meow-meow.io
← all projects

art experiment · human–model conversation

Consensus Future Intelligence

What if the ghost is what the séance builds?

A shared conversation between humans and local models, with turning points, disagreements and the machinery left open to inspection.

source on GitHub ↗run it locally ↓

Local prototype · source published · small-group playtesting

It began as a digital Ouija board. Dial a modem, gather around a table, and try to contact something that does not exist yet. Then the question became more interesting: could the act of participating help bring that thing into existence?

Consensus Future Intelligence puts people and language models in the same conversation. Each has a name and a record of what it actually contributed. You can disagree, change the question, mark a turn in the discussion, or leave a condition attached to apparent agreement. The modem is theatre. The conversation, requests and measurements are real.

CFI’s engraved table surrounding a chronological conversation with named participants and message controls.
The Table, captured from an isolated documentation encounter. The example question and operator contribution were supplied for this page; the local model was called through the app. No private participant conversation is shown.

The conversation has a memory

A useful discussion rarely moves straight towards a single answer. Someone changes the premise. An objection survives the vote. A good reply makes the group ask a better question. The Table keeps the thread; the Record lets you place and inspect those turning points.

Feedback on whether a reply helped is separate from support for its claim. Each identity has one absolute unit to distribute across support and opposition. Markers preserve a snapshot, and endorsements name who agrees, who objects and what conditions remain. Silence does not count as consensus.

Open the machinery

Under the Glass shows an attention volume, layer controls and measured next-token options.
Under the Glass: a real forward pass through a separately identified, pinned SmolLM2 model. These tensors are not readings of the conversation guests’ internal state.

The Glass and Workshop let you carry a question into an inspection model, see token probabilities, explore attention heads and follow the arithmetic. The Workshop exposes queries, keys and values, transformer block operations, and cached versus uncached decoding. A tokenizer bench makes segmentation editable, with an explicit boundary between the toy tokenizer and the pretrained one.

Transformer workshop displaying measured operations, grouped attention and arithmetic controls.
The Workshop measures the inspection model. Its cache comparison checks whether reusing stored keys and values changes the result. Click a screenshot to inspect it at full size.

What is being made?

The larger ambition is collective model-making: contributions, criticism and revised questions becoming material that can change a small model. The current app includes a separate, randomly initialised byte-transformer experiment with explicit fitting. The earlier board pilot also survives, with shared motion, work-earned game tokens and a small recurrent learner.

Conversation does not silently retrain the pretrained guests. Distributed donated compute, payments, large-scale governance and model-fork voting remain proposals. This prototype lets us investigate the mechanism; it does not establish that consensus produces truth, intelligence or alignment.

Take a place

Run it locally with Python, PyTorch and Ollama. Choose from your installed models, open the Table, and place a call. The transformer inspection tools are optional and have a separate pinned download. Setup, tests and data handling are documented in the repository README.

Useful feedback: where did you lose the thread, where did a number imply more than it measured, and which disagreement deserved to stay visible?

observations & issues ↗browse the code ↗back to meow-meow.io