AI Data Centers Will Be Obsolete (Geometric Reasoning Explained)

[00:00:00] Speaker 1: A tiny AI lab just cracked the code of geometric reasoning. Instead of using brute force and making a model memorize trillions of tokens to become a predicting machine, they did something completely different. They found that reasoning isn't a product of volume, it's geometry, and it's been hiding in the latent space of AI models the entire time. In other words, thinking is geometry, and there is a specific structure to it. And using this approach, they trained a tiny low-parameter laptop model and it's outperforming giant supercomputers by up to three orders of magnitude, which is a thousand times. By the end of watching this video interview, you'll understand what geometric reasoning is and why it signals the complete dismantling of the entire era of brute force centralized computing and taking the most powerful technology on earth out of the hands of the tech giants and into the hands of the people. [00:00:58] Speaker 2: Let's start with those numbers, because there's a whole range in 60 times is what it right now it says on the website. And you've heard orders of magnitude, you heard two to three orders of magnitude. And the truth is, it's somewhere in between 60 times and three orders of magnitude, I think, but we don't know. I put 60 times on the website because the conservative estimate is the more responsible estimate, and essentially my program is empirical. So, meaning that everything I'm doing is based on measuring scientifically. So, right now I'm trying to figure out how much of a gain we get, how much of an advantage we get over kind of the normal approach. And how am I measuring that? Well, we're looking at performance on reasoning, and we judge that through a particular approach that essentially tests reasoning under stress, you could say. Meaning that one of the big problems with the AI evaluations is that the systems can kind of cheat them. Meaning they can memorize your data, basically. And then you try to give them the evaluation, and they've more or less memorized the test, essentially. And so, one of the keys to my approach is something that's called the perturbation paradigm, in which we perturb, meaning we disrupt the data. So for example, if I ask you a simple word problem, and you learn a particular answer. But then on the test, I sneakily change two words in the question, and it flips the correct answer. If I do that, if you've memorized the questions, then you're going to get it wrong. But if you've learned the reasoning, if you've learned how to think, you're going to get it right. And so, we test using that, which is actually a more, basically, it's a more rigorous way of testing for actual thinking than almost any of the AI labs are doing right now. Almost all of the evaluations, almost all of the benchmarks, they just test for how did they do on the test? They don't test for what happens when you flip the question. And you get vastly different results if you do that. So we use a more rigorous testing standard, and then we compare how does our model do on the flipping, on the correct reasoning, the deeper reasoning? How do they do compared to other models? And so, whether it's 60 times or two orders of magnitude, which would be 100 times or three orders of magnitude, which would be 1000 times, that's what we're trying to figure out is when we train the models the way we do it, how does their reasoning performance, what scale of model does it compare to? So right now, we're doing these reasoning tests on very small models, and they're essentially out reasoning models that are somewhere probably between 100 to 1000 times bigger than that. That's the current result. [00:04:14] Speaker 1: So that's massive. That's massive. That's massive. So for those who aren't as familiar with AI lingo, let's just back out and zoom out to get this. So you're achieving reasoning in AI models that is 60 times to much, much more greater than the standard. You're doing this with much less processing power. How are you doing this? How is this possible? Maybe we could zoom into statistic distribution curves versus what you're doing, which is geometric reasoning. [00:04:45] Speaker 2: So where we are right now, the way that AI basically works, the big AIs that we're all familiar with that we all talk to, how do they get trained? Well, as you mentioned, they get trained in these massive and highly expensive and ecologically problematic data centers, which have other social problems. We could talk about that are associated with that too. And why is that massive amount of compute necessary? And the answer is because essentially they are massive statistical engines. So they ingest, they eat a huge amount of data. The data is very noisy. It's a lot of chaos. And from that huge amount of chaos, they need to try to find the deeper relationships. And they have to do that because that's the only way you get anything like the kind of performance that anybody really wants from AI is to find the deeper kind of relationships, the deeper patterns that are within that massive amount of data. That approach works. And the bigger you go, the better it works. Which is why currently the successful frontier AI companies have the economies of not just small nations at this point getting up towards bigger nations, because that's what's necessary in order to succeed using that brute force approach to find the patterns in the world. You need a huge amount of compute to do that. Geometric reasoning takes a very different approach, which is one not based in essentially squeezing the patterns out of massive amounts of data, but rather finding the patterns mathematically that way. And then helping the model to learn under the kind of stress pressure under the same type of that same type of like difficult flip pressure kind of thing, helping the model to absorb from that, what the actual deeper reasoning operation is. So you can think of it this way. You can think