Loom

AI Dungeon kept one history. A continuation became the story, and there was no way to hold two candidate next paragraphs open side by side and pick between them later. Janus built Loom in 2020 to fix that. Loom shows a language model's possible continuations as a branching tree rather than one line of chat.

Set against a chat interface, the difference is what gets kept. A chat UI holds one history and advances it, so every reply discards the alternatives a different sampling run would have produced. Loom keeps the tree. A continuation is one branch among several, and you can go back up and take a different one without losing the first.

That is a claim about an interface, not about a model. Loom does not improve a model's performance. It changes what you can see of what the model already produces: many possible continuations rather than one answer.

Same author, two artifacts, two dates

Janus also wrote "Simulators", posted to LessWrong on 2 September 2022, which argues that a base model is best read as a simulator that can roll out many characters rather than as an agent.

The two get collapsed into one dated thing, and they are not the same. The tool is 2020. The essay is 2022. The fact pass behind The p(doom) song, annotated caught exactly this slip in an earlier draft, which had dated the tool by the essay. If you see Loom described as a 2022 tool, that is the error.

Who ended up using it

Loom became part of the cyborgism circle's kit for working with base models, which is where the tool and the essay meet in practice. The argument and the interface fit each other: if what a base model holds is many possible continuations rather than one intended answer, then a view that shows one line at a time discards most of what is there.