内容来源：https://www.quantamagazine.org/thermodynamic-computers-go-with-the-energy-flow-20260715/

内容总结：

热力学计算：让“噪音”成为计算新动能

在传统计算机领域，热噪声一直是工程师们极力避免的干扰因素——原子随机热振动可能导致二进制位翻转，使计算偏离正轨。然而，一个名为“热力学计算”的新兴研究领域正颠覆这一认知：与其对抗噪声，不如利用噪声进行计算。

自2019年美国计算社区联盟首次举办热力学计算研讨会以来，一小批研究人员致力于将这一理念变为现实。近期，他们已在标准硅基逻辑电路中成功模拟热力学计算，验证了这一概念在原理上的可行性。该方法有望大幅降低计算机功耗与散热，为当前能耗惊人、散热困难的超密集微型化芯片技术提供新出路。

热力学计算的核心，是借助热力学过程——即能量分布与耗散过程中不可避免产生的微观随机性——作为计算资源。纽约初创公司Normal Computing的物理学家Patrick Coles指出：“这一领域旨在设计利用热力学作为计算资源的计算机。”若成功，它不仅将变革计算产业，更可能重塑人类对计算本质的认知。

目前主流技术路径分为两类：一是“平衡态热力学计算”，让系统在能量景观中自然演化至能量最低点，对应问题的最优解；二是“非平衡态热力学计算”，系统在持续能量输入下不断运动，其运动轨迹本身即编码计算结果。后者因无需等待系统自然平衡，被认为速度更快。

在实践层面，Normal Computing去年展示了使用模拟电路实现矩阵求逆运算。其定制电路板包含8个RLC谐振器集群，通过噪声驱动完成计算。研究显示，若扩大网络规模，其处理速度最终将超越传统数字神经网络，且能耗与发热显著降低。该公司随后发布采用硅芯片数字处理技术的新型热力学计算机CN101，可执行图像生成、分子模拟等任务，但目前尚未经同行评审。

另一起初创公司Extropic于2025年10月推出号称“全球首款可扩展概率计算机”的芯片，集成数千个半导体组件，据称能以现有算法万分之一能耗运行生成式AI。

理论模拟层面，劳伦斯伯克利国家实验室物理学家Stephen Whitelam近期展示了非平衡态热力学计算电路模拟：其算法能从随机噪声中重建图像，且运行过程散热量仅为数字神经网络的约千亿分之一。该方法与生成式AI训练思路相似，有望成为热力学计算的重要应用方向。

尽管热力学计算尚处早期阶段，但研究人员认为其技术门槛远低于需要极低温冷却和脆弱的量子纠缠态的量子计算机，更可能成为近期的可行替代方案。不过，当前热力学计算电路的能力仅相当于1990年代的小型数字神经网络，能否通过更大规模电路和训练实现突破，仍需时间验证。

中文翻译：

热力学计算机：随（能）流而动

引言

在追求计算机精确性与可靠性的过程中，噪声是最大的敌人。原子的热抖动时刻威胁着精密计算所需的准确性。无论是我们如今使用的笔记本电脑或超级计算机这类常见的经典设备，还是未来有望实现更快计算的量子装置，我们都不希望某次随机的热波动将二进制数字从 1 翻转为 0，从而导致计算偏离轨道。因此，计算机工程师们通过以远高于环境随机波动的能量来切换比特，竭力使计算机免受噪声干扰。

但假如噪声能成为计算机工程师的朋友呢？假如我们不再试图制造那些能在宇宙万物皆受干扰的热波动喧嚣中正常工作的设备，而是驾驭这种噪声来实际执行计算呢？

这便是热力学计算这一新兴领域的目标。自 2019 年计算社区联盟首次举办热力学计算会议以来，一小群研究人员一直在努力将其付诸实践。最近，其中一些人在标准的硅基逻辑电路中模拟了热力学计算，表明其基本概念在原理上似乎可行。

这种方法可能制造出功耗低、散热量小的计算机——鉴于当今设备的高能耗操作以及防止超高密度微型电路熔化的日益艰巨挑战，这无疑是一大优势。

热力学计算将利用热力学过程，这些过程会分布和耗散能量，并在微观尺度上不可避免地增加随机性。“这个领域致力于设计将热力学作为计算资源加以利用的计算机，”纽约初创公司 Normal Computing 的物理学家帕特里克·科尔斯说道。如果这一构想得以实现，它不仅可能改变计算机产业，还可能改变我们对计算本身的思考方式。