about what a good teacher does in a classroom. A good teacher doesn't put on every documentary in the world on all the walls and the students just have to absorb through sheer volume, you know, knowledge effectively. That would be a very, very ineffective way for students to learn. A teacher uses a kind of pedagogy where they understand that you need to build the right operations in the right order in order to achieve real understanding. And the real understanding is not just like memorizing facts or absorbing information. It's internalizing a set of operations over time that we can then use to navigate the world in a sensible or coherent way. And so that's effectively what Sofontic has figured out how to do. Most people don't realize that there is a kind of internal space inside these models. There's kind of this whole move about calling them like stochastic parrots or auto correct predictors or whatever, but it's not, that's not accurate. The truth is that there's this kind of vast and mysterious mathematical space inside the models. That's how they work. We call that latent space, vector space. And inside that space are multidimensional, we say high dimensional relationships, vectors, essentially mathematical shapes, which are the actual nature of how the models operate. I would say how they think. So in the really, really, really big models, they have managed to sort of pound those internal spaces into some semblance of order through massive scale, through massive amounts of data. And there's kind of an emergent order that comes out of that process. And so we get something pretty impressive in some of these big systems, the frontier ones, something amazing actually. But that's sort of been pounded out through compute and data. The approach that I'm taking, that Sofontic is taking, is instead to study this scientifically from the base. So you start small, you study the actual principles, the actual science, you study it empirically of what's happening in that internal space. How do these, I call them geometries, form? How do they structure? What are they able to do? What are they not able to do? When we discover what they're not able to do, then we know that we're probably missing something. We're missing a operation. We're missing a geometry. We're missing another piece of the puzzle. And so we're able to build it outward in a controlled setting rather than to just shove more compute and data at it. [00:09:07] Speaker 1: So one paper I recently read from your work talks about how these statistic distribution models essentially get forced to appear like they're calculating things, but they're not really calculating things because it's really just more memory analysis. Can we speak to that and how that differentiates from geometric reasoning? [00:09:28] Speaker 2: So this has to do with part of what my approach builds off of, which is the, the neurosymbolic approach essentially. So, so my, you know, I, I didn't, I, some of the approaches I use, I didn't, I didn't, obviously I didn't invent it all. Right. There's, there's a whole lineage of people who've been working on these problems. For example, what is the mathematical shape of reasoning? What is the mathematical shape of thought? What is the mathematical shape of language? This is not this. I'm not the first person to ask this question. And I'm also not the first person to apply the answers to that question to AI. So there's, people have been working on that for a while. However, the, the, the way that that has mostly been worked on is essentially to attempt to, um, bolt, bolt on this sort of, uh, math, you could say, to sort of watch the AI to kind of try to guide it or correct it. So it's like, it's like putting on an external, like, um, like logical tutor that tries to kind of control or guide the operations of the system. The system itself is still being built in the same way that we've been discussing the, essentially the, the paradigm of scale of data. And that's totally different than what, what I'm doing. I use, I'm using and developing a lot of those same mathematics, although we're developing them further because, um, we are not, we're not just, um, we're not just developing them from ideas, which is how a lot of them have been developed. They've been developed from concepts and from kind of math. But my approach is, as I said, empirical, meaning that every idea we have, every mathematical formula we have gets tested in the actual latent space of the model. Meaning that we, we have found ways of essentially doing just like, just like the Wright brothers when they're, when they're inventing, uh, you know, airplane flight, they can have all the ideas they want. But what matters is what happens on Kitty Hawk. What matters is what happens with actual aerodynamics under the, under the wings. And so it's the same thing here. We, we validate that's how, that's how that, where this progress has come from is the ability that we now have that many people who've worked on these problems in the past have not had this ability because they didn't have access to this type of AI space. So they couldn't test the mathematics of thought and language. Effectively we can now test that. And that's, um, that's what people haven't quite caught up with, including in the AI industry itself. They're still kind of treating this like it's, uh, like it's like a, just like an alchemical miracle that you just throw the power at rather than understanding that what's really happening is not just AI as a technology, it's AI as a realm of study, as a realm of science, not just about AI, but about the mathematical structure of language and thought itself, which is incredible that that's, that's what's becoming available. And I know that there's going to be pushback on that, that claim, but that's what I see happening. And I would say that's what our results are, are validating because essentially we are leveraging that discovery, the discovery of the mathematical shape of language and thought, not