Unconventional AI Designs Non-Silicon Chip to Prevent Global Energy Wall

Unconventional AI 设计非硅芯片以应对全球能源墙 随着 AI 训练需求超出全球供电能力，该公司正利用非线性动力学替代矩阵数学以绕过冯·诺依曼架构限制，力求将能效提升三个数量级。 Naveen Rao 解释称，与生物大脑相比，拥有 80 年历史的数字及冯·诺依曼计算架构极其低效，因此当前的 AI 扩展势必会在两到四年内遭遇重大能源危机。为克服这一难题，Unconventional AI 正在从物理学第一性原理出发重构计算，利用非线性动力学和随时间变化的相互作用，来取代标准的矩阵数学与浮点运算。通过消除存储与计算分离所带来的能量浪费，该公司已成功研制出一款非冯·诺依曼芯片原型，旨在逼近每瓦智能的真正热力学极限。 前 MosaicML 创始人 Naveen Rao 成立 Unconventional AI，致力于将神经科学原理引入底层芯片设计；依托无历史包袱的初创团队与 AI 工具加速，团队仅用 6 个月就完成了原型芯片设计并计划在夏季流片。 (0:02) Rao 警告称传统 AI 硬件将在 2 至 4 年内撞上硬性“能源之墙”；全球总电网容量约 9,000 GW，而全人类 80 亿个大脑总耗电仅 160 GW，继续依赖 80 年前的传统数字计算机与浮点抽象将面临能源枯竭瓶颈。 (1:21) 生物系统的计算效率距离兰道尔原理（Landauer principle）的热力学极限仅差 1 到 2 个数量级（如松鼠实现复杂视觉与运动控制功耗低于 10 毫瓦），Rao 认为跳出矩阵乘法限制可将当前计算能效提升 3 个数量级。 (4:51) Unconventional AI 采用非线性动力学与振荡器同步机制，利用物理连续时间本身替代冯·诺依曼架构中昂贵的数据搬运，成功在实验中训练并演示了从随机像素平滑收敛的生成模型。 (8:27)

Naveen Rao, CEO of Unconventional AI, argues that artificial intelligence faces a hard energy ceiling within two to four years because digital, matrix-based computing cannot match the thermodynamic efficiency of biology. He proposes bypassing traditional von Neumann architectures with physical electronic chips that compute directly through non-linear dynamics. The Impending Global Energy Wall (1:21) Rao asserts that artificial intelligence is rapidly colliding with the physical limits of global power generation. AI training and inference clusters already consume multiple gigawatts, and Rao forecasts that within two to four years, the world will lack the energy required to support continued scaling. Total global power capacity stands at approximately 9,000 gigawatts, with the United States accounting for about 1,000 gigawatts across all municipal, heating, and industrial needs. In contrast, the collective intelligence of Earth's eight billion humans runs on only 160 gigawatts, with individual brains operating at roughly 20 watts. Rao contends that alternative generation methods like fusion or space-based solar fail to address the core problem: modern computing remains chained to 1940s-era digital and floating-point abstractions that were never intended for intelligence workloads. Thermodynamic Benchmarks and Biological Limits (4:51) Computing efficiency is bounded by the Landauer principle, which dictates the minimum energy required to alter a bit of information. According to Rao, biological systems function within one to two orders of magnitude of this absolute physical asymptote, whereas current 2D lithography digital chips operate far below it. For example, a squirrel calculating a ten-foot leap in windy conditions consumes less than 10 milliwatts—one-hundredth the power of a standard smartphone—while a macaque monkey operates on under one watt. While market leaders have scaled matrix-math hardware, Rao argues that gains in FP8 efficiency (accounting for memory access) have been incremental. Cost reductions have largely stemmed from manufacturing and packaging improvements rather than fundamental thermodynamic breakthroughs, leaving room for a three-order-of-magnitude leap through new physical architectures. Non-Linear Dynamics Replacing Von Neumann Computing (8:27) To escape digital bottlenecks, Rao models hardware on biological computation, which relies on noise-resilient, non-linear dynamical interactions rather than rigid floating-point matrix operations. Traditional von Neumann machines waste substantial energy shuttling data back and forth between processors, caches, and memory. In contrast, Unconventional AI leverages the physical time axis for computation by applying generalized Kuramoto synchronization—the physical principle where coupled oscillators synchronize over time—to electronic circuits with trainable coupling. In