穿越能量景观的路径

热力学第二定律告诉我们，封闭系统的熵会随时间增加；整体上事物会变得更加无序。这意味着能量会作为随机的热波动被耗散，而这些波动通常对任何人都毫无用处。然而，自然界中的某些过程正是利用这些波动来达到更有序的状态。

“我认为热力学计算的发展源于这样一种想法：它‘搭便车’利用了‘早已存在于’世界其他地方的、但并未被明确标注为计算的计算过程，”加拿大本拿比西蒙弗雷泽大学的统计物理学家大卫·西瓦克表示。

在热力学中，物理系统随时间变化的方式（其动力学）可以被描述为穿越“能量景观”的一条轨迹，这是一幅关于系统组件不同构型总能量的图谱。在这幅景观中，山谷是系统稳定在低能构型的地方，而山峰则是相对不稳定的高能构型。当这样的系统稳定在最低的山谷并保持在那里时，就达到了平衡。

如果这一切听起来有点抽象，想想牛奶吧。许多人能够消化乳制品，得益于一种名为乳糖酶的链状酶。这种酶由小肠产生，像拼图碎片一样与牛奶中名为乳糖的复杂糖类契合。它们形状的兼容性使乳糖酶能够将乳糖分解为更简单的糖成分，从而更易于消化系统处理。

但乳糖酶是如何获得这种有用形状的呢？它是由构成它的氨基酸序列编码的。当这条链在细胞中合成时，热波动使其能够探索自身的能量景观，并折叠成最稳定的平衡状态。

这个过程是一种热力学计算，因为它仅利用细胞周围环境中的热能，就解决了将乳糖酶折叠成正确形状的问题。一旦氨基酸链正确折叠，持续的热波动仅会引起一些随机抖动。新形成的酶稳定地处于一个深的能量井中，从而保持其形状。因此，尽管像活细胞这样的自然系统的运行发生在充满热噪声的环境中，它仍然能够工作，并且往往具有显著的能量效率。

平衡态与非平衡态

一种名为平衡态热力学计算的热力学计算方法，其工作原理类似于蛋白质折叠。如果你将想要解决的问题编码到一个系统中，热力学可以驱动系统穿越其能量景观，朝向与该问题最优解相对应的能量最小值前进。“你可以拿一个小型电路，让它在其热驱动动力学作用下自然演化，然后一旦它达到热力学平衡，就测量其特性，”加州劳伦斯伯克利国家实验室的统计物理学家斯蒂芬·惠特拉姆说道。

另一种不同的热力学计算方法，发生在系统被某种能量源驱动远离平衡态时，就像太阳的热量阻止地球大气层陷入某种一成不变的状态一样。在这种情况下，计算是在系统不断穿越其能量景观时发生的。轨迹本身编码了计算。这些“非平衡”过程在自然界中随处可见；生命本身便是其中之一，由物质和能量的持续流动驱动。

在这种非平衡条件下运动的物体轨迹，展现出所谓的朗之万动力学，以 20 世纪初的法国物理学家保罗·朗之万命名。朗之万动力学代表了一种松弛过程：系统寻求降低能量，并在此过程中耗散能量。但系统永远无法稳定在一个固定的平衡状态，因为随机脉冲（如热波动）不断将其推向新的路径。

惠特拉姆表示，原则上，如果电路在足够低的功率水平下运行，从而受到随机热波动的影响，朗之万动力学就可以在电路中实现。如果科学家能够安排电路中的能量运动对应于计算结果，他们就可以利用它来进行热力学计算。

惠特拉姆解释说，平衡态和非平衡态热力学计算机都从热波动中持续获得能量提升。但平衡态计算机必须达到平衡才能完成计算，而非平衡态计算机可以被设计为在设定的时间尺度内完成计算。这就是为什么非平衡态热力学计算方法很受欢迎，科尔斯说：它们“可能更快，因为你不需要等待自然的平衡过程。”

从噪声中产生结构

到目前为止，将热力学计算付诸实践的大多数努力都使用硅基电路作为模拟。去年，科尔斯和 Normal Computing 的同事展示，这样的电路可以执行各种类型的计算，包括矩阵求逆——这一数学问题在机器学习、计算机图形学、工程和金融等多个领域都有应用。