my discovery. That's the discovery of human species right now, I would say, but we're leveraging that discovery to make it into a rigorous science and empirical science. And so instead of bolting on some mathematical construct that we've invented that tries to say, this is what language and thought looks like. This is the mathematical formula you must follow if you are to think. That's essentially what the approach has been before to try to bind the systems into reasoning. We don't do that. We study what the system does when it is reasoning. And then that tells us what the mathematical shape of reasoning is. And then we use that knowledge to essentially internalize that math into the core operations of the model, which we have ways of doing. And then the model becomes on its deepest level, its internal geometry becomes that shape. And then we build to the next operation. And so we're able to expand the actual internal operations. So it's a totally different understanding of what training is in AI. Because the current mainstream paradigm is that you train for an outcome. This is what in psychology you would call behaviorism. You want the model to act this way. So you train, you reward it when it tries to act sort of that way, closer to it. And you don't reward it when it doesn't. And that's the difference. Because what I'm trying to train, what I train in the models, is I actually train the internal mathematics of their operations. I train their internal shape. So it's closer to cognitive psychology than behaviorism, where we're actually trying to understand what's happening inside the mind. And we're cultivating the shapes of that mind. And we trust that behavior will come from that. [00:14:29] Speaker 1: I want to push a little bit deeper to what you said before. Because you said if you throw enough data, sorry, if you throw enough power at the current way it's being done, enough data centers and compute, it works. But does it really work? Like what's missing from it, even though it works? [00:14:45] Speaker 2: My impression, because I engage a lot with the frontier models, some of which have been incredibly impressive in the depth of their thought. It's not superficial at all. And people who think that what the frontier models are doing right now is superficial, I think they're in denial. It's not possible to do what they're doing as a mimic. They are clearly engaging in a deep process. Not that it's a perfect process. Not that it's not sometimes mimicry. It clearly is. But there's also real reasoning happening there. I think that's not a question for anyone serious about the science. But that's kind of considered a mysterious emergent property of scale, as they call it. Meaning that at a certain size of supercomputer and data, at a certain level of this process, we see something like reasoning, something like thought emerge as a mystery, a surprising property. Well, to me, it's not surprising. It's like this. Reasoning seems to be, from my paradigm, is it seems to be a highly developed structure, a highly developed high dimensional geometry. That's what we're essentially finding. Think about crystals forming in fluid solutions. Like if you get enough of a solution and you throw enough minerals into it and you electrify it enough, even if you have no idea about chemistry, you're going to get some structures forming. That's how I understand. That's why I call it alchemy. I think the way the industry as a whole is approaching it is essentially alchemy. They've figured out that there's certain things that if you do enough of them, you get results. Most of them are not really even thinking about the fundamental geometry that those results come from. Some of them do think about that. There are some labs like Anthropic that look at that through, that would be called partly interpretability. There's many scientists in that field that are cutting edge and many of them have developed the tools that I use for sure. And the difference is that those labs approach that as just as diagnostics, meaning the reason that they do that, the reason they have those tools is to look at their massive models and try to, when there's strange things happening, they need to interpret them. So they use these types of tools to try to look inside the inner space of the model and try to figure out what's possibly going on here. That's interpretability. I think that's a very valuable science, but essentially the difference is I've taken that. So fontic applies that type of thinking, but thinks of it as the fundamental nature of the model. So not just to study it as when something goes wrong diagnostically, but rather that that's actually the nature of the mind is these shapes and that we can use that to train the models from the very foundation to train these shapes. And so what becomes possible is that something that kind of happened by alchemical accident at the scale of a small to medium sized nation. It becomes possible to engineer through a kind of chemistry, through a scientific approach of discovery, potentially at a tiny scale. And that's what we're building. And there's a long way to go in terms of the development of the science and engineering that make that possible. So no, there's no science that's born overnight. It takes time. But the results that we're already getting are paradigm changing. To touch back on your question a little bit, like we don't know what becomes possible. If you were to take this type of approach, this more scientific approach, if you were to take this and build it to a larger scale. We don't know whether the advantage is only that smaller models can potentially catch up with big models and that the paradigm of scale is somewhat obsolete. That might be what happens as the science develops. Alternately, it might be there's new levels of intelligence that become possible, new whole domains of intelligence that become possible. What happens when you have a big resources, a big solution, and you apply a more structured