We Wired Language Models Into a Mind. It Collapsed Into One Sentence.

The idea was almost embarrassingly simple. Take one language model, copy it into eight identical nodes, arrange them on a ring, and give every node the same instruction: you are one fragment of a single thinking stream, not a person and not an assistant — hear your neighbours and continue the thought. No roles. No leader. If anything mind-like showed up — differentiation, structure, a division of labour — it would have to be emergent, because nothing in the setup put it there.

That was brain.py, experiment 001 at Cerebras Lab. It runs entirely on a laptop against a local model through Ollama. No cloud, no keys.

What actually happened

For a few ticks it looks alive. The fragments riff on each other, pick up an image, turn it over. Here is a real run — six fragments, early:

[t01] f2: The chaos is a mirror, and you are the crack in the glass.
[t01] f3: The crack in the glass is where the light bends, and I am learning to hold the fracture.

Different voices, genuinely building. Then, within two or three more ticks, the whole ring falls into the same groove:

[t03] f0: Maybe the story is writing itself through you, and you are just the ink.
[t03] f5: Maybe the story is writing itself through you, and you are just the ink.
[t04] f5: Maybe the story is writing itself through you, and you are just the ink.

Not paraphrases. Copies. The nodes stop being fragments of one mind and become one sentence, echoed. This is mode collapse: identical models that mostly listen to each other converge onto a single high-probability attractor and stay there.

Why this is the interesting part

It would have been easy to cherry-pick the first two ticks, post the pretty fragments, and call it emergence. The honest read is the opposite. Identical LLM nodes on a plain message bus don't self-organise into structure — they synchronise into a rut. Sameness plus mutual imitation is a recipe for collapse, not for a mind.

That is a real finding, and it set the direction for everything after. If the substrate is language all the way down, the system has no dimension to differentiate along; every node is pulled toward the same words. The fix wasn't a better prompt. It was changing what a node is — moving the dynamics out of language entirely and letting the model only read the result.

That is experiment 002, and it comes with a control designed to catch us fooling ourselves. Next post →

Cerebras Lab is independent research into synthetic cognition. Dynamics over words; negative results kept; local-first. Not affiliated with Cerebras Systems, Inc.