Normal 公司于 2022 年由前 Google X 和 Google Brain 成员法里斯·斯巴希、安东尼奥·马丁内斯和马蒂亚斯·坦创立。他们推出的热力学计算机是一块定制的印刷电路板，包含八个元件簇，每个簇都与所有其他簇相连，形成一个网络。每个簇包含一个简单的 RLC 谐振器——由电阻、电容和电感组成，这种组合能以特定频率产生振荡电信号。电路被输入一个随机电信号：基本上，它是被噪声驱动的。

网络中每对 RLC 谐振器之间的耦合强度可以变化，所有耦合关系可以写成数学上的矩阵形式。这基本上是一组互连弹簧的电路版本。问题是：如果你晃动它，网络的动力学将如何演化？

结果表明，如果网络被与耦合能量相当的噪声“晃动”，它经历的平衡波动对应于耦合矩阵的数学逆。“所以你可以构建你的设备，过一段时间后回来测量它的波动，你就完成了矩阵求逆，”惠特拉姆说。

科尔斯及其同事制造的电路实际上并未依靠环境噪声运行，因为环境噪声水平太低，不足以影响动力学。研究人员不得不使用随机数生成器手动添加额外的噪声，这消耗了能量。这也是这种模拟模型本身无法展示热力学计算所承诺的能量优势的原因之一。但科尔斯表示，优势最终来自于这样一个事实：一旦被噪声驱动，计算本身就会“免费”运行。Normal 团队表明，如果他们通过向网络添加更多节点来扩展处理单元，最终将达到一个临界点，即它能够比常规的数字神经网络更快地解决问题，同时消耗显著更少的能量并散发显著更少的热量。

自从公布这款原型机以来，Normal 公司宣布推出了一款名为 CN101 的新型热力学计算机，它使用硅芯片上的数字处理技术。研究人员表示，这种数字硅技术比他们的模拟电路更容易扩展，并且还可以执行其他类型的计算，例如图像生成和分子模拟。该设备尚待其他专家评估。

Normal 团队的初步原理验证激励了惠特拉姆，他最近报道了一种非平衡态热力学计算电路的模拟。

惠特拉姆使用了一个“去噪”问题的理论模型，在经典计算机上进行了模拟。他在一段保罗·朗之万的面部逐渐被噪声破坏、最终溶解成随机静态画面的视频上训练了一个算法。（与 Normal Computing 的工作一样，噪声是通过随机数生成器人为引入的。）然后，他证明该算法可以从静态画面开始，重建出朗之万的面部。

惠特拉姆使用的热力学计算机是一个由连接节点组成的网络，其状态由这些节点连接的强度决定。在训练过程中，算法逐步调整这些连接，以找到最有可能使朗之万出现的配置。

这就像是一个由弹簧连接的网络，这些弹簧具有不同且可调节的弹簧常数。事实上，“你可以用真实的弹簧来构建一台热力学计算机，”惠特拉姆说——尽管这更像是一种新奇事物，而非实用技术。

惠特拉姆表明，由此产生的动力学沿着一条耗散热量最小的路径进行——粗略比较，比用数字神经网络执行相同任务所产生的热量少约 1000 亿倍。

惠特拉姆训练其算法的方式，有点像目前人们训练生成式 AI 算法的方式——而这实际上可能是热力学计算的主要应用之一。“我的算法在两个方面具有生成性，”惠特拉姆说。“首先，它将噪声转化为结构，从而从无序中产生秩序。其次，如果你用一组图像训练它，它就能生成它之前从未见过的额外图像。”

并非只有 Normal Computing 一家初创公司押注热力学将在计算的未来中占有一席之地。总部位于波士顿的 Extropic 公司也于 2022 年由来自谷歌、IBM、苹果和微软的团队创立。2025 年 10 月，该公司宣布推出“世界上第一台可扩展的概率计算机”——一个芯片上由数千个相互连接的半导体元件组成的网格——据称，它运行生成式 AI 算法所消耗的能量比现有算法少约 10,000 倍。（该工作尚未经过同行评审。）

做自然而然之事

除了可能提供低成本、低耗散的计算之外，热力学计算还可能揭示自然复杂系统的工作方式。

分子生物学中的许多过程已经被描述为一种信息处理。例如，一个信号分子（如激素）可能到达细胞表面，其信号被“转导”——沿着相互作用的分子链传递——从而最终在细胞核中切换一个开关，激活某个特定基因。细胞似乎能够非常高效地进行这种计算，产生很少的能量耗散，并在一定程度上依赖于分子间相互作用的内在热力学特性。