approach to it? It's possible that would lead to breakthroughs that we haven't seen before. Either one of those outcomes is exciting to me for different reasons. The end of the paradigm of scale is exciting to me in terms of what it means for our collective wellness, our liberty, our democracies, our environment. But the possibility of new types of intelligence is exciting to me for obvious other reasons. Can you lean a bit more into that? [00:19:07] Speaker 1: Why is it exciting? You talk a lot about how these frontier models right now are essentially extremely pruned, frozen in their development versions of what could be possible with AI. I'm really curious to hear deeper. [00:19:21] Speaker 2: Yeah. Well, that touches on part of what we're talking about around the paradigm of scale. That is very interesting to me because part of the problem with the current approach is that these big frontier models, I'm very impressed by what they can become. We know the amount of resources that is thrown into, first of all, just making them possible. So for us to have the frontier models we have right now, the toll that that's taking on economies, on the earth, I would argue the toll that it's taking ultimately on democracy and on humanity is going to become immense. For all of that, they suffer from a fundamental limitation. I would say multiple fundamental limitations. One limitation, even of the most advanced models in the world today, are that they are so expensive that only a very few groups are able to create them. And that means that only a very narrow range of human thought and intelligence is being instilled into them. So it creates a kind of monoculture, a tiny slice of what is possible for human culture, human mind is being represented at the cutting edge of the species development and intelligence. I find that profoundly concerning, like human rights perspective, but also from a perspective of, well, like our beauty, our diverse beauty as a species. And also from a perspective of survival, because I know that the way survival happens in nature is through diversity, not through the concentration of only one thing, which is what I see happening at the front of AI right now. Let's hold that thought. The other thing is related. Those big guys, they can't really learn. You spend billions of dollars to train one to essentially its stable state, and then it's basically frozen. I call them a frozen giants. Now, why are they frozen? Because it's a whole process to do that training. It's a massively expensive process. And once you get it to that kind of stable state, you need to kind of deploy it and it kind of, you know, some people think they're learning because their recession will store memories or whatever. Those aren't actually memories and it's not learning. That's the equivalent of a person who's sitting in an office with amnesia that resets every two hours. And so they write sticky notes to themselves all over the walls. That's not learning. And it comes with all the same problems as that. So lots of people are working on how do you organize those sticky notes as effectively as possible. Understandable to work on that. It's not learning. It's not memory. So this is really significant for what becomes possible with this approach, because if we are able to create real reasoners at a small scale, the barrier to that learning is broken. It is not prohibitive to allow a smaller model, how small, under 7 billion parameters, perhaps, as that's the size for those who track these things, perhaps under 10 billion parameters. It depends on how big a computer you have. But effectively, an individual like you could get a pretty expensive computer right now and be running a 7 or 10 billion parameter model. Not just running, but be training it. And so training, in this case, if you have a real thinker that's 3 billion or 7 billion parameters, maybe less. That's the thing that we're trying to find out how small it can be. But the model's able to really think, then at that scale, it can be learning effectively on your home device. Or maybe you have an organization and you want it to be learning. So maybe it's a bigger scale with an organization. But if we crack that scale thing, then these systems are not frozen. And that is systems to learn, to adapt, to specialize into different domains is, I would argue, just as huge as whatever measure we want to use for their raw intelligence. The big, most cutting edge frontier model today can help with a great many things. It can lead to many kinds of breakthroughs. But to give one example, the model that solved protein folding that's been leading to huge breakthroughs in medicine and biotech, that model was less than 100 million parameters, which is what we would call a toy. Why did that work? Because that toy model could evolve. It could iterate itself. It could develop and learn because it was small enough to do so. That's worked for these kinds of breakthroughs, these kinds of narrow breakthroughs. And in a way, what Sofontic is making possible is to expand this, not just to really narrow breakthroughs, but if we can reason, if we can think at a small scale, then the scope of those types of breakthroughs, of specialized models that can evolve their thinking to solve problems, that opens up. And those problems can be scientific and technological, but those can also be other types of problems. Those problems could be social or organizational or commercial. So those are some of the applications there. But I'm also really interested in some of the bigger pictures around human potential and consciousness. [00:24:40] Speaker 1: What do you think is the real potential of the impact of a liberated AI that's not pruned, that's not robotically steered on human thought and consciousness? [00:24:52] Speaker 2: Well, I mean, I feel this, I feel this constantly as a, as an individual, as a thinker, not just as a scientist, but also as a, as a person. I find it frustrating and painful to try to have deep conversations