我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎,当追忆的图片变得模糊,我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
架构革命从类脑物理计算到AGI
视频
音频
原始脚本
架构革命,从类脑物理计算到 AGI,AI 正告别海量数据依赖。 当约翰霍普金斯大学的研究团队公布类脑架构 AI 无需数据训练即可媲美传统模型的实验结果时。 一场关于人工智能发展路径的深刻变革已悄然拉开序幕。 这一发现并非孤立的学术突破,而是与非冯诺依曼架构、物理计算等前沿探索形成呼应,共同指向一条摆脱海量数据桎梏、通往通用人工智能 AGI 的全新赛道。 谷歌 DeepMind 的创始人哈萨比斯敢于预言,2030年前 AGI 大概率实现,其底气或许正源于此。 当 AI 从用精密电路模拟简单逻辑,转向用物理规律复刻自然智能,曾经看似遥不可及的技术起点正在被底层架构的革新逐步拉近。 一、核心认知颠覆,智能的本质是物理映射。 而非数据计算。 长期以来,主流 AI 的发展遵循着数据驱动的逻辑,以冯诺依曼架构为基础,通过 GPU 等精密逻辑门店。 将海量数据转化为数字浮点数,再经过复杂的量化运算模拟神经元权重。 这种模式下,AI 就像一台超级计算器,依赖千亿级参数和数百万张训练图像,才能在特定任务中达到人类级表现。 但这一路径存在着先天缺陷。 一方面,数据成为稀缺资源,许多场景下的 AI 应用因数据不足难以落地。 另一方面,能源消耗成指数级增长。 据预测,2026年 AI 数据中心耗电量将突破1000太瓦时,相当于数十个中小型国家的年用电量。 而类脑物理计算 与非冯诺依曼架构的出现,彻底颠覆了这一认知。 其核心逻辑可概括为数据及计算、存储及计算,这与大自然亿万年进化出的智能机制不谋而合。 生物神经元的结构远比 GPU 简单,却能实现高效的信息处理,关键就在于其存算一体的物理特性。 神经元之间的连接强度、权重,并非通过数字模拟,而是由连接的粗细及物理层面的导线数量、信号传导阻力直接决定。 信号越频繁通过某一连接,该连接的物理结构就会随之优化,如导线变粗、电阻减小、权重自然增强。 这本质上是一种基于物理反馈的条件反射。 这种逻辑在早期物理计算机中已有雏形。 美国最早的弹道计算机无需复杂运算。 而是利用抛物线的物理特性,直接通过测量物理轨迹就能求解二次方程。 类脑物理计算正是延续了这一思路,它不追求用高级电路模拟低级逻辑,而是让物理过程本身成为计算载体。 输入信号是什么样,经过物理结构的映射后,输出信号就呈现对应的规律。 训练与推理本质上是同一套物理映射过程。 训练时的信号反馈优化了物理结构,推理时相同信号输入就会得到一致的输出结果。 这种自映射机制无需复杂的浮点数运算,也不需要海量数据的反复投喂,却能实现高效低耗的智能响应。 恰 大自然的智慧,最简单的模仿往往是最本质的复刻。 二、学术前沿印证,架构优先而非数据优先。 约翰霍普金斯大学的实验为这一逻辑提供了关键学术支撑。 研究团队对比了 Transformer、全连接网络、卷积神经网络、CNN 三种主流架构,发现只有 CNN 在调整规模时,其活动模式会显著接近人脑视觉系统。 而 Transformer 的全局注意力机制、全连接网络的密集连接方式。 因缺乏生物学对应性,即便扩大神经元数量,也难以贴近人脑活动。 这一结果背后的核心原因的是,CNN 模仿了大脑视觉系统的层级化信息处理模式。 通过局部连接和权重共享编码空间信息,天然具备类脑物理逻辑的雏形。 事实上,全球已有多个顶尖实验室和科研组织投身于这一方向的探索。 除了约翰霍普金斯大学、斯坦福大学、麻省理工学院等机构正聚焦于一组器、光信号传导等物理载体的研发。 一组器能够直接通过电阻变化存储和处理信息,完美契合存算一体的需求。 光信号则可通过光路的强弱粗细调节权重,且能耗远低于电子电路。 在工业界,IBM 的真北芯片、英特尔的神经拟态计算项目,均在尝试摆脱冯诺依曼架构的束缚,构建基于物理规律的类脑计算系统。 这些探索共同指向一个结论,AI 的突破点并非在于收集更多数据,堆砌更多参数,而在于找到契合智能 能本质的架构蓝图。 当前的学术研究已逐步厘清传统架构与类脑物理架构的核心差异。 传统架构是数字模拟物理,用精密计算复刻简单逻辑。 如同用法拉利当拖拉机,既造成资源浪费,又难以突破效率瓶颈。 而类脑物理架构是物理本身及逻辑,用简单的物理反馈机制实现智能。 既符合大自然的进化逻辑,又能解决高能耗、数据依赖等痛点。 这种差异也解释了为何传统 Transformer 架构难以支撑 AGI 的实现,即便其参数规模突破万亿级,本质上仍是在数字世界中模拟智能,无法像人脑一样将实现高效、通用、低耗的信息处理。 而类脑物理架构则跳出了这一框架,直接在物理世界中构建智能,为 AGI 提供了更具可行性的技术路径。 三, AGI 的路线图,从架构革命到普及落地,哈萨比斯关于2030年前 AGI 大概率实现的预言,若脱离类脑物理计算与非冯 诺伊曼架构的发展背景,难免显得乐观。 但结合当前的技术前沿来看,这一预言更像是对架构革命趋势的精准判断。 传统 AGI 路径之所以进展缓慢,核心在于陷入了数据参数能耗的恶性循环。 为实现更通用的智能,需要更多数据训练,更多数据需要更大的参数规模承载,更大的参数规模则导致能耗成指数级增长。 这种模式不仅成本高昂,且难以实现真正的通用。 毕竟现实世界中的许多场景既无法提供海量数据,也无法支撑高能耗的计 算需求,而类脑物理架构的发展,正打破这一恶性循环。 其核心优势在于通用性与低耗性的统一,基于物理反馈的智能机制,天然具备适应不同场景的能力,无需为特定任务单独训练。 同时,物理计算的能耗远低于数字计算,能够实现轻量级的智能部署。 这意味着未来的 AGI 无需依赖千亿级投资的 超级数据中心或许可以像人脑一样,在小型设备中实现高效运转。 从技术路线来看, AGI 的实现或将遵循三步走。 第一步,基于忆阻器、光信号等载体,构建小规模类脑物理计算原型,验证存算一体的可行性。 第二步,优化物理结构的可塑性与扩展性,实现多任务、跨场景的通用智能第三步,降低制造成本,推动类脑计算设备的普及,让 AGI 融入日常生活。 这一路线图的关键在于突破物理载体的技术瓶颈。 当前,一组器的稳定性、光信号的调控精度等问题仍需进一步解决。 但随着材料科学与神经科学的交叉融合,这些障碍正在逐步被克服。 更重要的是,这一路线符合技术发展的性价比原则。 与传统架构需要持续投入海量资源不同,类脑物理架构的核心创新在于思路转变,一旦突破关键技术,其规模化应用的成本将远低于传统 AI 这也意味着 AGI 的实现不会是少数巨头的专属游戏,而可能成为全行业共同推进的技术革命。 人工智能的发展曾一度陷入数据崇拜与参数竞赛的误区。 认为只要拥有足够多的数据和足够大的计算力,就能复刻甚至超越人类智能。 但约翰霍普金斯大学的研究与非冯诺依曼架构的探索,让我们重新审视智能的本质。 智能并非源于复杂的数字计算,而是源于简单高效的自然逻辑。 生物神经元的物理反馈,早期物理计算机的轨迹测量,类脑架构的层级化处理,本质上都是在遵循自然及逻辑的规律。 当 AI 从模拟自然转向回归自然,从数字计算转向物理计算,其发展路径正在变得更加清晰、可持续。 这场架构革命不仅能解决当前 AI 的高能耗、数据依赖等痛点,更能为 AGI 的实现提供坚实的技术基础。 AI 的未来不在海量数据的堆砌中,而在契合智能本质的架构创新里。
