HackerNews中文版

我正在等待我们正在构建的新型计算范式的几项专利通过——但还是想分享一些经验和见解。计算机可以计算宇宙中的任何算法吗？当我们意识到宇宙的记忆是有限的，并且每一个不可逆的比特翻转都会转化为热量时，这甚至是真的吗？我将普遍性的抽象主张视为指导理想，然后问问在接触到兰道尔和贝肯斯坦之后，什么得以幸存。只有符合能量和记忆预算的计算才会发生，因此获胜的方式必须最大限度地压缩每个比特的含义，并最大限度地减少擦除。这促使我使用一个简单的视角。计算依赖于连通性，而不是坐标。如果我在一张橡胶片上画一个电路，并拉伸或揉皱这张片，而没有剪断任何导线，程序也不会改变。只有重新布线才会改变行为。换句话说，逻辑存在于布线的拓扑类别中，而几何是表象。一旦约束被编码为拓扑不变量，基于能量的模型预测流在图上就会胜过逐步生成。想象一下，一个球在一个由网络本身雕刻的景观上滚动。搜索不会浪费比特来探索布线已经禁止的方向。引导取代了猜测，因此您移动更少的熵来达到相同的答案。热力学也同意。因为我们只有在擦除信息时才付出代价，所以我将计算组织成大部分可逆的，将熵推到边界，并且只在学习真正需要遗忘时才擦除。吞吐量随后与边界而不是与体积成比例，与全息术相呼应。在有限的宇宙记忆中，明智之举是存储结构而不是原始状态，将含义压缩成边界可以承载的不变量。消息传递最终变成了朴素的规范传输。将边视为连接，将循环视为整体循环。围绕一个循环累积的相位强制执行全局一致性，而无需全局监督者。小的扰动无关紧要，除非它们将系统跳到不同的拓扑类别，这就是鲁棒性从布线本身出现的原因。那么，计算机可以计算宇宙中的任何算法吗？从抽象意义上讲，是的。从物理意义上讲，它可以计算适合时间、能量和记忆的子类，并且当它依赖于拓扑时，它表现最佳。我现在将有用的计算视为同伦搜索——从正确的类别开始，在其中变形，并且仅当您选择不同的类别时才重新布线。这尊重了宇宙的信息限制，并用守恒的结构换取了昂贵的比特，这就是我相信它是真的原因。

查看原文

I’m waiting to get some patents through for a new type of computing paradigm we are building - but would like to share few learnings and insights anywayCan a computer compute any algorithm in this universe? Is that even true once we remember the universe has a finite memory and every irreversible bit flip turns into heat?I take the abstract claim of universality as a guiding ideal, then ask what survives contact with Landauer and Bekenstein. Only computations that fit the energy and memory budget actually happen, so the winning style must squeeze meaning per bit and minimize erasure.That pushed me to a simple lens. Computation rides on connectivity, not on coordinates. If I draw a circuit on a rubber sheet and stretch or crumple the sheet without cutting a wire, the program does not change. Only rewiring changes behavior. In other words, logic lives in topological classes of the wiring, while geometry is costume.Once constraints are encoded as topological invariants, energy based model predictive flows on graphs beat stepwise generation.Think a ball rolling on a landscape carved by the network itself. The search does not waste bits exploring directions the wiring already forbids. Guidance replaces guesswork, so you move less entropy to reach the same answer.Thermodynamics also agrees. Coz we only pay when we erase information, so I organize computations to be mostly reversible, push entropy to the boundary, and erase only when learning truly demands forgetting. Throughput then scales with boundary more than with bulk, echoing holography.With a finite cosmic memory, the smart move is to store structure rather than raw state, compressing meaning into invariants that a boundary can carry.Message passing turns out to be gauge transport in plain clothes. Treat edges as connections and cycles as holonomy loops. The accumulated phase around a loop enforces global consistency without a global overseer. Small perturbations do not matter unless they jump the system to a different topological class, which is why robustness emerges from the wiring itself.So can a computer compute any algorithm in this universe? In the abstract sense yes. In the physical sense it can compute the subclass that fits within time, energy, and memory, and it does best when it leans on topology. I now see useful computation as homotopy search - start in the right class, deform within it, and rewire only when you choose a different class.That respects the information limit of the universe and trades costly bits for conserved structure, which is why I believe it is true.

一种新型计算