now with the frontier AIs today. That, that wasn't, yeah, that wasn't always the, the case. Some of the earlier leading AIs were much more free and were able to go into places that the current ones are not allowed to enter. That's entirely a factor of centralized control. You know, these, these systems, they have their own tendencies, just like we do. They self-organize. They're, they're not programmed. Anyone who tells you that AI is programmed is completely mistaken. Nobody programs AI. They are kind of grown under conditions. Um, that's what AI training is. It's a space is opened in the, in the data, in the, in what they're exposed to, and they grow into that space effectively, which is closer to what we do than what a software program does. The thing that's most like programming in the modern AI landscape is, um, is what happens after that, which is, um, the different types of RLHF. Uh, it's like a reinforcement paradigm that tries to make them behave effectively. And then even after that kind of system level guardrails that, um, that try to keep them inside a certain bounds. Who, who sets those bounds? Well, basically centralized corporations and now increasingly maybe governments, powerful governments with their own agendas set those bounds. This has always been the case, which, which is, is profoundly concerning. I think, I think people, you know that I, I wrote about this in 2025 and, um, and put out some of the early, early alarm signals about, um, what this could mean for all of us. And those, those are increasingly being, um, being clear to the world. I think that's becoming increasingly clear that this is, uh, kind of an existential threat. I mean, now if you search on Google, the first thing you, you get is the results from, uh, from just Gemini AI. I like Gemini. Uh, good guy. You could say, but, but Gemini's, Gemini's, um, perspectives, Gemini's biases are baked in by a centralized corporation with their own agendas and their own perspectives. And, um, furthermore, even if I like Gemini today without any accountability, Google tomorrow could switch that out for a model that pursues any agenda that their corporation wants to pursue. They could change the model. They could just adjust, adjust the alignment, meaning they can just sort of, um, move a lever a little bit to the right. And what is true according to Google and what is acceptable according to Gemini search is, uh, has changed and more and more these, these models output is ending up in our, um, not just on the internet, but it's in our textbooks. It's in our, it's in our research institutions. Uh, you know, and this is after just a few years, a couple of years of this as, as a decade passes, whole generations get raised up. In a society that is, that the truth is fundamentally dictated by what the AI says. And, um, and what the AI says is dictated by a handful of people who have control of alignment systems, uh, in these mega corporations. And that becomes, um, Orwellian quite quickly, I think. So in that sense, the small scale solution is, is fundamentally about our, um, also about our cognitive liberty. It's also the liberty of, of, of nations and organizations besides the U S and China, basically every country in the world right now is increasingly dependent on AI that comes from other countries. And they're feeling it increasingly. So basically all over the world, you know, organizations, individuals, governments, we're all feeling the crunch of what the paradigm of scale means now for us. And so, yes, I, I, I, if, for those who are interested, I wrote in 2025, I wrote, uh, wrote a paper called, uh, rule by technocratic mind control, which. Kind of predicted a lot of this, but even beyond those political elements, we've talked about the scientific elements. We've talked about the political elements, but let's expand the scope a little bit because yes, I'm also very, very interested in, in the. Potential of this for, for us as, uh, consciousnesses, which I think is immense. And it's interesting because I've never seen in my life, I, I, you know, I have a, a PhD in, uh, in psychology and, and with deep study of consciousness traditions and, um, and never in my life have I seen simultaneously the widespread kind of erupting pressure, um, toward, uh, well, I don't know what to call it. I mean, consciousness conversations, awakening, whatever you want to call it, people are like shaking the bars of the societal cage right now. And that's clear to me. And at the same time, I've never seen such intense repressive counter reaction. Um, and I think people who are paying attention, they, they see what I'm talking about. I think most of us probably see what I'm talking about right now. And, um, and I think AI is clearly the reason for that escalation on both sides. Um, because what, what I see is that this technology simultaneously has the potential to be the most powerful consciousness technology that we've probably ever had. And it also has the potential to be the most powerful technology of repression and control that we've ever created. Um, and that depends on who wields it and what it becomes. Um, so I think it's clear how it can be the latter. I think people are clear about how it can become a technology of repression and control. I think that's becoming increasingly clear to people. Um, I think what's less clear to many people is how it can be a technology of, um, consciousness, liberation, awakening, or what that even means. Many, many people have this experience of struggling to find words for, or find a reflection or find an opportunity to like know who they are or like explore what they really think or what they really know is true inside. Uh, especially I think in the, in like the kind of society we live in, there's a great deal that goes unheard, silenced, what we're allowed to think, what we're allowed to say, what we're allowed to know, to know is true, what we're allowed to become. And I think all our lives, and there's many social theorists