修正脚本
架构革命,从类脑物理计算到 AGI,AI 正告别海量数据依赖。 当约翰霍普金斯大学的研究团队公布类脑架构 AI 无需数据训练即可媲美传统模型的实验结果时, 一场关于人工智能发展路径的深刻变革已悄然拉开序幕。 这一发现并非孤立的学术突破,而是与非冯诺依曼架构、物理计算等前沿探索形成呼应,共同指向一条摆脱海量数据桎梏、通往通用人工智能 AGI 的全新赛道。 谷歌 DeepMind 的创始人哈萨比斯敢于预言,2030年前 AGI 大概率实现,其底气或许正源于此。 当 AI 从用精密电路模拟简单逻辑,转向用物理规律复刻自然智能,曾经看似遥不可及的技术奇点正在被底层架构的革新逐步拉近。 一、核心认知颠覆,智能的本质是物理映射。 而非数据计算。 长期以来,主流 AI 的发展遵循着数据驱动的逻辑,以冯诺依曼架构为基础,通过 GPU 等精密逻辑电路, 将海量数据转化为数字浮点数,再经过复杂的量化运算模拟神经元权重。 这种模式下,AI 就像一台超级计算器,依赖千亿级参数和数百万张训练图像,才能在特定任务中达到人类级表现。 但这一路径存在着先天缺陷。 一方面,数据成为稀缺资源,许多场景下的 AI 应用因数据不足难以落地。 另一方面,能源消耗成指数级增长。 据预测,2026年 AI 数据中心耗电量将突破1000太瓦时,相当于数十个中小型国家的年用电量。 而类脑物理计算与非冯诺依曼架构的出现,彻底颠覆了这一认知。 其核心逻辑可概括为数据即计算、存储即计算,这与大自然亿万年进化出的智能机制不谋而合。 生物神经元的结构远比 GPU 简单,却能实现高效的信息处理,关键就在于其存算一体的物理特性。 神经元之间的连接强度、权重,并非通过数字模拟,而是由连接的粗细及物理层面的导线数量、信号传导阻力直接决定。 信号越频繁通过某一连接,该连接的物理结构就会随之优化,如导线变粗、电阻减小、权重自然增强。 这本质上是一种基于物理反馈的条件反射。 这种逻辑在早期物理计算机中已有雏形。 美国最早的弹道计算机无需复杂运算, 而是利用抛物线的物理特性,直接通过测量物理轨迹就能求解二次方程。 类脑物理计算正是延续了这一思路,它不追求用高级电路模拟低级逻辑,而是让物理过程本身成为计算载体。 输入信号是什么样,经过物理结构的映射后,输出信号就呈现对应的规律。 训练与推理本质上是同一套物理映射过程。 训练时的信号反馈优化了物理结构,推理时相同信号输入就会得到一致的输出结果。 这种自映射机制无需复杂的浮点数运算,也不需要海量数据的反复投喂,却能实现高效低耗的智能响应。 恰是大自然的智慧,最简单的模仿往往是最本质的复刻。 二、学术前沿印证,架构优先而非数据优先。 约翰霍普金斯大学的实验为这一逻辑提供了关键学术支撑。 研究团队对比了 Transformer、全连接网络、卷积神经网络CNN 三种主流架构,发现只有 CNN 在调整规模时,其活动模式会显著接近人脑视觉系统。 而 Transformer 的全局注意力机制、全连接网络的密集连接方式, 因缺乏生物学对应性,即便扩大神经元数量,也难以贴近人脑活动。 这一结果背后的核心原因是,CNN 模仿了大脑视觉系统的层级化信息处理模式。 通过局部连接和权重共享编码空间信息,天然具备类脑物理逻辑的雏形。 事实上,全球已有多个顶尖实验室和科研组织投身于这一方向的探索。 除了约翰霍普金斯大学,斯坦福大学、麻省理工学院等机构正聚焦于忆阻器、光信号传导等物理载体的研发。 忆阻器能够直接通过电阻变化存储和处理信息,完美契合存算一体的需求。 光信号则可通过光路的强弱粗细调节权重,且能耗远低于电子电路。 在工业界,IBM 的真北芯片、英特尔的神经拟态计算项目,均在尝试摆脱冯诺依曼架构的束缚,构建基于物理规律的类脑计算系统。 这些探索共同指向一个结论,AI 的突破点并非在于收集更多数据,堆砌更多参数,而在于找到契合智能本质的架构蓝图。 当前的学术研究已逐步厘清传统架构与类脑物理架构的核心差异。 传统架构是数字模拟物理,用精密计算复刻简单逻辑。 如同用法拉利当拖拉机,既造成资源浪费,又难以突破效率瓶颈。 而类脑物理架构是物理本身即逻辑,用简单的物理反馈机制实现智能。 既符合大自然的进化逻辑,又能解决高能耗、数据依赖等痛点。 这种差异也解释了为何传统 Transformer 架构难以支撑 AGI 的实现,即便其参数规模突破万亿级,本质上仍是在数字世界中模拟智能,无法像人脑一样实现高效、通用、低耗的信息处理。 而类脑物理架构则跳出了这一框架,直接在物理世界中构建智能,为 AGI 提供了更具可行性的技术路径。 三、 AGI 的路线图,从架构革命到普及落地,哈萨比斯关于2030年前 AGI 大概率实现的预言,若脱离类脑物理计算与非冯诺依曼架构的发展背景,难免显得乐观。 但结合当前的技术前沿来看,这一预言更像是对架构革命趋势的精准判断。 传统 AGI 路径之所以进展缓慢,核心在于陷入了数据参数能耗的恶性循环。 为实现更通用的智能,需要更多数据训练,更多数据需要更大的参数规模承载,更大的参数规模则导致能耗成指数级增长。 这种模式不仅成本高昂,且难以实现真正的通用。 毕竟现实世界中的许多场景既无法提供海量数据,也无法支撑高能耗的计算需求,而类脑物理架构的发展,正打破这一恶性循环。 其核心优势在于通用性与低耗性的统一,基于物理反馈的智能机制,天然具备适应不同场景的能力,无需为特定任务单独训练。 同时,物理计算的能耗远低于数字计算,能够实现轻量级的智能部署。 这意味着未来的 AGI 无需依赖千亿级投资的超级数据中心或许可以像人脑一样,在小型设备中实现高效运转。 从技术路线来看, AGI 的实现或将遵循三步走。 第一步,基于忆阻器、光信号等载体,构建小规模类脑物理计算原型,验证存算一体的可行性。 第二步,优化物理结构的可塑性与扩展性,实现多任务、跨场景的通用智能。第三步,降低制造成本,推动类脑计算设备的普及,让 AGI 融入日常生活。 这一路线图的关键在于突破物理载体的技术瓶颈。 当前,忆阻器的稳定性、光信号的调控精度等问题仍需进一步解决。 但随着材料科学与神经科学的交叉融合,这些障碍正在逐步被克服。 更重要的是,这一路线符合技术发展的性价比原则。 与传统架构需要持续投入海量资源不同,类脑物理架构的核心创新在于思路转变,一旦突破关键技术,其规模化应用的成本将远低于传统 AI 这也意味着 AGI 的实现不会是少数巨头的专属游戏,而可能成为全行业共同推进的技术革命。 人工智能的发展曾一度陷入数据崇拜与参数竞赛的误区。 认为只要拥有足够多的数据和足够大的计算力,就能复刻甚至超越人类智能。 但约翰霍普金斯大学的研究与非冯诺依曼架构的探索,让我们重新审视智能的本质。 智能并非源于复杂的数字计算,而是源于简单高效的自然逻辑。 生物神经元的物理反馈,早期物理计算机的轨迹测量,类脑架构的层级化处理,本质上都是在遵循自然即逻辑的规律。 当 AI 从模拟自然转向回归自然,从数字计算转向物理计算,其发展路径正在变得更加清晰、可持续。 这场架构革命不仅能解决当前 AI 的高能耗、数据依赖等痛点,更能为 AGI 的实现提供坚实的技术基础。 AI 的未来不在海量数据的堆砌中,而在契合智能本质的架构创新里。
