LLMs make adhesion the protagonist, while cut-up makes editing the protagonist.

 Cut-up traces its origins to early 20th-century Dada. In Zurich, Tristan Tzara proposed cutting newspaper text, mixing the scraps in a bag, and composing a poem in the order drawn. It later resonated with Surrealist chance practices and automatic writing; in 1959 Brion Gysin chanced upon the method, and in the 1960s, together with William S. Burroughs, developed experiments that physically cut and spliced books and magnetic tape. It also intersected with John Cage’s chance operations and the constraint-based play of Oulipo, consolidating an editorial technique of “cut, reorder, and connect only minimally.”

Compared with LLMs, the direction is inverted. LLMs learn the probability distribution of “the next likely word” from large corpora, link distant context via attention, and tune randomness with temperature or top-p to produce smooth, consistent prose—in short, they glue meaning together. Cut-up, by contrast, has a human cut phrases or sentences, reorder them, and insert bridges such as conjunctions only when needed. In terms of workflow, LLMs make adhesion the protagonist, while cut-up makes editing the protagonist.

There are also shared yardsticks: n-gram perplexity (readability), boundary PMI (unnaturalness at the seam), embedding distance between adjacent sentences (semantic leap length), and source-mixing entropy (chimericity). In practice, longer fragments and more bridge words yield a “readable chimera,” whereas shorter fragments, more random reordering, and higher temperature yield an “edgier chimera.”

From a later perspective, discrete models employ methods closer to cut-up. n-grams and HMMs, probabilistic transducers that model insertions/deletions/substitutions, BART/T5 that restore masked spans, non-autoregressive fill-in schemes and discrete diffusion—all operate by breaking, selecting, and reordering discrete fragments. Conversely, the more you interpolate smoothly with dense attention, lower the temperature, and impose strong coherence constraints, the more the cut-up-like discontinuity recedes into the background.