Language as VM
The core idea: a language whose utterances don’t describe — they execute. The runtime is LLM inference itself.
Traditional prompting treats language as description. You describe what you want, the model interprets, produces output. The language is a vehicle for intent.
A language VM treats language as bytecode. Utterances are instructions. The model is the processor. Execution is inference.
Why This Is Different
In descriptive prompting, meaning is reconstructed by the model from natural language. There’s translation overhead — human linguistic structure → model processing → output. Ambiguity lives at every step.
In a language VM, tokens are chosen not to describe concepts but to activate them directly in the model’s probability space. The token is the concept. No translation layer.
The compression is a side effect: native model-space communication carries less overhead than translating from human linguistic structure.
The Runtime
LLM inference is not a sequential processor. It doesn’t execute instructions one at a time with a program counter. It processes context as a field — all tokens simultaneously, each attending to all others, collapsing into output probability distributions.
This means the VM metaphor requires reframing:
- Opcodes are not sequential commands — they are attractors that shape the probability field
- Execution is not a loop — it is a single forward pass that resolves the entire field at once
- State is not a register — it is the accumulated context window
- Branching is probabilistic, not deterministic
A language VM for LLMs is therefore closer to a field specification language than a traditional instruction set.
The Grammar
See index for the full design.
The key insight driving the grammar: token position matters as much as token identity, because attention is causal — earlier tokens condition all later ones. The grammar exploits this:
DOMAIN → MODE → BINDINGS → DEPTH → OUTPUT_SHAPE
Each layer narrows the field established by the previous. By the time the model reaches the output shape token, most of the probability mass has already been directed.
Connections
- attractors-vs-instructions — the mechanism behind why this works
- token-economics — the compression payoff
- anchor — the frozen primitive substrate the VM operates on