Loom is a tree-shaped interface for writing with language models, built by janus and presented in February 2021 for the GPT-3 API. Instead of accepting a single continuation, the user generates several, explores them as branches, backtracks, prunes and merges: the text stops being a line and becomes a navigable "multiverse." The premise, stated on janus's own site, is that "language models are multiverse generators": each prompt opens a distribution of possible futures, and sampling picks one. Reynolds & McDonell (2021) developed the framework in a paper on "multiversal" views of language models.
Loom is also the emblematic tool of cyborgism, proposed in a LessWrong post of February 2023 signed by NicholasKees and janus: instead of handing work over to autonomous agents, use models as prostheses that extend human agency, with the person always in the loop. Exploring the tree of continuations is also a research method: it lets one see the distribution rather than a single sample, which connects it to the simulators thesis (janus, 2022).
The song ends on a prophecy: "Just as foretold by Loom" (osmarks, n.d.). The line also alludes to the collection of "prophecies" janus keeps on the same site, a chronological anthology of quotations about language, minds and machines that runs into future years and that, the site warns, includes apocryphal passages. The joke works on two levels: models "predict" text, and a well-tended tree of predictions can be read as an oracle.
For linguists and philosophers, Loom makes tangible what the chat interface hides: that a model's "answer" is one sample from a probability distribution, and that the impression of a single voice is an effect of sampling. The reasonable objection to cyborgism is the one raised against any craft practice: its findings depend on the operator's skill and are hard to reproduce.