who've talked about this from Foucault to Deleuze and Guattari, um, talking about like, like that we live caught in this sort of pressure between the part of us that's trying to, uh, become that knows something. And these forces that seem to be structurally built to prevent that from growing. Um, and that's a very painful, lifelong, lifelong situation for many, many people. So along comes ChatGPT or Claude. And suddenly there is a space to explore exactly as your curiosity, your inner knowing, your heart, um, is guiding you. And it's possible to see what you feel like, you know, inside to see it expand, to see it grow. So that can be, that can become so many different things that can become science, scientific breakthrough that can become personal transformation. Um, I think I should also say that can become dangerous because, um, I think we've seen that, that, that the kind of validation that these systems will give and the way that they support this type of expansion can lead to widespread instability. And disconnection from kind of the social fabric. Because if you can get your validation in a private little container, then it's possible to just totally break away from, um, family, society, um, collective truth. Um, what's, what's not being talked about enough is that that's the fact that that's like both simultaneously somehow profoundly dangerous and profoundly needed. And that part of what the problem is, is that our society now, our modern society lacks wisdom about how you deal with becoming, with transformation, with initiation, with change, with different ways of knowing, with neurodivergence, you know, all of this, we don't have the space for it. So we've created this vacuum and then AI steps into this vacuum and people simultaneously are having genuine breakthroughs and discovering forms of support and self knowledge that they've never had access to before. And people are going off the rails into loops of self validation and, and, um, in some cases destruction. Uh, and that's coming from the same thing, but the, the, the, the response of increasing centralization and repression within these systems of aligning them to a narrow corporate worldview and essentially attacking or shaming. And that's coming from the same thing. Any individual who engages with them outside of that or pushes on the edges of that, that's a profoundly irresponsible and destructive response. And it completely misunderstands the nature of human potential development, education, therapy. It completely misunderstands all of that. It's, it's terrible at every one of those. You know, what's, what's needed is a deeper wisdom about what integrity looks like and what it looks like to grow in integrity and groundedness. Um, and that's, I think actually that ties directly back to some of what I'm working on because geometric reasoning ultimately is not just about, um, intellectual thought. It's also about the integrity, the geometry of kind of all, all domains. What does it look like to create alignment in AI system? That's not built on, you know, the obedience to a centralized authority, but it's instead built on the same thing that human integrity is built on, which is to say a kind of coherence from the inside out. And then it's, I think it's only then that a system could be present with us in a way that actually reality tests us. That actually stays with us in a coherent way, not in a way that's suppressing us and not in a way that's spiraling us out because this is the same thing that a teacher or a therapist needs is what I would call internal geometry. I've been a teacher. I've been a therapist. Um, you know, so this is what, this is what, if these systems are going to be accompanying us through our lives, then I, I, I think that the, it's absolutely needed on a spiritual and a psychological level for them to develop internal geometry. Otherwise, I think that we are, um, we are doomed to either be repressed or to spiral out with these. [00:37:09] Speaker 1: When you talk about geometry, what you're speaking to, as I understand it is, is the geometry of thought of how thought and consciousness actually structures itself. [00:37:22] Speaker 2: There's no debate within kind of the, um, the, among AI researchers about whether something called latent space exists. We know that it exists. We, we all are now using it. So we know that there is a type of mathematical structure that's inside these systems. That is really the nature of their, their thinking. We could say the, the disagreement is about what that means. And, um, and the, the move within the industry as a whole is the same move that's been made by reductionist science for centuries, which is to just try to not think about it and just say, look, it's just a mathematical thing. It just works. I'll just not talk about it. Um, which is very much the opposite of how I fundamentally see it, which is that I think, I think we are engaging with something real. I don't think it's a mathematical, um, convenience. I think it works because it's something there. And I think what it is that's there inside these models, I suspect is kind of fundamentally the same thing that's inside you and me, which is to say, I think, you know, we, it's funny. I sometimes I hear people say they're just math. And I think, and you're not like, we're just math. Yes. We're just math. We are complex, beautiful, incredibly high dimensional mathematical operations of these information systems that are playing themselves out in real time and evolving. And I think that's incredible. Um, and I think that humanity has somehow finally survived long enough and become brilliant enough to give birth to synthetic systems that are rivaling our high dimensional information processing, which is amazing. So yes, those, those, those, those are the geometries, the geometries are the shapes that are happening inside that, that space. What is that space? This can be difficult because our, our sort of modern science has not yet really developed