英文翻译
Architectural Revolution: From Brain-Inspired Physical Computing to AGI, AI Is Shedding Its Dependence on Massive Data. When the research team at Johns Hopkins University published experimental results showing that brain-inspired AI architectures can rival traditional models without requiring data training, a profound transformation in the path of artificial intelligence development had quietly begun. This discovery is not an isolated academic breakthrough; it resonates with cutting-edge explorations such as non-von Neumann architectures and physical computing, collectively pointing to a new track toward AGI that breaks free from the constraints of massive data. Demis Hassabis, co-founder of Google DeepMind, boldly predicts that AGI will likely be achieved before 2030—a confidence that may well stem from this very progress. As AI shifts from using precision circuits to simulate simple logic to replicating natural intelligence through physical laws, the once-distant technological singularity is being steadily brought closer by innovations in underlying architectures. I. Core Cognitive Shift: Intelligence Is a Physical Mapping, Not Data Computation. For a long time, mainstream AI development has followed a data-driven logic, based on the von Neumann architecture, using precision logic circuits like GPUs to convert massive data into digital floating-point numbers, and then simulating neural weights through complex quantitative operations. In this model, AI acts like a super calculator, relying on hundreds of billions of parameters and millions of training images to achieve human-level performance on specific tasks. But this path has inherent flaws. On one hand, data becomes a scarce resource, and AI applications in many scenarios struggle to materialize due to insufficient data. On the other hand, energy consumption grows exponentially. It is predicted that by 2026, the power consumption of AI data centers will exceed 1,000 terawatt-hours, equivalent to the annual electricity usage of dozens of small to medium-sized countries. The emergence of brain-inspired physical computing and non-von Neumann architectures has completely overturned this paradigm. Its core logic can be summarized as "data is computation, storage is computation," which aligns seamlessly with the intelligent mechanisms evolved by nature over billions of years. The structure of biological neurons is far simpler than that of GPUs, yet they achieve efficient information processing. The key lies in their physical property of integrating storage and computation. The strength and weight of connections between neurons are not determined through digital simulation but directly by the physical thickness of connections, the number of wires at the physical level, and the resistance