这个项目开始于把The Waste Land的声音以及意义放在同一个几何空间里,通过声音以及意义之间的矛盾,试图让计算机理解现代诗歌。但是很快我就发现这是不可能的。
大模型理解语言的方式是将文字符号的结构作为句法(syntax),将词语在embedding空间中的位置作为承载理解的初始条件以及意义的可能性空间。 而在大模型运行过程中,通过attention mechanism的不断加权,表示上下文之间的关系,这些在语义(semantics)层面词语的位置会不断地在空间中发生漂移,而trajectory本身就是大模型理解语言的过程。 大模型在运行的过程中有一个loss function,也就是它赖以存在目标,这个目标是预测下一个token,或者说是将下一token预测的不确定性降到最小。 这一点本身其实是语用(pragmatics)压力被编码成了一种能量函数的硬约束。 通过大量文本、语言应用习惯的平均,语言的理解方式被强制固定在了确定的轨道上面。 这是为什么它拥有整个语言的可能性空间,同时有能力在这个空间里表示多个意义系统(这个空间包含但丁、莎士比亚以及T.S. Eliot的所有),但不可避免地会倒向某个看似确定的意义。意义的并行,一个空间中有多个奇异吸引子的现代诗的结构的理解是大模型结构上不兼容的。
这正是我们生而为人独享的一件事。我们能够读懂The Waste Land, 我们能感受到The Waste Land里面大模型所不能感受到的东西。正是因为我们并不会为了追逐下一个Token的确定性而强制要求意义闭环, 我们并不会把不同的相互矛盾的感受浓缩成一个极其自洽的表述。 而诗里那些神话的时间、语言以及我们私人的时间和语言可以共存在同一段话、同一个词里。 比如,梵语的告诫既可以是古典传统的告诫本身,又可以是非常身体化感受的雷声;又比如,我们可以把对战争记忆的恐惧和握在手上的一抔土放在一起理解。
我觉得相似的,这种功用上的压力同时也发生在当代资本主义的市场价值系统里,这并不是简单的劳动的异化或者是资本的剥削。而是说我们被抛掷到这个世界上来,如果说我们也有一个完整的可能的意义空间的话,在我们人生的选择的过程当中,通过一个一个行动之间的上下文的关系,以及行动的因果的对照,我们本来是可以有更加丰富的可能的方向,而且它们是有可能共存的。但是在单一的价值函数的压力之下,人往往会被合理化收敛到一个单一的轨道上,忘记了空间里本身对于我们来说,本应该拥有的那些奇异吸引子结构。
如果说前现代的空间里还有作為主體的作者、诗人、制度的推動者、啟蒙者、革命者这些山峰让我们反墒攀登的话,后现代的人则是在山峰崩塌或者是被遮蔽之后在系统里迷失、滑移的点而已。
This visualization maps each word in "The Waste Land III: The Fire Sermon" across 10 linguistic dimensions. Words are positioned in 3D space based on their values on three axes (X, Y, Z) selected from the framework below.
Axes vs Dimensions:
How the word feels in the mouth and ear during performance.
"The river's tent is broken: the last fingers of leaf
Clutch and sink into the wet bank."
Closed vowels (−1): "broken" /oʊ/ → /ə/, "fingers" /ɪ/, "leaf" /i/ — These create constriction, withholding. The "i" in "sink" feels tight, drowning.
Open vowels (+1): "clutch" /ʌ/, "bank" /æ/ — These feel expansive, grasping. "Clutch" opens the throat before the hard stop.
Why it matters: Open vowels in "The Waste Land" often appear at moments of loss or exposure (drought, barren land). Closed vowels appear in constricted, interior states.
"Sweet Thames, run softly, till I end my song."
Liquid/flowing (+1): "Sweet" /sw/, "run" /r/, "softly" /s/ + vowels — The refrain flows like water itself. "Thames" has liquid /m/ and /z/.
Obstructed (−1): When the poem shifts to pollution: "bottles, cardboard, cigarette" — hard stops /t/, /k/, clusters.
Eliot's core tension: River (flow, liquid, life) vs. Drought (obstruction, blockage, sterility). This axis directly maps Waste Land's water/desert dialectic.
"At the violet hour, when the eyes and back
Turn upward from the desk, when the human engine waits"
Strained (+1): "back" /bæk/ — hard /k/ stop after open vowel. "desk" /dɛsk/ — cluster /sk/. These words require tongue/lip pressure.
Easy (−1): "waits" glides on /w/, though its meaning is passive tension.
Body perception metric: This isn't IPA phonetics but felt performance. "Clutch" feels broken/explosive. "Sweet" slides out effortlessly. The mechanization passage uses strained articulation to mirror bodily alienation.
Not "what it means" but how the word orients toward presence, body, time, action.
"A rat crept softly through the vegetation
Dragging its slimy belly on the bank"
Highly embodied (+1): "belly" — visceral body image. "slimy" — tactile, repulsive. "dragging" — physical motion.
Abstract (−1): "memory", "thought", "meaning" — cognitive, non-sensory.
Special case — Mythic/Symbolic: "Tiresias", "Shantih", "Phoenix" — These are CONCRETE (not abstract) but non-bodily. They exist in ritual/mythic space. We annotate high concreteness but note symbolic function.
The Waste Land's dialectic: Hyper-embodiment (visceral decay) vs. mythic displacement (quoted rituals). The body is both intensely present and alienated.
"The time is now propitious, as he guesses,
The meal is ended, she is bored and tired,
Endeavours to engage her in caresses
Which still are unreproved, if undesired."
Critical insight: Many words in The Waste Land are grammatically ACTIVE (verbs) but existentially PASSIVE or NULL.
Passive (−1): "unreproved", "undesired", "bored", "tired" — These describe states of being acted upon or negated will.