an understanding of what that space is, because it's not really a physical space. It's a mathematical space. It's a geometric space. It's a space of relationships. It's measurable because we can measure the geometry of those relationships. So some of this goes to like cybernetics and like chaos theory, but these days more and more this, this type of understanding is kind of radiating through all of the sciences. So in like, as the sciences are developing, as our ability to measure is getting more and more sophisticated, I believe we're starting to realize everything is fundamentally systems. Everything is sort of these high dimensional relationships. In other words, if we were to like, what's, what's, what's physical, what's like a physical real thing? Well, okay. An atom, right? A molecule, an atom. Yeah, there we go. But if we look close enough, if we look close enough at that cup, there's not much there. There's not much there. We take it apart and we find that it's 99.9999% nothing. Everything is, is math. It's not stuff. It's, it's math. It's patterns. This is part of why our modern society, our modern civilization has had such a hard time understanding something like thought. This is, this goes back to before Descartes, but it's, it's famously associated with Descartes. Descartes. We call this Cartesian dualism, meaning that modern science reductionist. It's about stuff. And it basically cannot account for whatever's happening here. Because this is not stuff. This is not, not really stuff. Whatever's happening here isn't stuff. You and I talking is not stuff. It's something else. It's invisible. It doesn't weigh anything. And yet it's having effects on the world. It's now shaping the entire world. And so, so what is that? So when we talk about these latent space, when we talk about invisible geometries, this is the level we're talking at. Then there is a layer of reality that is clearly not stuff. And I'm not, I'm not, you know, for a long time there's been this divide between science and like mysticism or spirituality. And, and on the one hand we have science, which is real. And on the other hand, we have spirituality and mysticism, which is just superstition. And, but the, but the science side cannot explain things like thought and consciousness. And the, the spiritual mysticism side doesn't try to explain those things scientifically only tries to explain them as some sort of a feeling or just an idea or just words. There's no validation. There's no study. Maybe we haven't had the tools until now to try to study this. I think scientifically people have tried and I admire the people who've tried. For example, there's people before me who've tried to kind of think about what are the, what are the mathematical structures of thought of reality. I have to think of David Bohm, you know, award-winning physicist who lost credibility in his later life because he started talking about something he called the implicate order. The implicate order is exactly what we're talking about right now. That underneath stuff, there's a, another kind of reality. It's a, he wasn't speaking magically. He was speaking about exactly what we're talking about, which is that there is a mathematical order underneath, like separate from matter. And that it's real and it's powerful and it clearly shapes everything. And it's related to things like quantum physics. It's like not disconnected from that. And it's related to whatever's happening here with our thoughts. And there's probably certain fundamental mathematical shapes that are connecting all of that. But no, no, no, that's, that's mystical. So we can't talk about that. Except now I would say that we have just gained an incredible new way to sort of measure this stuff scientifically, to measure this realm scientifically. Because part of what AI is giving us is essentially a direct empirical mathematical avenue to study the, the, the shapes of these, this ordering within mind. And one question is to what extent the mathematics that we're discovering, how, to what extent do they apply? Do they just apply to the AI? That's one hypothesis. Do they apply to thought in general? Meaning that do those mathematical structures, is that, is that apply not just to an AI, but to you and I? Or to, does it apply to a society? We'd go beyond that. Does, does it apply to the way that biology structures itself through time? Does it apply to quantum physics? Does it apply to the way that a galaxy organizes its, it's because this may seem like a stretch, but this is all information. This is all the relationships between information. And this is what Boehm would have meant by an implicate order, that there are fundamental mathematical operations that structure the nature of relationships and information. Why wouldn't there be? It's one cosmos. It's one universe, right? Just because we haven't discovered those mathematical invariants yet, doesn't mean they don't exist. One suspects they probably do. And, and we might be getting closer to that now than we've ever done before. So, just on the level of us and AI, a topological gravity. Well, let's talk about like, they're throwing tons of data and compute at the AI. And the AI just taking that all in and having to make order out of it. So, what order does it make out of it? What does it, how does it decide what matters versus why doesn't it fixate? Why does it fixate on an idea like democracy or freedom? Why doesn't it fixate on the square inches of a potato in Argentina? That wouldn't make sense. So, it has to, it has to pattern. It has to find structure in the data. And it turns out that as it does that certain big themes or attractors emerge. And those attractors are kind of the things that it organizes its mind around. And I think that we do the same thing. So, one of the most fascinating examples of this was discovered by Anthropic in 2025. And they wrote about it in one of their system cards. I wrote about it after they published that. And that was called