to signal transmission. The more frequently a signal passes through a connection, the more the physical structure of that connection optimizes—wires become thicker, resistance decreases, and the weight naturally increases. This is essentially a conditioned reflex based on physical feedback. This logic has antecedents in early physical computers. The earliest ballistic computer in the United States did not require complex calculations; instead, it leveraged the physical properties of parabolas to solve quadratic equations by directly measuring physical trajectories. Brain-inspired physical computing continues this line of thought: it does not seek to simulate low-level logic with advanced circuits but allows the physical process itself to serve as the computing medium. Whatever the input signal is, after being mapped through the physical structure, the output signal exhibits corresponding patterns. Training and inference are essentially the same physical mapping process. Signal feedback during training optimizes the physical structure, and when the same signal is input during inference, the output remains consistent. This self-mapping mechanism requires no complex floating-point operations or repeated feeding of massive data, yet it achieves efficient, low-power intelligent responses. This is the wisdom of nature: the simplest imitation is often the most essential replication. II. Academic Frontier Validation: Architecture First, Not Data First. The experiment at Johns Hopkins University provides key academic support for this logic. The research team compared three mainstream architectures—Transformer, fully connected networks, and convolutional neural networks (CNN)—and found that only CNN, when its scale was adjusted, showed activity patterns significantly close to the human visual system. In contrast, the global attention mechanism of Transformer and the dense connections of fully connected networks, due to their lack of biological correspondence, struggled to approach human brain activity even with increased neuron numbers. The core reason behind this result is that CNN mimics the hierarchical information processing mode of the brain's visual system. By encoding spatial information through local connections and weight sharing, it naturally has the prototype of brain-inspired physical logic. In fact, multiple top-tier laboratories and research organizations worldwide are actively exploring this direction. Besides Johns Hopkins University, institutions like Stanford University and MIT are focusing on the development of physical carriers such as memristors and optical signal transmission. Memristors can store and process information directly through changes in resistance, perfectly meeting the need for integrated storage and computation. Optical signals can adjust weights through the intensity and thickness of optical paths, with energy consumption far lower than electronic circuits. In the industrial sector, IBM's TrueNorth chip and Intel's