this dynamical paradigm, operators establish an initial state and allow physical laws to execute the computation autonomously, eliminating the energetic overhead of continuous state transfer. Silicon Prototyping and Continuous Generative Modeling (11:44) Unconventional AI developed a prototype chip design within six months and scheduled a physical silicon tape-out for the summer. The team verified that their dynamical systems can be steered along arbitrary state-space trajectories, demonstrating generative capabilities using animal image models. By initializing the system with random noise and backpropagating error at a specific time step, the hardware naturally converged into distinct image classes and smoothly transitioned between states. Rao describes this design as a true non-von Neumann architecture where state and computational function overlap completely within the device's physical substrate. Rao envisions that building non-von Neumann systems where physics performs the computation will circumvent the power grid constraints threatening AI development. However, proving that these dynamical electronic prototypes can match the versatility, programming models, and manufacturing scale of conventional GPUs remains an essential hurdle as the company moves toward physical silicon.

Unconventional AI首席执行官Naveen Rao指出，基于20世纪40年代浮点数与冯·诺依曼架构的现代数字计算正逼近物理能源极限，预计两到四年内全球能源将难以支撑AI的算力扩张。他主张摆脱传统矩阵数学体系，转向借鉴神经科学的非线性动力学物理架构，实现计算与状态在物理层面的深度融合。 算力扩张的物理瓶颈与能源鸿沟 (1:21) Naveen Rao警告，当前AI训练与推理的能耗已达吉瓦（GW）级，按当前轨迹发展，全球将在两至四年内遭遇AI电力供给的硬性墙壁。他列举数据指出，全球总发电装机容量约为9,000吉瓦（其中美国约为1,000吉瓦），而支撑全人类80亿大脑运转的总体功率仅约160吉瓦（单人脑仅约20瓦）。相比之下，机器智能达到等效负载需耗费兆瓦乃至吉瓦级电力。 Rao强调，解决该问题不能仅依赖算法与数据效率的改进，因为核心症结在于信息处理底层的物理基底。当前的数字抽象与浮点运算机制可追溯至80年前，这些机制最初并非为机器智能设计。若不从根本物理层重构计算范式，单靠太空光伏或受控核聚变等远期能源设想，无法扭转传统计算基底低效的现实。 朗道尔极限与生物智能的高能效启示 (4:51) 从热力学角度审视，朗道尔原理（Landauer principle）界定了改变一位信息所需消耗能量的绝对物理下限。Rao指出，生物演化出的神经系统在能效上距离这一物理渐近线仅差一到两个数量级，而当前数字芯片仍受制于二维光刻工艺与传统架构，能效远落后于生物水平。 自然界展现了极高的能耗效率：松鼠在风中跨越10英尺树枝跳跃时，其实时完成的复杂视觉与物理动力学计算耗电不足10毫瓦，仅为普通1瓦智能手机的百分之一；猕猴大脑整体运行功耗不足1瓦。Rao分析认为，尽管以Nvidia为代表的芯片企业在矩阵计算上做到了极致，但在计入内存访问开销后，FP8浮点运算的物理能效提升实则有限，芯片成本降低主要依赖半导体制造与封装技术改进，而非能量物理层面的突破。他认为通过转向全新物理基底，计算能效理论上存在提升三个数量级的空间。 以非线性动力学重塑计算架构 (8:27) 为逼近生物级能效，Rao带领团队从神经科学中汲取方案。大脑的计算本质并非执行矩阵乘法或严格的二进制操作，而是依赖神经元之间随时间变化的相互作用——即非线性动力学系统。同时，生物计算具备随机性并对物理噪声具有天然韧性，与数字计算中单比特翻转可能导致系统崩溃的特性形成鲜明对比。 团队利用物理振子自发同步的藏本模型（Kuramoto synchronization）原理，将其拓展至具有可调节耦合强度的电子电路中，使系统能够利用物理学本身的时间轴开展演化计算。在传统冯·诺依曼架构中，大量能量消耗在内存、缓存与处理器之间搬运数据的过程中；而在动力学系统中，设计者只需设定初始状态并施加触发信号，物理过程本身即可隐式完成计算。Rao透露，得益于无技术历史包袱及AI辅助设计工具，团队在6个月内完成了原型芯片设计，正推进夏季芯片流片，并已验证能够引导系统在状态空间内描绘任意预设轨迹。 动力学生成实验与无冯·诺依曼状态融合 (11:44) 在技术验证演示中，Rao展示了基于该动力学系统训练的生成模型：系统从随机像素初始状态出发，在预定时间步反向传播误差信号后交由系统自然演化，状态空间顺利收敛至特定图像类别（如马匹图像），并在不同类别间展现出平滑过渡的能力。 Rao将这种设计界定为新型计算范式的确立。他对比指出，CPU长于单线程冯·诺依曼运算，GPU长于大块操作数在存储进出间的高度并行冯·诺依曼计算，而Unconventional AI构建的架构使功能与状态在物理层面完全重叠，消除了存储与运算的物理界限。Rao总结称，在经过三十年的理论探索后，计算工程学正迎来通过工程构建实体系统来真正解构生物大脑运转机理的转折点。 Naveen Rao提出了跳出传统矩阵与浮点架构、利用非线性物理动力学再造类脑计算芯片的激进路线。然而，这种将状态与计算完全融合的非冯·诺依曼方案，其在实际夏季流片后的良率表现、面对通用大语言模型负载时的可扩展性，以及配套软件栈对现有AI生态的迁移成本，仍是该范式面临的关键工程挑战。

[0:02] Naveen Rao: bringing cool back to neuroscience. [0:02] Naveen Rao: unconventional [0:02] Naveen Rao: Conventional AI. [1:21] Naveen Rao: we just don't have any more energy in the world for AI. [8:27] Naveen Rao: non-linear dynamics [8:27] Naveen Rao: kick it and let it run, [8:27] Naveen Rao: Unconventional [11:44] Naveen Rao: galaxy braininess

