with the human mind.

[00:30:59]Now I showed you not only caution with the human mind but maybe the methodology that they use for JSpace the mapping here of G&W the global neuron workspace is not at all an established neuroscience evidence it is maybe a theory but this is where it ends. So yeah I think you have enough cautious word that you understand even neuroscience tell you caution in drawing parallels with the human mind. So yeah, Entropic burries this somewhere here at this particular link if you want to read it. So okay, I understand it. I understand exactly what they expecting next in their development of their of global corporation and they have to prepare for the IPO and therefore they have to have some material here to go to the press and tell them hey look our malls are just outstanding.

[00:31:53]Do I think the MS appear to possess a causally active verbal semantic workspace? Yes, it is a cone. It is part of the residual stream. Does it surprise me honestly between you and me? No, not at all. I was expecting that this is this mathematical operation matrix multiplication are happening in the residual stream. But for sure it does not demonstrate any form of consciousness of doi models. I'm so sorry if you're looking for this conspir this theory but no way no scientific facts now I want to be and end on a positive note so I thought what is this what is it special and I think it is special because it appears to be a vocabulary aligned causally active interface through which some intermediate results become available to different downstream computation because remember we're doing this somewhere in the middle of our layer architecture of the transformer former but again if you ask me is this a surprise I would say no I expected that somewhere in the middle of the layers and we understand that here the specialization of the layer and even of the attention head we have the specialization so yes some intermediate results will become available to different downstream computation this is what I also expected without having read the study but now we know it okay so h Yeah.

[00:33:22]What is it? What is tangible on this JS space? No. And I thought it resembles Yeah. If I can explain this to somebody, it resembles a combination of a computational scratch pad deep inside the blackbox, a shared variable interface where hopefully the variable is something similar to spider or aunt.

[00:33:41]And if you want a working memory like channel, yeah, of course, we are living here in the residual stream. we have not a subspace but a but a convex cone of this. So yeah h so interesting what is the strongest result I thought what what is really the main idea that is outstanding in this paper apart from that it is a beautiful marketing paper for the IPO I think the strongest result is not that the researchers can attach the word spider now to a particular activation but maybe it is that manipulating here the spider aligned component in the middle can change a later computational state in a let's call it sematically appropriate way. So this means neither the particular vector for spider or ant needs to appear needs to appear in the output but in the middle of the layer reasoning we have a semantic appropriate understanding of what this animal could be.

[00:34:46]Now either you are now totally excited or you say yes and what is special about this. It depends on you personally if you value this paper. But you see h and I think jpspace is not special and here contradict here what tropic writes because it contains literal words as a token representation here large stand. I think it is special because some activation directions as we just noticed define by future verbal effects also align with naturally occurring intermediate computation and can be causally reused across task where the corresponding word is never produced.

[00:35:27]So the idea started of this unspoken thinking we are now talking about a mathematical representation on a large language model in the middle of a transformer layer that was pre-trained here on specific embedding vectors of guess what language elements let's call that word on sentences and now we can causally reuse them across further downstream task. Okay.

[00:35:59]So, I'm sorry, but my objection here removes much of this rhetorical magic in the paper by Entropic about verbalized representation if you look at the title.

[00:36:11]But what still remains scientifically interesting, I think, is is there, but it is narrower. No, a partial alignment between the mall's output semantics and some of the internal computational variables. I don't think we really understand everything down to the lowest level of conformer transformer dynamics now no and we do have an alignment between those token defined direction and multiple downstream computation and I think this is a non-trivial result if you're surprised by it or not but however we have to be honest no scientifically what the preprint still does not prove I'm so sorry even a successful semantic intervention here does not establish established that a model has a discrete internal variable literally equivalent to the human concepts let's say of an aunt or of a spider.

[00:37:03]So there you have it. My first reflection here. It was published today.

[00:37:07]The paper I sit down. I read it and I thought I give you my first ideas. Maybe I'm completely wrong. Please read the paper yourself. Have your own opinion about it. Maybe feed it to your preferred AI system. Talk with your friends about it. Get an idea. Is there a JSpace deep inside a transformer or is this just a marketing gag by Entropic to show us here or prepare here the the investors here for some very specific element that distinguishes entropic from Open EI.

[00:37:40]I hope you enjoyed it. I know some of my viewers I got an email I will stop watching your video exactly at 30 minutes. This is not acceptable to have a video longer than 30 minutes. So, okay. I hope I'm not at 30 minutes and I hope you enjoyed it.