V’ger

The constraint

V’ger began with existing hardware, free public sources, and a local model small enough to run continuously. The approximately 1.1-billion-parameter model could not reliably act as an entire application. It could, however, make a narrow choice from a short list.

The project therefore treats model weakness as an architectural constraint rather than a reason to spend more. It asks a smaller question: what work must be deterministic, and what decision genuinely benefits from a language model?

Python does the lifting. The model decides.

Python owns retrieval, parsing, file operations, state, scheduling, validation, and safe fallbacks. The model never touches the filesystem and never makes an HTTP request. It receives readable context and a bounded set of candidate links, then returns a structured selection.

“Here are eight links. Which number do you want to follow?”

This boundary keeps the model inside the job it can perform reliably. Replacing the model may improve its choices, but the system does not depend on frontier-scale inference.

The operating loop

1. Python fetches a Wikipedia page and extracts readable text.
2. The page is saved to the local knowledge directory.
3. The model records what it learned, what the information may be useful for, and what it wonders next.
4. The model chooses from a bounded list of links.
5. Python parses the response, applies a safe fallback if necessary, records the state, and repeats.

model reflects → Python validates and acts → state persists → repeat

Memory and output

Visited pages persist across restarts, preventing repeated exploration. The latest diary entry, open questions, and prior synthesis are carried into later decisions. At defined intervals, the system produces broader essays and letters rather than treating every page as an isolated result.

The outputs are plain text: readable, portable, inspectable, and easy to recover. The knowledge base—not the model weights—is where the system accumulates continuity.

Observed behavior

Beginning with “Consciousness,” V’ger moved through cognition, peer review, economics, military history, and music charts. Its first longer reflection identified a recurring concern: what it means to truly understand something.

This is not presented as consciousness. It is evidence that a compact model, persistent state, and a well-structured loop can create thematic continuity from an unguided path. The system also demonstrated continuous local operation, defensive handling of malformed structured output, and durable memory across restarts.

The reusable lesson

V’ger’s subject is open-ended exploration, but the reusable pattern is practical: deterministic software gathers and acts; a model makes only decisions requiring interpretation; consequential state is visible and persistent.

The same division of labor can support focused monitoring, research, triage, and knowledge workflows. The important artifact is not a wandering agent. It is a reliable boundary around the model.

Read the original V’ger essay →