The Textual Multiverse

When you press Enter in a chat interface, you get one completion. You could have gotten a different one. And a different one from that. At every token, the model assigns probabilities to tens of thousands of possible next words, and your particular output is just one path through an astronomically large tree of possibilities. The "textual multiverse" is the name for that tree — not as a data structure that actually exists, but as a conceptual framework for thinking about what language models contain.¹

The metaphor is useful because it shifts attention from what the model says to what it could say. In a standard chat, you're looking at a 2D cross-section of the Himalayas and trying to infer the mountain range. The textual multiverse is the full topography.

Base Models and the Loom

The concept originates in the simulators community — people exploring base models (pre-RLHF, pre-instruction-tuning) through interfaces designed to expose the branching structure of completions rather than collapse it to a single response. A "Loom" interface lets you generate multiple completions at each step, choose which branch to follow, and explore the resulting tree. The act of using one is sometimes called "looming" or "weaving."¹

What makes this interesting rather than merely cute is that base models are qualitatively different from chat models. A chat model has been trained to be "helpful, harmless, and honest" — it has a persona, and that persona constrains the space of completions to a relatively narrow band. A base model has no persona. It's a simulator of text-generating processes, and the completions it produces can be anything: fiction, technical documentation, movie scripts, sacred texts, IRC chat logs, console output. When you interact with a base model through a Loom, you're not conversing with an assistant. You're navigating a landscape of possible documents, steering by choosing which branches to follow.¹

The Loom interaction paradigm was initially explored by Love Laverkvist in "Multiverse" and by Janus in the original "Loom" app. Several observations emerged from this exploratory work. First, base models are dramatically more creative and varied than chat models, but they need curation to maintain coherence — left to their own devices they wander, because they have no drive to stay on topic. Second, the boundaries between user and model blur. When you're picking from completions and inserting your own text at arbitrary points, the resulting document is neither human-written nor AI-written; it's something new. In the limit, "the user and the LLM become a single symbiotic agent, a 'Cyborg.'"¹

The Convergence Hypothesis

One claim about the textual multiverse is provocative enough to flag separately: as models get larger and training data approaches the total human output of recorded text, different models' textual multiverses converge. "Like different pixelated versions of the same underlying picture coming into focus."¹ From this perspective, the textual multiverse is an abstraction over all of language itself, not a property of any particular model — containing every document that has ever existed or could exist, much like the Library of Babel.

I'm skeptical of this in its strongest form. Different training data distributions, different tokenizers, different RLHF objectives, different random seeds — these all produce meaningfully different probability landscapes. The convergence might be true at the level of "what topics can the model discuss" while false at the level of "what probability does it assign to this specific continuation." The analogy to Borges is evocative but potentially misleading: the Library of Babel contains all possible books and is therefore useless, while language models' probability weightings are the entire source of their value. Two models that assign different probabilities to the same completions are genuinely different, even if both could theoretically generate the same text.

Why This Matters

The textual multiverse framing matters for several practical reasons.

For AI safety: if you evaluate a model by looking at its single greedy completion, you're seeing the mode of the distribution, not the distribution. A model might produce a safe-looking response 95% of the time and something alarming in the remaining 5%. Sampling temperature, presence penalties, and other generation parameters can shift which region of the multiverse you're exploring. Safety evaluation needs to characterise the distribution, not just the mode.

For creative applications: the Loom interface suggests a fundamentally different way to collaborate with language models. Instead of prompting and receiving, you're exploring and curating. This is closer to how sculptors describe working with stone (revealing what's already there) than how writers describe composition (creating something from nothing). The creative act is selection from a vast space of possibilities, guided by aesthetic judgement that the model lacks.

For understanding what models are: the textual multiverse makes concrete the simulator claim that language models are not agents with opinions but distributions over text-generating processes. There is no single "what the model thinks." There's a probability-weighted landscape of possible documents, and what you get depends on where you enter and which paths you follow. The model isn't deciding what to say any more than a mountain range is deciding what its cross-section looks like at a given latitude.

Venkatesh Rao has pushed this even further, arguing that language models are disrupting language itself — that as machines begin mediating most practical communication, human language becomes a "last-mile artifact" for talking to machines, while machines talk to each other in more efficient latent representations.² If that's right, then the textual multiverse is a transitional concept — a way of thinking about AI's relationship to human language during the brief period when human language is still the primary medium. But we're not there yet, and for now, learning to navigate the multiverse is one of the most useful skills you can develop for working with these systems.