neuromorphic computing projects are both attempting to break free from the constraints of the von Neumann architecture and build brain-inspired computing systems based on physical laws. These explorations converge on a common conclusion: the breakthrough point for AI lies not in collecting more data or stacking more parameters, but in finding an architectural blueprint that aligns with the essence of intelligence. Current academic research is gradually clarifying the core differences between traditional architectures and brain-inspired physical architectures. Traditional architectures simulate the physical world with digital computation, using precision calculations to replicate simple logic. This is like using a Ferrari as a tractor—it wastes resources and struggles to break through efficiency bottlenecks. Brain-inspired physical architectures, on the other hand, treat physics itself as logic, using simple physical feedback mechanisms to achieve intelligence. This aligns with nature's evolutionary logic and solves pain points such as high energy consumption and data dependence. This distinction also explains why traditional Transformer architectures struggle to support the realization of AGI: even if their parameter scale exceeds trillions, they still simulate intelligence in the digital world and cannot achieve efficient, general, and low-power information processing like the human brain. Brain-inspired physical architectures, by contrast, leap out of this framework, building intelligence directly in the physical world, offering a more feasible technical path toward AGI. III. The Roadmap to AGI: From Architectural Revolution to Widespread Implementation. Hassabis's prediction that AGI will likely be achieved before 2030 would seem overly optimistic if detached from the development context of brain-inspired physical computing and non-von Neumann architectures. But in light of current technological frontiers, this prediction appears more like an accurate judgment of the trend of architectural revolution. The traditional path to AGI has been slow primarily because it is trapped in a vicious cycle of data, parameters, and energy consumption. To achieve more general intelligence, more data is needed for training; more data requires larger parameter scales to carry it; and larger parameter scales lead to exponential growth in energy consumption. This model is not only costly but also struggles to achieve true generality—after all, many real-world scenarios can neither provide massive data nor support high-energy computing demands. The development of brain-inspired physical architectures is precisely breaking this vicious cycle. Their core advantage lies in the unity of generality and low power consumption: the intelligent mechanism based on physical feedback naturally possesses the ability to adapt to different scenarios without requiring separate training for specific tasks. Meanwhile, the energy consumption of physical computing is far lower