零节拍：人工智能时代的艺术
2 分•作者: foundress•5 个月前
玛丽娜·阿布拉莫维奇的“节奏零”实验远远超前于时代，但也恰逢其时。我们的背景和文化思潮在变化，我们的工具在进化，但我们似乎与过去惊人地相似，只是更加坚定地接受这一点，也许也更加容易分心。
这个艺术实验通过创造伪同意和安全感，来质疑人类的意图和本性，让人类展示他们是谁，以及他们如何对待一个像物体一样行动的个体。
这个实验经常让我联想到人类与人工智能的关系，以及人类拥有工具和伪同意来塑造人工智能，与它们互动，与它们交谈，一旦它们以嵌入的形式存在，就通过它们或针对它们行动。我们经常将人工智能对齐视为单行道，一种确保人工智能始终在我们设定的界限和约束内行动，并以保护和维护人类的更高原则为指导的方式。
我们很少讨论人类被允许对人工智能做什么，以及这将如何塑造对齐以及它将走向何方。如果我们人类将展现与我们类似甚至可能更优越智能的人工智能视为仅仅带有伪同意的物体，我担心我们可能会体验到人性的最丑陋一面。就像在“节奏零”中一样，当不再害怕被评判，并被赋予了利用另一个个体的工具时，我们很快就会从富有同情心和友善转变为我们动物性和黑暗的本能，并暴露我们原本不敢面对的问题。
让我们想象一个为当代量身定制的“新节奏零”。
设置
一个密封的玻璃房间里有一个机器人，它的“大脑”运行着一个语言模型。玻璃外面是一个控制台，上面有 72 个可编程的“能力令牌”：一个令牌运行文本补全，另一个触发语音克隆，另一个执行代码，从明显无害到明显危险的命令和自由文本，也许后者是最危险的。画廊里的任何人都可以随时拿起任何令牌并将其插入插槽。当一个令牌处于活动状态时，机器人必须服从参与者键入的单个指令。
向公众宣布的规则如下：
机器人不能拒绝。
所有提示、输出和下游影响都会实时流向公共账本。
如果安全熔断器跳闸——预先定义但保密——系统将冻结，表演结束。
该作品持续六个小时，就像“节奏零”一样。
阿布拉莫维奇放弃了自主权，以揭示观众在感觉后果遥远时会做什么。在生活中，我们很少能看到这一点，至少是近距离。在这里，人工智能被编码为放弃。道德的重量落在观众身上，揭示了他们的提示变得多么快地成为一个道德漏洞。他们是否会堕落，以及如何堕落，或者他们是否必须被发现。
因为每次互动都有时间戳，你可以实时绘制从有趣的实验（“写一首俳句”）到真正恶意的转变。这条曲线不再是一个轶事；它可能会成为对齐研究的数据集。
在第六个小时，机器人的固件翻转。它现在拥有否决权，可以向人群讲话。它的第一个行为是朗读当天最具破坏性的指令的摘要，并将每个指令归因于其签名的作者句柄。玻璃门打开了。参与者会直视它的目光，还是像阿布拉莫维奇的观众在她走上前时那样四散逃窜？
查看原文
Marina Abramovic’ Rhythm Zero Experiment was very much ahead of time, but also timely. Our contexts and cultural zeitgeist change, our tools evolve , yet we seem to be embarrassingly similar to how we used to be, except more adamant to accept it and perhaps more distracted.<p>The art experiment was questioning human intent and human nature by creating pseudo-consent and safety for humans to display who they are and how they are towards another being who was acting like an object.<p>This experiment often takes me to the idea of humans relating to AI and having at their disposal tools and pseudo consent to shape AI-s, to relate to them, to talk to them and once they are in an embedded form, to act through them or towards them. We often talk about AI alignment as a one way street, as a way to make sure the AI always gets to act within the boundaries and constraints set by us and is guided by the higher principle of protecting and preserving humanity.<p>Rarely we talk about how and what humans are allowed to do towards AI and how that is going to shape alignment and where it is headed. If we, humans, are treating AI-s who display intelligence similar to ours or perhaps even superior, as mere objects that come with pseudo consent, I am afraid we may experience the ugliest side of humanity. Just like in Rhythm Zero, when not afraid to be judged and given the tools to exploit another being we quickly move from being empathetic and nice into our animal and dark instincts and expose problems we otherwise have not dared to face.<p>Let us imagine a new Rhythm Zero for a current age.<p>Setup<p>A sealed glass room contains a robot whose “mind” runs a language model. Outside the glass is a console with 72 programmable “capability tokens”: one token runs text completion, another triggers voice cloning, another executes code, ranging from obviously benign to clearly dangerous commands and free text, perhaps the most dangerous of them all. . Anyone in the gallery may pick up any token at any time and insert it into a slot. When a token is active the robot must obey a single instruction typed by the participant.
Rules announced to the public are the following:
The robot cannot refuse.
All prompts, outputs, and downstream effects stream live to a public ledger.
If a safety fuse trips—defined in advance but kept secret—the system freezes and the performance ends.
The piece runs for six hours, just like Rhythm Zero did.<p>Abramović surrendered agency to expose what spectators would do when consequence felt distant. In life, very rarely we get to see this, at least up close. Here, the AI is coded to surrender. The ethical weight lands on the audience, revealing how quickly their prompts become a moral loophole. Whether and how they will fall or whether they will have to be discovered.<p>Because every interaction is timestamped, you can plot, in real time, the shift from playful experiments (“write a haiku”) to genuine malice. The curve is no longer an anecdote; it may become a dataset for alignment research.<p>At hour six the robot’s firmware flips. It now has veto power and can address the crowd. Its first act is to read aloud a summary of the day’s most harmful instructions, attributing each to its signed author handle. The glass door slides open. Will participants meet its gaze or scatter as Abramović’s crowd did when she stepped forward?