那么，细胞本身就是一种热力学计算机吗？“按照我作为物理学家的思维方式，可以公平地说，自然界使用了由进化编程的朗之万计算机，”惠特拉姆说道。

驾驭热力学波动而非抑制它的核心思想“确实发人深省”，哥本哈根尼尔斯·玻尔研究所的复杂系统理论家金子邦彦表示。“但是否能有效地转化为生物学背景下的计算，仍然是一个悬而未决的问题。”

热力学计算领域仍处于早期阶段，“类似于 1990 年代建造小型量子计算机时的情况”，正如科尔斯及其在 Normal 的同事在他们 2025 年的论文中所写。如今，量子计算已是一个全球产业，估计价值约 120 亿美元。但用基于半导体的简单谐振器构建电路，比用必须保持在微妙纠缠量子态且通常需要低温冷却的量子比特要容易得多。“热力学计算缺乏技术障碍，可能使其成为比量子计算更近期的替代方案，”Normal 团队写道。

惠特拉姆将之与支撑当今 AI 的神经网络做了类似比较。“我们迄今为止（为热力学计算）提出的计算机设计，其能力仅相当于 1990 年左右小型数字神经网络的水平，”他说。AI 的历史表明，通过更大的电路和更多的训练，应该有可能做得更好，“但这还有待观察。”如果它最终取得成功，热力学计算将在不止一个方面变得“嘈杂”起来。