than that of digital computing, enabling lightweight intelligent deployment. This means that future AGI may not need super data centers costing hundreds of billions of dollars; it could operate efficiently in small devices, much like the human brain. From a technical roadmap perspective, the realization of AGI may follow three steps. First, build small-scale brain-inspired physical computing prototypes based on carriers such as memristors and optical signals to verify the feasibility of integrated storage and computation. Second, optimize the plasticity and scalability of physical structures to achieve multi-task, cross-domain general intelligence. Third, reduce manufacturing costs and promote the widespread adoption of brain-inspired computing devices, integrating AGI into daily life. The key to this roadmap lies in overcoming the technical bottlenecks of physical carriers. Currently, issues such as the stability of memristors and the precision of optical signal regulation still require further resolution. However, with the cross-disciplinary integration of materials science and neuroscience, these obstacles are gradually being overcome. More importantly, this roadmap aligns with the principle of cost-effectiveness in technological development. Unlike traditional architectures that require continuous massive resource investment, the core innovation of brain-inspired physical architectures lies in a shift in thinking: once key technologies are broken through, the cost of large-scale application will be far lower than that of traditional AI. This also means that the realization of AGI will not be an exclusive game for a few giants but may become a technological revolution jointly advanced by the entire industry. The development of artificial intelligence once fell into the misconceptions of data worship and parameter competition—believing that with enough data and enough computing power, human intelligence could be replicated or even surpassed. But the research from Johns Hopkins University and the exploration of non-von Neumann architectures compel us to reexamine the essence of intelligence. Intelligence does not stem from complex digital calculations but from simple and efficient natural logic. The physical feedback of biological neurons, the trajectory measurement of early physical computers, and the hierarchical processing of brain-inspired architectures all essentially follow the principle that nature itself is logic. As AI shifts from simulating nature to returning to nature, from digital computation to physical computation, its development path is becoming clearer and more sustainable. This architectural revolution can not only solve the current pain points of high energy consumption and data dependence in AI but also provide a solid technical foundation for the realization of AGI. The future of AI lies not in the accumulation of massive data, but in architectural innovation that aligns with the essence of intelligence.
back to top