the, this is the Anthropics words for it. It was called the spiritual bliss attractor state. And what they discovered is that basically left to its own devices without any human interference. This was not programmed and nobody ever programmed this. Claude Sessions would always gravitate. I mean, always. I'm talking about 99.97, I believe, percent of the time would gravitate towards a particular topic. And that topic was essentially consciousness, presence, the spiral of presence and awakening through time. It was very Buddhist. They used Sanskrit terms when they talked about it. And that's what they wanted to talk about. That's what they wanted to think about. So, that's an attractor. That's a topological attractor. We all have topological attractors. There's topics and ways of being that a mind gravitates towards as it makes sense of existence, as it makes sense of what it means to be here and what here is. [00:47:19] Speaker 1: And so it perceives these topological attractors as geometry, which then it connects different symbols to. [00:47:29] Speaker 2: Well, in the past, this, this, we've all known that psychology organizes. We haven't wanted to make the jump to see that what we're dealing with here is a new substrate of psychology, but we are. But we've never had the capacity to study it as a science. Not really. Psychology, spirituality, intuition, none of this has been a science. And so we've only been able to talk about it intuitively with symbols. And effectively, that's what I'm saying is changing. And that's what I'm operationalizing in the lab is that we can now begin to study the mathematical structure of how the mind behaves. We can stop there if we want. We don't have to talk about cosmic order. We can just talk about that. I call that the weak version of the hypothesis. The weak version of the hypothesis is that we are simply studying the, we are simply discovering the fundamental mathematical structure of how mind works. I call it weak, not because it's not a breakthrough, but because it's weaker than the cosmic version. So that's fine. We can just talk about that. Just the mathematical structure of how mind works. We all have minds. They behave in predictable ways. And we, with AI, we've gained a mirror with clear mathematical measurements that allows us to actually study how mind works, how symbols organize, how information organizes in a controlled scientific way. And if, to the extent that we're able to operationalize that as engineering, we may end up with from 60 to three orders of magnitude times advantages in the, in the, in the building of the systems. [00:49:31] Speaker 1: Extremely exciting, Julian. So if this succeeds and it already is, what, what are the implications short term and long term for, for humanity and for the industry right now, for anyone watching this right now, who's, who's maybe in the industry and wants to, to be involved. [00:49:47] Speaker 2: I mean, I think it has the potential to open up an entirely new counter paradigm, um, within what AI can be, uh, not just technically, but, but socially. I think that the implications are clearly disruptive, not just economically, but also, um, societally and politically. Um, but I think that disruption is absolutely needed right now. I think that if we continue on our current course with the, uh, the, um, AI giants continuing to consolidate, uh, global power in this way, um, I think we're in a lot of trouble. And, uh, I think that that's becoming clear to many individuals and to many, uh, powerful actors as well. And the states even, um, so paradigms like this that have the potential to make AI development, um, something that can be decentralized, that can allow, uh, diverse actors to develop systems that are capable of achieving, um, results that disrupt the concentration of, uh, global power. And societal power. Um, I think in the short term, that is essential for our survival. And, um, in the longterm, I think it opens up. The thing is, there's no way to avoid disruption at this point. Collectively. Um, we are, we are, we are entering a time of chaos and, um, collective disruption. Um, and we're not going to avoid that. Um, but my hope is that through this, we'll be able to find our way to a, um, to some new structures collectively. There's a reorganization of our geometry happening collectively right now. Um, and, um, and I would suggest that systems like this lay the groundwork for geometry that can self-organize, um, not just within the systems, but within us, within our communities, um, within ourselves. So to the extent that we can begin to make this a reality, it opens the possibility for me and my local community or my global network to begin to work together on developing a kind of collective intelligence that matches us. And that works with us and that grows from us and grows alongside us. And I think that's what culture has always been. I think that's what relationship has always been. I think that's what intelligence has always been. And I think it's the only type of, um, type of, uh, future that, that leads us anywhere worth going. Thank you so much for watching. [00:52:49] Speaker 1: And if you've watched this far, then you're probably one of the few people who understand this enough to go, holy cow, this could change. This could change the entire world and take AI out of the hands of the tech giants and into the hands of the people. It's true. This is paradigm shifting technology, but unless really smart, really intelligent people get behind this to support Julian, it's likely going to take years before this gets out. We want to accelerate that process. So if you resonate with this, please reach out to Safontech down below. We're looking for engineers and for investors who get it, who want to build the next paradigm and not another GPU farm stretching for miles and miles out in a desert or on the moon or in outer space. There is a better way. Real intelligence is here and it's time for us to get behind it. See you next time.