[0:02] We're going to jump into our next set of [0:04] frontier talks. The first is which from [0:07] Naveen Rao. Naveen is a pioneer in the [0:10] AI space. He did his PhD in [0:12] neuroscience. Then he started one of the [0:15] first ever AI chip companies way before [0:18] it was cool. He's actually the person [0:21] who ran Mosaic ML, one of the first AI [0:24] training companies, and then built all [0:27] of Data Bricks AI. He left that amazing [0:29] role to do something new and redefine [0:32] the future of computing bringing cool [0:34] back to neuroscience. Naveen, come on [0:36] up. [0:38] >> [applause] [0:44] >> All right. Afternoon everyone. Yeah, [0:46] super excited to be here. I'm Naveen Rao [0:49] and CEO of Unconventional AI. So we are [0:51] unconventional because maybe it's [0:53] actually the wrong word to use cuz I [0:54] think we're going to change the name to [0:55] conventional. It's a actually a great [0:57] time to be a startup as as Boris alluded [1:00] Boris alluded to, right? Like having no [1:01] baggage actually I think is a true [1:03] competitive advantage. We can do things [1:06] so much faster than traditional [1:09] you know, sort of chip companies and [1:10] full stack companies can do. And I think [1:12] that's what's very exciting. We can get [1:14] to tape outs in months instead of years, [1:16] things like this. So anyway, let's get [1:19] get started. [1:21] I put that up there. I can guarantee you [1:23] some of you are like trying to prove me [1:24] wrong. All of a sudden you're like, oh [1:26] my god, what's you know, I can hear the [1:28] I can hear the gears turning like that [1:29] that that can't be true. Well, let me [1:31] let me explain my logic a little bit and [1:33] uh and and maybe the definition of what [1:35] I call ASI. So [1:37] really I think we need to get to much [1:40] greater amount of computer efficiency. [1:42] And when I say computer efficiency, I [1:43] don't mean algorithmic computer [1:45] efficiency or data efficiency. There's [1:47] lots of people working on these [1:48] problems. I actually mean the [1:50] fundamental substrate. Actually how I do [1:53] information processing at the like [1:55] physics level. [1:57] We chose a path of making computers work [2:00] the way they did for a lot of reasons [2:01] about 80 years ago. If you think about [2:03] it, like in the tech industry, how many [2:05] things have existed for 80 years? Like [2:07] not a lot. The digital abstraction sort [2:10] of floating point numbers, those were [2:11] around from from the 1940s. And for [2:14] machine that was built from a [2:15] completely different substrate and for a [2:18] machine that was built for completely [2:19] different purpose. And now we're [2:20] building machines for intelligence. So [2:22] let's think through this a little bit. [2:24] So part of this is that AI will make us [2:27] more efficient. I mean, we've talked [2:29] about coding and you know, running a [2:31] thousand agents on your phone, all this [2:33] kind of stuff, right? It does make us [2:34] more efficient. But at some point we [2:36] start to get to a place where what what [2:38] what is efficiency really mean? If you [2:40] think about energy, actual energy, maybe [2:43] we're actually not more efficient. And [2:44] we're kind of budding up against those [2:46] limitations of the physical world now. [2:48] Um you know, already today we're we're [2:52] using many gigawatts for AI inference [2:54] and training. And we're going to get to [2:56] a point within the next couple of years. [2:58] This is not 10 years away. This is two, [3:00] three, four years where we just don't [3:02] have any more energy in the world for [3:03] AI. Then this topic starts to become [3:06] very very important. Right now you can [3:08] kind of look at it like, well [3:09] electrical energy versus like food. [3:11] These are two energy sources for [3:12] intelligence. [3:14] Um and you know, there's no real [3:16] limitation on the on the electrical [3:17] side, but that's going to hit a pretty [3:19] solid wall very soon. And you know, [3:21] we're talking about going to space, [3:22] we're talking about building fusion [3:23] reactors. Great. Let's do all those [3:24] things, but still these fundamental [3:26] physics apply. [3:28] So [3:29] if we think through this a little bit, [3:30] we have about 8 billion people in the [3:31] world. Our brains use about 20 watts [3:33] each. That's only 160 gigawatts. So the [3:36] entirety of humanity is 160 gigawatts. [3:39] So just put that into comparison, we [3:41] have about 9,000 gigawatts of capacity [3:44] in the world today. The US has about a [3:45] thousand. [3:46] And you know, this runs everything. This [3:47] is like heating in your home and you [3:49] know, all the things, electric cars, all [3:51] that kind of stuff. But if we said we [3:53] you know, maybe you got 50% more of this [3:56] and we say okay, great. Now we got like, [3:58] you know, 4,000 plus gigawatts. But the [4:01] problem is our current [clears throat] [4:03] paradigm of compute is just vastly more [4:05] inefficient. So a computer I'm I'm [4:07] making some numbers up here. I mean, I [4:09] can say like if I'm running inference [4:11] per token, I can come up with numbers, [4:12] but you know, fully loaded like the [4:14] amount of energy that goes into [4:15] inference, building the model, running [4:17] the model, all of that. Let's call it a [4:19] gigawatt, something like that or you [4:21] know, definitely in the megawatt range. [4:23] But humans are on the order of of 20 [4:25] watts. And you could also argue that [4:27] evolution over 4 billion years has [4:28] created what we are. But the reality is [4:31] today our constraint becomes how quickly [4:33] can we get learnings happen? How quickly [4:36] can we build intelligence on a given [4:38] amount of energy? [4:39] So [4:40] if we want to build this future where we [4:41] have lots of intelligence in the world [4:43] and we're automating all kinds of things [4:45] that we want to really be more efficient [4:47] from an energy standpoint, we're we're [4:48] going to need a lot more we're going to [4:49] need a lot more watts. [4:51] Or we can think about building a vastly [4:53] more power efficient computer. And [4:55] that's where we come in. [4:57] So I love this curve here and I think [5:00] most people really haven't really [5:01] haven't thought about this cuz you just [5:02] sort of assume the computer is a [5:03] computer. And we haven't really [5:05] questioned that. Again, this is the [5:06] unconventional part is like, let's break [5:08] that apart. The [5:09] the assumptions we made 80 years ago are [5:11] actually not quite valid anymore. We [5:13] just choose to keep building on them [5:14] because I can build a product in two [5:16] years. I can make something that I can [5:17] sell in two years. But we're kind of [5:19] taking a different tax like, let's go [5:20] back to those first principles and see [5:22] if we can build something much much [5:23] better. [5:25] So there is a thermodynamic limit to [5:27] intelligence per watt. Okay, that means [5:29] you just can't do any better. There's [5:31] there's something called the Landauer [5:32] principle which some of you may know, [5:34] which basically suggests how much [5:35] compute could happen within a certain [5:37] amount of energy. So there is a physical [5:40] reality that we can't we can't get past [5:42] and that's sort of this asymptote here. [5:44] Now biology is somewhere up here. It's [5:47] actually pretty darn efficient. [5:49] 4 billion years of evolution have [5:51] created something that is actually very [5:52] efficient. However, it's not at the [5:54] asymptote yet. There's probably an order [5:56] of two of magnitude between those two. [5:59] We're actually here by the way. We're [6:02] we're down here. And I think limits of [6:04] 2D lithography, that's what chips are [6:05] built on today is call it somewhere down [6:07] here. [6:08] And I think with focused effort we can [6:10] get to the point where we're pushing the [6:12] limit of that. [6:13] And so we'll have something that's this [6:15] is this by the way is something like [6:16] three orders of magnitude from where we [6:17] are. It is very far away from where we [6:20] where we could be in terms of energy [6:21] efficiency. And so really that's what [6:22] we're focused on today. [6:24] So how do we do it? I mean, yeah, this [6:26] is great. Make something more power [6:28] efficient and wonderful, right? But the [6:29] reality is we it's not so simple. We [6:31] can't be thinking about the computer in [6:33] exactly the same way as we have been. [6:34] It's not about a machine that runs off [6:36] of runs matrix math. That's been the [6:38] simple way to move forward. Nvidia of [6:40] course has owned that market and [6:42] continue to push the envelope. But if [6:43] you look at the power efficiency [6:45] numbers, actual power efficiency of [6:46] delivering an FP8 flop for instance, [6:50] it's not that much better. Costs have [6:52] gotten better cuz manufacturing's gotten [6:54] better, our ability to package has [6:55] gotten better, but actual energy per [6:58] flop with memory access has not gotten [7:00] better. It's very very incremental now. [7:04] So I am a [7:06] a neuroscientist. I was a computer [7:07] architect for 10 years before that. So [7:09] I've been thinking about this problem [7:11] for a long time, actually on the order [7:12] of 30 years. So it's a very exciting [7:14] time for me personally. And you know [7:16] biology really does provide an existence [7:18] proof. I mean, you can argue that okay, [7:20] the tokens per second out of a human are [7:21] lower than a machine, but the [7:23] intelligence is higher. We still haven't [7:25] gotten to the point with these gigawatts [7:27] that we're throwing at it that we're [7:28] we're rivaling a human's intelligence in [7:30] terms of discovery. We'll get there. [7:32] We're going to get there in a very short [7:33] amount of time. But it's going to come [7:34] at the cost of a lot of energy. [7:36] So what I think is most interesting here [7:38] is actually not just [7:40] that brains are human brains are 20 [7:42] watts, but that the watt is kind of [7:44] scales with the with the weight. A [7:46] macaque monkey's brain is probably less [7:48] than a watt. And actually you see this [7:50] all through the mammalian world and I've [7:51] also the insect world. Like you have [7:53] very complex behavior for milliwatts. [7:56] Just for reference, your phone in your [7:57] pocket is about one watt. So you know, a [8:00] squirrel jumping from branch to branch [8:03] is running on a less than 10 milliwatts. [8:05] That's 1/100 of your phone. We can't [8:08] actually do this perfectly. Like [8:10] a squirrel's jumping 10 feet across from [8:12] branch to branch within wind and all [8:13] that stuff. We can't do that with a much [8:15] much larger computer. So biology still [8:17] created something quite amazing. And I [8:20] just I just don't feel like there's been [8:21] an appreciation for that. So I'm just [8:23] you know, just just just a little [8:24] reminder there. So now [8:27] great. We we see this kind of [8:29] phenomenologically like [8:32] biology is efficient can do can do [8:33] amazing things, but how does it actually [8:35] work? We don't really know. I will be [8:37] honest as a computer scientist and a [8:40] neuroscientist, but there are some ideas [8:42] that we can harvest from neuroscience. [8:44] And one of them is that the brain is [8:46] dynamic. It does not use matrix math to [8:49] do compute. It uses what's called [8:50] non-linear dynamics to do compute. What [8:53] this means is that there's a time [8:54] varying [8:55] interaction between neurons and that's [8:57] actually where the compute lies. So can [8:59] we extract that and actually apply it to [9:01] to synthetic circuits? Maybe. [9:04] They don't do floating point math. They [9:05] don't do matrix math. They do something [9:07] that can be characterized as such, but [9:09] it's actually much richer than that [9:11] because of these non-linear dynamics. [9:14] And they're stochastic. Brain's compute [9:16] is not a strict one and zero. In a [9:18] digital computer, if we're off by a one [9:20] or zero, the whole system falls apart. [9:23] So they're really not computers. So I'm [9:26] going to try to go through this quickly. [9:28] This is a thing called a Kuramoto [9:30] synchronization. So if you look at a [9:31] bunch of oscillators here and they're [9:33] kind of rigidly [9:34] coupled to each other on this plank, you [9:36] know, you'll see over time that [9:38] they start off in any state and then [9:42] they actually synchronize. This is an [9:44] example of a of a of a contracting or or [9:46] or converging dynamical system. So no [9:49] matter how you start it, it converges [9:51] and it's only based upon the coupling [9:52] between them. We can generalize this to [9:54] something that has kind of a flexible [9:56] coupling, call it a trainable coupling [9:57] between those those things. And then it [9:59] can have all kinds of interesting [10:01] dynamics. It can move through the state [10:02] space of dynamics in many, many [10:04] different ways. [10:05] So, if you generalize this, you can [10:06] actually think about an electrified [10:07] electronic circuit. You can say I have a [10:09] bunch of oscillators and they have a [10:10] fabric on which they're coupled. And now [10:13] when I when I make this fabric [10:14] trainable, I can actually see something [10:16] that's it starts to look a little little [10:18] bit more like the dynamics of the brain. [10:20] It actually has nonlinearities and they [10:22] interact with each other in very [10:23] interesting ways. It's actually very [10:24] rich represent a lot of information. [10:27] This is actual uh chip that we're going [10:30] to be building this summer. So, we went [10:31] from basically no team in January to a [10:34] to a full prototype in 6 months and [10:36] that's because of AI. So, this is what's [10:39] really cool about not having baggage is [10:40] you can do things in completely [10:41] different ways. [10:42] And the way you compute with something [10:44] like this is the traditional way would [10:45] be basically loop over some sort of [10:48] linearized time. This is how we do [10:49] things in a in a von Neumann machine. We [10:52] write state out, we retrieve it, we [10:54] operate on it, we write it back. So, we [10:55] keep going back and forth. Turns out [10:57] that's what burns most of the energy in [10:58] an existing computing system. [11:00] With something with nonlinear dynamics, [11:02] I actually just say, "Here's the initial [11:04] state, kick it and let it run." So, the [11:07] the physics themselves basically do this [11:09] computation and it does it sort of the [11:11] the state is an implicit. It's not an [11:13] explicit right. [11:16] So, in some ways you can think about it [11:17] if you take anything from this talk that [11:19] we use the time the time access of the [11:21] physics to do computing and existing [11:24] computing constructs do not. [11:27] And so, the question then is, "Can I [11:28] train this?" And the answer is yes. Um I [11:30] can actually steer the system into [11:31] multiple multiple different things. In [11:32] fact, we've sort of traced out in state [11:34] space a unconventional logo. That's the [11:36] idea here and we can train it in a few [11:38] different ways. So, yes, we can train [11:39] these systems and steer them into [11:40] basically any arbitrary set of [11:42] trajectories. [11:44] And can we compute uh can we connect it [11:46] to AI problems like uh image generation? [11:49] So, actually I have a better version of [11:50] this I can go to in a just a really [11:52] quick demo here. [11:54] Uh let's go to the next one. [11:56] We'll go to the demo. [11:59] Yeah, so basically what you see here is [12:02] uh something that's running on a on a [12:04] model of dynamics that was trained on [12:06] these different images. So, basically I [12:08] can say, "Okay, I have to use cats." I [12:09] think Andrew Ng is here. So, [12:12] >> [laughter] [12:13] >> homage to him, but we can do anything. [12:14] But basically this is a pretty simple [12:15] generative model. And I can basically [12:18] say like, "Okay, at time t equals 1, I'm [12:20] going to I'm going to backprop an error [12:23] um from randomness to to a particular [12:25] image class." And after that point, we [12:27] let the we let the system just run [12:29] naturally. And you what you'll find is [12:30] that it actually has clumped its [12:33] representation into places that are [12:35] meaningful. They're no longer just [12:36] random [12:38] pixels, but they're actually pixel [12:39] pixels that make different kind of [12:40] machine or different kind of animals or [12:41] whatever. So, for horses, I start off as [12:44] random and at t equals 1, you should see [12:46] it kind of converge into horse-like [12:47] things. [12:49] And then over time you'll actually see [12:51] it sort of morph between those. So, it's [12:53] already learned in the state space that [12:55] I can move between these different [12:56] things. [12:58] So, let's go ahead and move out of this. [13:02] So, this is really the emergence of [13:03] something new. So, CPUs actually do um [13:06] very fast single-threaded things the [13:08] best. Even today, uh it's faster than a [13:10] GPU. And really what you're doing is [13:12] this kind of von Neumann machine where [13:14] you're moving in and out of uh memory [13:16] and cache and doing operations. GPU [13:18] basically do this with multiple operands [13:20] at once. So, you move a bunch of [13:21] operands from memory, do some stuff to [13:23] it, write it back. [13:25] Compute in memory like Groq sort of did [13:27] the same thing, but just did it on chip. [13:28] It's kind of a a more fine-grained [13:30] version of this. [13:31] Now, what we're talking about is doing [13:32] something in a dynamical system. The [13:34] state and the the function are [13:36] overlapped with the physics themselves. [13:38] So, you now have no separation between [13:41] state and computation. [13:43] And you know, compute efficiency goes [13:45] goes up, of course, galaxy braininess [13:47] goes up. And uh this is truly non-von [13:50] Neumann. [13:52] So, with that, I'm just going to leave [13:53] you with this quote. It's been something [13:55] I've guided my entire life by and I'm [13:57] I'm really excited about this time [13:58] because I've been thinking about this [13:59] problem for 30 years. And we're at this [14:01] point where I think we can actually [14:02] start to understand how brains work cuz [14:04] now we can build them. Thank you. [14:07] >> [applause]