Close × project concept 計畫概念
about · 關於
Protean Language Model centers on four specific agents: the immortelle flower, the human, Rhodobacter soil bacteria, and the ESM2 protein language model. These are figures chosen to stage how human desire moves through biotechnological systems.
The immortelle flower points to longstanding fantasies of rejuvenation and immortality, from Emperor Qin's search for eternal life to contemporary techno-capitalist longevity culture. The human enters as the desiring subject that projects fears of death and fantasies of transcendence onto science and technology.
Rhodobacter figures the microbial labor through which these desires are materially processed and metabolized. ESM2 functions as the machinic framework that renders life computable.
Together, these four agents form a speculative ecology.
異形蛋白質語言模型聚焦於四位主角:永生花(蠟菊)、人、Rhodobacter 土壤細菌,以及 ESM2 蛋白質語言模型。這些角色被選來演繹人類的慾望如何穿越生物科技系統。
永生花指向長久以來對返老還童與不死的幻想，從秦始皇求長生到當代科技資本主義的長壽文化。人作為慾望主體進入，把對死亡的恐懼與對超越的幻想投射到科學與技術之上。
Rhodobacter 指涉微生物的勞動。這些慾望透過它，被物質性地處理與代謝。ESM2 則作為機器框架，使生命變得可計算。
這四個行動者共同構成一個推測性的生態。在此「抗老」作為一種分散於不同系統的文化執念，也形塑不同系統。
Close × methodology · 方法
The entered sentence becomes a score for 3-10 voices. The preserved fragment remains recognisable while the surrounding words change. These are explicit compositional translations made by HSURAE after taking part in the deCYPher project as artist-in-residence. Below are the mappings.
1 · Human text to amino acids Entered text is mapped to UTF-8 bytes, and encoded into codons. Stop codons (TAA, TAG, and TGA) are recorded but omitted from the ESM-2 sequence. This is an authored, lossy mapping; the entered sentence is translated into a synthetic protein.
2 · ESM-2 to text disintegration The synthetic protein from step 1 enters a live ESM-2 protein language model. ESM-2 was trained on millions of natural protein sequences. During training, some amino acids (AAs) in each sequence were hidden, and the model learned to guess them from the surrounding sequence. A similar test is applied here to your synthetic protein. Each AA position is hidden in turn, and the model predicts how likely each of the 20 AAs is to belong there. When the model finds the original AA unlikely, the syllable linked to it begins to erode earlier. Each syllable takes new sounds from the model's guesses at one representative position, through an explicit sound dictionary. The top three guesses are sampled for each voice . (The ESM-2 predictions and sound dictionary are in the drop-down at the bottom of the score.) As it erodes, the articulation passes through a softened or stretched version, model-derived sounds and humming. The timing and sound dictionary are artistic rules, not a biological decoding.
Each voice receives short passages from the sentence. Voices sustain, repeat, call and answer, allowing different passages and stages of disintegration to overlap. The preserved word or phrase is the only part that is never mutated.
3 · Immortelle to desired compound The compound selection changes what happens as the language breaks down. Neryl Acetate regenerates passages: every third phrase restores the surrounding words to their original articulation. The sentence briefly comes back, then starts deteriorating again. Italidiones revisits damaged passages through successive stages of repair, with each version repeated before the next change. Depending on how far the material has deteriorated, this recovery may be partial. Gamma-Curcumene has a protective effect, the words resist breakdown longer. When they change, they stretch into longer vowels and hums, creating a more sustained texture. Each reproduces attributed benefits as artistic prompts based on loose clinical claims.
4 · Rhodobacter to time Rhodobacter supplies the idea of one becoming two, then four, then eight. A study of R. sphaeroides WS8 reports an approximately 160 minute doubling time under certain conditions. That reference is compressed into eight beats per generation: 20 culture minutes per score beat. At 60 BPM, eight beats take eight seconds. It is a chosen reference and an artistic compression, not a universal growth rate or a live microbial measurement.
The population expands 1 / 2 / 4 / 8 voices, capped at the chosen group size. Voice entries occur on beats 1, 9, 17, 25, 33. All voices share 160 beats after a 8 beat count-in. .
5 · Score to body The system turns model predictions into instructions for sounding. Each participant lends a different respiratory tract, articulation, pitch and attention to the same requested persistence. Simultaneous voices enact proliferation.
References Rhodobacter response kinetics and culture conditions Neryl acetate: skin-barrier activity in skin explants Helichrysum constituents and computational VAP-1 docking