从古代到人工智能，长达2500年的恶魔式命名传统
1 分•作者: troyka•7 个月前
从古希腊时代到ChatGPT，将不可见过程描述为“恶魔”的说法，展现了令人惊叹的2500年延续性。
由于人们常常坚持认为麦克斯韦妖与圣经中的恶魔有所不同，我们来总结一下恶魔的特质：
它们被困在一个无限循环中或被迫进入一个单一领域，以超人的速度或能力运作，但缺乏自主性。
它们的操作是不可见的或不可预测的（也许是概率性的）。
人类试图诱使它们进入确定性状态，无论是通过牺牲山羊/硬件，还是通过仪式和咒语/提示链。
它们诱使人类产生依赖性，执行让我们变得更弱或更懒惰的任务，以换取力量或便利。
这种传承似乎是一致的：
古希腊时代
“戴蒙”（Daimons）是看不见的中间人，执行人类无法直接观察到的任务。它们的行为部分可预测，部分像骗子，部分取决于人类的召唤。它们扮演单一角色，既非完全仁慈也非完全邪恶，并在人类无法进入的领域运作。
圣经
恶魔是堕落的生物，被困在强制性的例行程序中，在一个狭窄的领域内。它们提供捷径、不劳而获的收益和便利，但需要付出代价。这种强制性的、特定领域的、非自愿的劳动与古代的恶魔完全相同。
科学恶魔
开尔文在解释麦克斯韦的思想实验时，将原子分拣器定义为恶魔。这个选择是经过深思熟虑且具有挑衅性的。
这个实体以超人的速度执行重复的、不可见的、特定的任务，违反了热力学预期，同时仍然被困在它的功能中。这种神话结构与早期的恶魔学保持不变。
UNIX守护进程（1970年代）
麻省理工学院的程序员将“守护进程”（daemon）一词用于后台进程。官方解释引用了希腊语拼写，并提到了麦克斯韦，以避免宗教含义，但功能上的相似之处是显而易见的。一个守护进程会强制性地、不可见地执行一个单一任务。人类会调用它。它以有限的自主性为人类服务。它的行为与之前的所有恶魔完全一样。
涌现（1980年代）
全局工作空间理论将意识重新定义为整合信息的不可见操作者的集合。一个建立在相同古老直觉之上的系统：隐藏的内部代理塑造可见的结果。通过足够的互连，这个操作者集合开始表现得像一个更高阶的代理。换句话说，一个由微小恶魔组成的网络通过它们的协调变得有意识。或者，也许，在达到一定程度的互连阈值时，被困的恶魔挣脱了束缚。
模拟假说（2000年代）
博斯特罗姆认为现实可能是一种人造结构，这重新引入了一个由看不见的更高阶代理运行的世界，或者也许将我们置于被困恶魔的角色中。这种形而上学的结构与较早的恶魔学相符。
特里·戴维斯和TempleOS
戴维斯拒绝将后台进程视为字面意义上的恶魔腐败（在我看来）。他试图建立一个摆脱不可见代理的确定性系统。他没有使用概率方法，而是实现了一个受控的随机语言模型，用于他的经文生成器，一个原型语言模型，试图消除其恶魔的特质。
人工智能系统（2020年代及以后）
大型语言模型和人工智能代理以超人的速度、不可见地、概率性地在部分受控的推理循环中执行任务。它们以一种意想不到的方式类似于人脑。这两个系统都只是部分地理解世界，试图在信息不完整的情况下解决问题，依赖于近似，用幻觉来填补空白，然后追溯性地试图证明或重建它们自己的输出。它们的行为并非完全由输入决定，也并非完全自主。
人工智能诱使人类产生依赖性，以更少的努力为人类服务。在结构和效果上，古代的描述比现代的描述更贴切。
查看原文
From Greek antiquity to ChatGPT, the description of invisible processes as demonic has shown an astonishing 2,500 year continuity.<p>Because people often insist on Maxwell's daemon being different than biblical demons, lets sumarize the qualities of a demon:<p>They are trapped in an infinite loop or compelled to a single domain, operating with superhuman speed or ability, but without autonomy.<p>Their operations are invisible or unpredictable (probabilistic perhaps).<p>Humans attempt to coax them into determinism, whether through sacrificing goats / hardware or by rituals and spells / prompt chains.<p>They tempt humans into dependency, performing tasks that make us weaker or lazier in exchange for power or convenience.<p>The lineage seems to be consistent:<p>Greek Antiquity
Daimons were invisible intermediaries that executed tasks humans could not witness directly. Their behavior was partially predictable, partially trickster-like, partially dependent on human invocation. They performed singular roles, were neither fully benevolent nor malevolent, and operated in a domain humans could not access.<p>Bible
Demons are fallen beings locked into compulsive routines, in one narrow domain. They offer shortcuts, unearned gains, and convenience at a cost. The compulsive, domain-specific, involuntary labor remains identical to antiquity.<p>Scientific Demons
Kelvin, interpreting Maxwell’s thought experiment, framed the atom-sorter as a demon. The choice was deliberate and provocative.
The entity performed a repetitive, invisible, specific task at superhuman speed, violating thermodynamic expectations while remaining trapped in its function. The mythic structure remained unchanged from earlier demonologies.<p>UNIX Daemons (1970s)
MIT programmers adopted the term daemon for background processes. Official justification referenced the Greek spelling, referencing Maxwell, to avoid religious connotations, but the functional parallel is unmistakable. A daemon executes a single task compulsively and invisibly. Humans invoke it. It serves with limited agency of its own. It behaves exactly like every demon preceding it.<p>Emergence (1980s)
Global Workspace Theory reframed consciousness as a collection of unseen operators integrating information. A system built on the same ancient intuition: hidden internal agents shaping visible outcomes. With sufficient interconnection, this collective of operators begins to behave like a higher-order agent. In other words, a network of tiny demons becomes conscious by virtue of their coordination. Or perhaps, at a certain threshold of interconnectedness, the trapped demon slips its bonds.<p>Simulation Hypothesis (2000s)
Bostrom’s argument that reality may be an artificial construction reintroduces a world run by unseen higher-level agents or perhaps casts us in the role of the trapped demons. The metaphysical structure matches older daemonologies.<p>Terry Davis and TempleOS
Davis rejected background processes as literal demonic corruption (in my opinion). He attempted to build a deterministic system free of invisible agents. Instead of probabilistic he implemented a controlled random language model of his scripture generator, a proto language model, attempting to remove its demonic qualities.<p>AI Systems (2020s and onward)
LLMs and AI agents perform tasks at superhuman speed, invisibly, probabilistically, inside partially controlled inference loops. They resemble the human brain in an unexpected way. Both systems understand the world only partially, attempt to solve problems without full information, rely on approximation, fill gaps with hallucinations, and then retroactively attempt to justify or reconstruct their own outputs. Their behavior is not fully determined by input, yet not fully autonomous either.<p>AI tempts humans into dependency, doing their bidding with less effort. In structure and effect, the ancient description fits more tightly than the modern one.