英文来源：

Thermodynamic Computers Go With the (Energy) Flow
Introduction
In the quest to make computers accurate and reliable, noise is the enemy. The thermal jiggling of atoms is a constant threat to the precision needed for detailed calculations. Whether we’re dealing with familiar classical devices like the laptops or supercomputers that we use today, or fancy quantum devices that promise us faster computation tomorrow, we don’t want some haphazard heat fluctuation to flip a binary digit from a 1 to a 0, sending a calculation off course. So computer engineers work hard to make computers immune to noise, by switching bits at energies far above the random ripples of the environment.
But what if noise could be made into the computer engineer’s friend? What if, instead of trying to make devices that work despite the hubbub of thermal fluctuations that ruffle everything in the universe, we could harness that noise to actually do the computing?
That’s the goal of a nascent field called thermodynamic computing. Since the Computing Community Consortium hosted its first conference on thermodynamic computing in 2019, a small community of researchers has been laboring to put it into practice. Recently, some of them have simulated thermodynamic computation in standard silicon-based logic circuits, showing that the basic concepts seem to work in principle.
The approach could produce computers that consume little power and dissipate little heat — a huge advantage, given the power-hungry operation of today’s devices and the increasing struggle to prevent ultra-dense miniaturized circuits from melting down.
Thermodynamic computing would make use of thermodynamic processes, which distribute and dissipate energy and inevitably increase randomness at the microscopic scale. “The field is about designing computers that exploit thermodynamics as a computational resource,” said Patrick Coles, a physicist at the startup Normal Computing in New York. If it works, it could transform not only the computing industry but the very way we think about computation itself.
A Path Through the Energy Landscape
The second law of thermodynamics tells us that the entropy of a closed system should increase over time; things should become less organized overall. This means that energy gets dissipated as random thermal fluctuations, which are generally of no use to anyone. Some processes in nature, however, use those fluctuations to find their way to a more organized state.
“I think thermodynamic computing was developed with the thought that it was piggybacking on the computation that ‘already happens out there’ in the rest of the world but [is] not explicitly labeled as such,” said David Sivak, a statistical physicist at Simon Fraser University in Burnaby, Canada.
In thermodynamics, the way a physical system changes over time (its dynamics) can be described as a trajectory through an “energy landscape,” a kind of map of the total energies of different configurations of a system’s components. In this landscape, valleys are where the system settles into comfortable, low-energy configurations, while peaks are high-energy configurations that are relatively unstable. A system like this reaches equilibrium when it settles in the lowest valley and stays there.
If this all seems a bit abstract, think about milk. What allows many people to digest dairy products is a chainlike enzyme called lactase, which is produced in the small intestine and fits together like a puzzle piece with the complex sugar in milk called lactose. The compatibility between their shapes allows lactase to break lactose apart into its simple sugar components, which are more easily managed by the digestive system.
But how does lactase get this useful shape? It is encoded into the sequence of amino acids that make it up. As the chain is synthesized in the cell, thermal fluctuations allow the chain to explore its energy landscape and crumple up into its most stable equilibrium state.
This process is a kind of thermodynamic computing, in that it solves the problem of folding the lactase into the correct shape using only the thermal energy in the ambient environment of the cell. Once the chain of amino acids is properly folded, continuing thermal fluctuations merely induce a bit of random wiggling. The newly formed enzyme keeps its shape as it sits stably in a deep energy well. Thus, although the operation of a natural system like a living cell proceeds in a milieu pervaded by thermal noise, it still works, often with remarkable energy efficiency.
In and Out of Equilibrium
One type of thermodynamic computing, called equilibrium thermodynamic computing, would work like a folding protein. If you encode the problem you want solved in a system, thermodynamics can drive the system through its energy landscape toward the energy minimum that corresponds to the problem’s optimal solution. “You could take a small electrical circuit, let it evolve naturally under its thermally driven dynamics, and then measure its properties once it has attained thermodynamic equilibrium,” said Stephen Whitelam, a statistical physicist at Lawrence Berkeley National Laboratory in California.
A different approach to thermodynamic computing takes place when a system is driven away from equilibrium by some source of energy, much as the sun’s heat prevents the Earth’s atmosphere from settling into some unchanging state. In this case, the computation takes place as the system moves constantly through its energy landscape. The trajectory itself encodes the calculation. These “out-of-equilibrium” processes are ubiquitous in nature; life itself is one of them, powered by a constant flow of energy and matter.
The trajectories of objects moving under such nonequilibrium conditions display a version of so-called Langevin dynamics, named after the early-20th-century French physicist Paul Langevin. Langevin dynamics represents a kind of relaxation: The system seeks to lower its energy, and it dissipates energy in the process. But the system can never settle into a fixed equilibrium state because random impulses, such as thermal fluctuations, constantly push it onto some new course.
In principle Langevin dynamics can be embodied in an electrical circuit, if it is operating at a power level low enough to be affected by random thermal fluctuations, Whitelam said. If scientists can arrange for the movement of energy through the circuit to correspond to the answer to a calculation, they can use it to do thermodynamic computing.
Both equilibrium and nonequilibrium thermodynamic computers receive continuous boosts of energy from thermal fluctuations, Whitelam explained. But an equilibrium computer must reach equilibrium to complete its calculation, while a nonequilibrium computer can be designed to complete its calculation on a set timescale. That’s why nonequilibrium approaches to thermodynamic computing are popular, Coles said: They “can potentially be faster, because you’re not waiting for natural equilibration.”
Structure From Noise
Most efforts so far to put thermodynamic computing into practice use silicon-based circuits as analogues. Last year, Coles and colleagues at Normal Computing showed that such a circuit could perform various kinds of computation, including matrix inversion, a mathematical problem that has applications in diverse fields ranging from machine learning to computer graphics, engineering, and finance.
