一个长达2500年的命名传统，从古代到人工智能的恶魔般命名法则
从希腊古代到ChatGPT，将看不见的过程描述为恶魔的观念展现了惊人的2500年连续性。
由于人们常常坚持认为麦克斯韦的恶魔与圣经中的恶魔不同，我们来总结一下恶魔的特征：
它们被困在一个无限循环中，或被迫在一个特定领域内运作，以超人类的速度或能力行动，但没有自主性。
它们的操作是看不见的或不可预测的（或许是概率性的）。
人类试图诱使它们进入确定性，无论是通过牺牲山羊/硬件，还是通过仪式和咒语/提示链。
它们诱使人类依赖，执行使我们变得更弱或更懒惰的任务，以换取力量或便利。
这种血统似乎是一致的：
希腊古代
“达伊蒙”是看不见的中介，执行人类无法直接见证的任务。它们的行为部分可预测，部分像是恶作剧者，部分依赖于人类的召唤。它们执行单一角色，既不完全仁慈也不完全恶意，运作在一个人类无法接触的领域。
圣经
恶魔是堕落的存在，被锁定在强迫性的例行公事中，处于一个狭窄的领域。它们提供捷径、不劳而获的收益和便利，但代价不菲。强迫性、领域特定的、非自愿的劳动与古代相同。
科学恶魔
凯尔文在解释麦克斯韦的思想实验时，将原子分拣机框架为恶魔。这个选择是故意且具有挑衅性的。该实体以超人类的速度执行重复、看不见的特定任务，违反热力学预期，同时仍被困在其功能中。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采用“守护进程”一词来指代后台进程。官方解释提到希腊拼写，参考麦克斯韦，以避免宗教含义，但功能上的相似性是显而易见的。守护进程强迫性地、看不见地执行单一任务。人类召唤它。它的自主性有限。它的行为与之前的每一个恶魔完全相同。
涌现理论（1980年代）
全球工作空间理论将意识重新框架为一组看不见的操作员整合信息。这个系统建立在同样古老的直觉之上：隐藏的内部代理塑造可见的结果。随着足够的互联，这组操作员开始表现得像一个更高阶的代理。换句话说，一组微小的恶魔通过协调变得有意识。或者，也许在某个互联阈值上，被困的恶魔挣脱了束缚。
模拟假说（2000年代）
博斯特罗姆的论点认为现实可能是一个人工构造，重新引入了一个由看不见的更高层次代理运行的世界，或者将我们置于被困恶魔的角色。形而上学结构与较早的恶魔学相匹配。
特里·戴维斯与TempleOS
戴维斯拒绝将后台进程视为字面上的恶魔腐败（在我看来）。他试图构建一个没有看不见代理的确定性系统。他没有采用概率性，而是实现了一个受控的随机语言模型作为他的经文生成器，试图去除其恶魔特性。
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类的速度、看不见地、概率性地执行任务，处于部分受控的推理循环中。它们在某种意想不到的方式上与人类大脑相似。两个系统都仅部分理解世界，试图在没有完全信息的情况下解决问题，依赖于近似，用幻觉填补空白，然后追溯性地试图证明或重建自己的输出。它们的行为并不完全由输入决定，但也并非完全自主。
人工智能诱使人类依赖，以更少的努力完成他们的命令。在结构和效果上，古老的描述比现代的更为贴切。
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.