Active (+1): "Endeavours" — grammatically active, but hollow. "engage" — volitional attempt, but mechanical.
Why this matters: Eliot's modernist horror isn't violence but the dissolution of agency. People don't act — they "wait", "throb", "are bored". The typist encounter is pure passivity masked as action.
"The river's tent is broken: the last fingers of leaf
Clutch and sink into the wet bank. The wind
Crosses the brown land, unheard. The nymphs are departed."
Absent (−1): "departed", "empty", "lost", "broken" (as absence of wholeness)
Present (+1): "here", "now", "this", "wet" (immediate sensory presence)
Quoted/Displaced (special): "Sweet Thames" — physically present river but displaced through Spenser quotation. Marks cultural fragment as both here and not-here.
Opening gesture: Section III begins with DEPARTURE. "The nymphs are departed" sets the tone: presence as always-already absence. Even embodied details (wet bank, brown land) exist in the wake of what's gone.
"I Tiresias, old man with wrinkled dugs
Perceived the scene, and foretold the rest—"
Timeless/Mythic (−1): "Tiresias" — exists across all time. "foretold" — prophetic, eternal return. "always", "death", "memory"
Continuous (0): "waiting", "throbbing", "drifting" — durational verbs, stretched time
Instant (+1): "crack", "now", "sudden", "violet hour" (specific moment)
Eliot's mythic method: Overlays mythic time (Tiresias, Philomel, Buddha) onto modern London NOW. The Waste Land exists in multiple temporal registers simultaneously. This axis captures that simultaneity.
This axis captures Eliot's collage structure at the word level. Words can be syntactically continuous but semantically fragmented.
Fragment (−1): Proper nouns, cultural references, interrupted phrases — "Shantih", "HURRY UP", isolated images
Continuum (+1): Conjunctions, flowing syntax, sustained imagery — "and", "while", lyrical refrains
How the word functions in Eliot's modernist collage. Who speaks? At what register?
"I Tiresias, though blind, throbbing between two lives,
Old man with wrinkled female breasts, can see
At the violet hour, the evening hour that strives
Homeward, and brings the sailor home from sea,
The typist home at teatime, clears her breakfast, lights
Her stove, and lays out food in tins."
Speaker-owned (+1): "I Tiresias" — But who IS Tiresias? Both speaker and mythic mask.
Anonymous (0): "The typist", "the sailor" — collective modern figures, no individual voice
Borrowed (−1): "Sweet Thames, run softly" — Spenser's "Prothalamion". "Shantih shantih shantih" — Upanishads. "HURRY UP PLEASE ITS TIME" — pub voice, class-marked
Why this eliminates sentiment analysis: The Waste Land has NO stable "I" to carry emotion. Voices are borrowed, fragmented, displaced. What sentiment belongs to a speaker who is Tiresias+typist+Dante+Buddha? This axis tracks that radical instability.
"Sweet Thames, run softly, till I end my song.
...
The river bears no empty bottles, sandwich papers,
Silk handkerchiefs, cardboard boxes, cigarette ends
Or other testimony of summer nights."
Lyrical/ritual (+1): "Sweet Thames, run softly" — Spenserian refrain, elevated language. "Shantih" — ritual closure. "violet hour" — poetic diction
Documentary/neutral (0): "cardboard boxes", "cigarette ends", "sandwich papers" — catalogue, itemized description
Colloquial (−1): "HURRY UP PLEASE ITS TIME" — pub speech. "Ta ta" — vernacular goodbye. Class-marked London voices.
Register jumps = Waste Land tension: Within ONE stanza, Eliot moves from ritual refrain to trash inventory. This isn't stylistic variation — it's the core method. High and low collapse into each other. Sacred and profane occupy the same line.
This visualization uses LLM-generated annotations. The positions you see represent the LLM's context-aware understanding of the poem — each word is annotated based on its surrounding context within "The Fire Sermon."
What the LLM does: It tries to collapse all meaning into the same kind of system, placing each meaning dimension on a very smooth, continuous axis. This allows the LLM to localize everything very precisely and coherently. Every word gets a definite position on each scale: embodiment from −1.0 to +1.0, presence from −1.0 to +1.0, and so on.
The smoothing effect: The LLM's architecture is fundamentally designed to reduce uncertainty — to predict the next token by converging toward a single coherent interpretation. When it encounters "Sweet Thames," it integrates the Spenser quotation, the river imagery, the pollution context, and produces a set of smooth coordinates. The meaning is localized in semantic space.
What this reveals about LLMs and poetry: As argued in the Manifesto, this approach is structurally incompatible with how modern poetry like The Waste Land actually works. Eliot's method creates multiple strange attractors — the mythic and the modern, the sacred and the profane, presence and absence — that coexist without collapsing. A human reader can hold "burning burning burning burning" as simultaneously purgatorial fire (Buddha), sensory intensity, mechanical repetition, and ritual incantation. The LLM must choose coordinates.
This visualization makes that tension visible. The smooth trajectories you see are not how meaning works in The Waste Land — they're how a prediction-optimizing system must represent meaning in order to function. The gap between these smooth paths and the fractal, non-collapsing structure of poetic meaning is the space where human understanding lives.
Each word is plotted in 3D space using three axes from the 10-dimensional framework:
Context is everything: A word's position changes based on surrounding words. "River" alone might be neutral embodiment, but "Sweet Thames" in the Spenser refrain shifts toward ritual register and quoted voice ownership.
Not smoothing — selective sharpening: Context doesn't blur distinctions. It heightens them. Eliot's collage structure creates CONTRAST. Position reflects how each word functions in its specific moment in the poem's unfolding — as interpreted by the LLM's context-aware but necessarily smoothing annotation process.