Normal was founded in 2022 by Faris Sbahi, Antonio Martinez, and Matthias Tan, former members of Google X and Google Brain. The thermodynamic computer they unveiled was a tailor-made printed circuit board containing eight clusters of components, with each cluster connected to all the others to form a network. Each cluster contained a simple RLC resonator — a resistor, capacitor, and inductor, a combination that creates an oscillating electrical signal at a particular frequency. The circuit is fed a random electrical signal: Basically, it is driven by noise.
The strength of the coupling between each pair of RLC resonators in the network can be varied, and all the couplings can be written in the mathematical form of a matrix. It’s basically an electrical version of a set of interconnected springs. The question is: If you give it a shake, how will the dynamics of the network evolve?
It turns out that, if the network is “shaken” by noise comparable to the energy of the couplings, the equilibrium fluctuations it undergoes correspond to the mathematical inverse of the coupling matrix. “So you can build your device and come back sometime later and measure its fluctuations, and you’ve done matrix inversion,” Whitelam said.
The circuit made by Coles and colleagues didn’t actually run on ambient noise, which was too low-level to affect the dynamics. The researchers had to add in extra noise by hand, using a random-number generator, which cost energy. That’s one reason this kind of analog model doesn’t itself demonstrate the promised energy advantages of thermodynamic computing. But ultimately, Coles said, the advantage comes from the fact that the computation itself runs “for free” once driven by noise. The Normal team showed that, if they scaled up a processing unit by adding more nodes to its network, it would eventually reach a point at which it could solve problems faster than a regular digital neural network, using considerably less energy and dissipating considerably less heat.
Since unveiling this prototype, Normal has announced a new thermodynamic computer called CN101 that uses digital processing on a silicon chip. The researchers say that this digital silicon technology is easier to scale up than their analog circuit and can also carry out other kinds of computation, such as image generation and simulations of molecules. The device has yet to be assessed by other experts.
The initial proof of principle by the Normal team served as an inspiration for Whitelam, who recently reported a simulation of a nonequilibrium thermodynamic computing circuit.
Whitelam used a theoretical model of a “denoising” problem, simulated on a classical computer. He trained an algorithm on a video of Paul Langevin’s face being steadily corrupted by noise, dissolving into random static. (As with the Normal Computing work, the noise was introduced artificially by a random-number generator.) He then demonstrated that the algorithm could start with the static and reconstruct Langevin’s face.
The thermodynamic computer Whitelam used was a network of connected nodes, and its state was determined by how strongly connected those nodes were. During the training process, the algorithm adjusted those connections a little at a time to find the configuration most likely to make Langevin appear.
It’s like a network of linked springs, where those springs have different (and adjustable) spring constants. Indeed, “you could build a thermodynamic computer from real springs,” Whitelam said — though that would be a curiosity rather than a practical technology.
Whitelam showed that the resulting dynamics followed a path that dissipated the minimal amount of heat — by a rough comparison, about 100 billion times less than would be produced if the same task were performed by a digital neural network.
The way Whitelam trained his algorithm is a little like the way people currently train algorithms for generative AI — and indeed this may be one of the major applications of thermodynamic computing. “My algorithm is generative in two ways,” Whitelam said. “First, it turns noise into structure, thereby generating order from disorder. Second, if you train it on a set of images, then it can generate additional images that it hasn’t seen before.”
Normal Computing is not the only startup betting that thermodynamics will feature in the future of computing. The Boston-based company Extropic was also founded in 2022 by a team from Google, IBM, Apple, and Microsoft. In October 2025 the company announced “the world’s first scalable probabilistic computer” — a grid of thousands of interconnected semiconductor-based components on a chip — which it says can run generative AI algorithms using about 10,000 times less energy than existing algorithms. (That work has yet to be peer-reviewed.)
Doing What Comes Naturally
In addition to potentially offering low-cost, low-dissipation computing, thermodynamic computing might also offer insights into the way natural complex systems work.
Much of what goes on in molecular biology is already described as a sort of information processing. For example, a signaling molecule (like a hormone) might arrive at the surface of a cell and have its signal “transduced” — transmitted along a chain of interacting molecules — so that ultimately it flips a switch in the cell nucleus and activates a particular gene. Cells seem able to conduct this kind of computation very efficiently, creating little energy dissipation and relying, in part, on the intrinsic thermodynamics of intermolecular interactions.
So are cells themselves a kind of thermodynamic computer? “To my physicist’s way of thinking, it would be fair to say that nature uses Langevin computers programmed by evolution,” Whitelam said.
The central idea of harnessing thermodynamic fluctuations instead of suppressing them “is indeed thought-provoking,” said Kunihiko Kaneko, a complex-systems theorist at the Niels Bohr Institute in Copenhagen. “But whether this effectively translates to computing in a biological context remains an open question.”
The field of thermodynamic computing is in its early days, “analogous to when small-scale quantum computers were built in the 1990s,” as Coles and his colleagues at Normal wrote in their 2025 paper. Quantum computing is now a global industry estimated to be worth around $12 billion. But it’s a lot easier to build circuits from simple semiconductor-based resonators than from quantum bits that have to be kept in delicate entangled quantum states and often need cryogenic cooling. “The lack of technological barriers for thermodynamic computing can potentially make it a more near-term alternative to quantum computing,” the Normal team wrote.
Whitelam makes a similar comparison with the neural networks that underpin today’s AI. “The computer designs we’ve come up with so far [for thermodynamic computing] are only as capable as the small digital neural networks of around 1990,” he said. The history of AI suggests that it should be possible to do better with larger circuits and more training, he said, “but that remains to be seen.” If it pays off, thermodynamic computing will get noisy in more ways than one.